!"#$%&'$()* ,(#-.

('$(/'0 1%&#'2&$("#
3$.4(*'0 5&/6(#* 1%&#'2&$("# 7 86&$ /&#
8* 2*&%# 9%": *&/6 "$6*%;

"#$%&'()$#$(&$ *'#+,-'.
/012 34556 789:;#<
3=>- 0$.>$9:$#6 3455

?''+ ') @:,>#8&>,


ABCD$# E;B'6 0CBDC8 78(,$(%0&-C##8
F'-8(($, /;>$(:$#<%G(CD$#,C>H> I8C(J












!"#$"%&


9:30-10:00 lnLroducLlon
MeLhodologlcal cross-ferLlllsaLlon: emplrlcal
meLhodologles ln (compuLaLlonal) llngulsLlcs
and LranslaLlon sLudles

Lrlch SLelner, Sllvla Pansen-Schlrra,
Cllver Culo
10:00-10:30 Can llngulsLlcs and LranslaLlon sLudles beneflL
from M1 quallLy barrlers? needs an
opporLunlLles from a Lechnology perspecLlve
(lnvlLed Lalk)
Pans uszkorelL



11:13-11:40 WhaL Can ConLrasLlve LlngulsLlcs 1ell us abouL
1ranslaLlng ulscourse SLrucLure?
lrn korzen, MorLen Cylllng
11:40-12:03 An analysls of LranslaLlonal complexlLy ln Lwo
LexL Lypes.
MarLha 1hunes
12:03-12:30 Web-based conLrasLlve comparlson of Age-
8elaLed 1emporal hrases ln Spanlsh and
lrench
Sofla n. Callcla-Paro, Alexander l.
Celbukh



14:00-14:23 A ConLrasLlve SLudy on AbsLracL Anaphors ln
Cerman and Lngllsh
SLefanle ulpper, ChrlsLlne 8leger,
Melanle Selss, Pelke ZlnsmelsLer
14:23-14:30 A Corpus-based ConLrasLlve Analysls Lo ueflne
Mlnlmal SemanLlcs of lnLer-senLenLlal
uependencles for Machlne 1ranslaLlon
1homas Meyer, Andrel opescu-
8ells, !eevanLhl LlyanapaLhlrana,
8runo CarLonl
14:30-13:13 lormallslng LranslaLlon behavlor wlLh parallel
Lreebanks
Cllver Culo, Sllvla Pansen-Schlrra



13:40-16:03 uslng annoLaLed corpora for rapld
developmenL of new language palrs ln M1
Susanne reuss, Pa[o keffer, aul
SchmldL
16:03-16:30 hrase 1able SupporL for Puman 1ranslaLlon Cerhard kremer, MaLLhlas ParLung,
SLefan 8lezler, SebasLlan ad
16:30-16:33 lnslde Lhe monlLor model: processes of defaulL
and challenged LranslaLlon producLlon
Mlchael Carl, 8arbara uragsLed



17:00
ulscusslon




!

Methodological cross-fertilization: empirical methodologies in (computational) linguistics and
translation studies
Erich Steiner
Universitt des Saarlandes, Saarbrcken


Recent years have seen a few, although still limited, attempts at improving empirical methodologies in
contrastive linguistics and in translation studies through interdisciplinary collaboration with multi-
layer corpus architectures as developed and refined in computational linguistics. These corpus
architectures provide data enriched by a variety of techniques ranging from shallow to deep processing
(Vela et al 2007, ulo et al 2008). They allow the posing of linguistic questions as empirical
questions even in areas which until recently were considered the province of hermeneutic debates
supported by hopefully representative examples.

At the same time, explanatory background for empirical results is increasingly sought in more
sophisticated models of language contact in typologically based comparative linguistics (e.g.
Thomason 2001, Teich 2003, Doherty 2006, Fabricius-Hansen and Ramm eds. 2008, Siemund and
Kintana. eds. 2008, Steiner 2008, Miestamo et al. eds. 2008, Dunn et al 2011) on the one hand, and in
language processing in situations of multilinguality, including translation, on the other (Alves et al
2010, Carl et al 2008). There remains a significant challenge, though, in closing the gap between the
often necessarily high level of abstraction of models, and the data provided through shallow (and
cheap), or else deeper (and more expensive), analysis and annotation of electronic corpora. This gap
has to be narrowed through concerted efforts involving methodologies from computational linguistics,
including machine translation, (contrastive) linguistics and translation studies.

We shall discuss two test cases from DFG-projects for such interdisciplinary work: one of them
investigates a key notion of translation (explicitation) and the other an under-researched area of
language contact (contact through cohesion). The gap to be closed consists between the notions of
explicitness/ explicitation and contact through cohesion on the one hand, and the level of the available
data (annotation layers, statistics on these, alignment phenomena such as crossing lines, and empty
links) on the other. Seen relative to existing approaches, we are attempting to synthesize individual
parameters of language comparisons and language contact into more general dependent variables
(explicitness, cohesion) on the one hand, and we suggest operationalizations in such a way as to
enable empirical corpus-based (and ultimately also experimental) investigations. An attempt is made
to identify achievements as well as persistent methodological gaps, and implications are identified for
research methodologies.

The first attempt subjected the hypothesis that translations as texts are characterized by the property of
explicitness relative to original texts, and that this explicitness is due to the translation process, rather
than to the factors of register and language (both of which play their independent roles) to elaborated
tests on a corpus of originals and translations, partitioned into registers, between English and German.
The corpora were compiled using sampling techniques (Biber et al 1998) and annotated for PoS,
morphology, chunks, syntactic functions, clauses and sentences. A second, and important, source of
data were alignments between originals and their translations on all of the levels annotated. The
notions of explicitness and explicitation were given a careful operationalization in terms of the types of
information contained in our data. We shall present a sub-set of the results and argue that it was
possible to show whether and to what extent explicitness and explicitation can be traced in the
available data. The independent variables language system, register and translation can be reasonably
isolated and related to the observed effects in the data, but the third one of these, if interpreted as
translation process, is inherently complex and at present still insufficiently-understood (cf. also
Becher 2010). This shortcoming can be systematically addressed by subjecting the notion of
translation process to a more detailed analysis and by independently testing its effect in processing
studies involving the cumulation and intersecting of data from key-stroke logging, eye-tracking and
post-hoc protocols. As a first evaluation of this line of research, it will be argued that the general
corpus-architecture and the processing employed can be trusted to yield more and also
methodologically refined results of the type indicated here, but that we need improvements in the areas
!

of modeling (internally over-complex variables, representativeness of data), operationalization of the
models in terms of linguistics features, and in processing techniques for corpus data (processing
pipelines, evaluation and significance of findings) and for experimental data (amount and naturalness
of data and experimental design).


The second attempt sets out from the diagnosis that our current knowledge about English-German
contrasts in cohesion is weak. We do have reasonably comprehensive system-based accounts for
contrastive grammar, yet even these are not yet backed-up by empirical validation. For cohesion, not
even a system-based comparison is available, much less an empirical foundation for such a
comparison. The tracing of contact phenomena on the level of cohesion is therefore necessarily still in
its infancy (but cf. Hansen-Schirra et al 2007 for an early attempt). We shall argue that substantial
advances in technologies using multi-layer annotated electronic corpora for text-based investigations
of phenomena of cohesion hold the promise of placing constrastive accounts on an empirical basis,
and beyond this comparison also allow us to trace contact phenomena in suitably configured corpora.
A multi-layer representation will again be used, approaching tree-bank functionality and including
aligned data for English and German translations in both directions as a crucial empirical base.
Extensive frequency information about cohesive configurations will be incorporated, tied to varieties
or registers of the language concerned.

One of the interesting questions is that of whether contrastive properties of cohesion in the two
languages point into the same direction as some assumed generalizations in contrastive grammar
(directness of mapping from semantics to grammar, different tolerance of various forms of ellipsis,
more explicit encoding in one of the languages in the clause, possibly the opposite tendency in the
verb phrase, etc.), or whether cohesion serves as a dialectic counterpart, distributing constraints not in
the same direction as in grammar, but possibly in the opposite one. A further interesting object of
investigation is the nature of cohesive chains (frequency, length, distance between elements, etc.).
Our corpus-linguistic analysis includes identifying various types of cohesive devices (reference,
substitution, ellipsis, coherence relations, lexical cohesion), the linguistic expressions to which they
connect (the antecedents), as well as the nature of the semantic ties established and properties of the
cohesive chains where appropriate. Including translations in the analysis should provide evidence for
analogies between cohesive devices in the two languages, but also show areas where one-to-one
equivalents are not preferred, or even non-existent.

The currently existing annotation requires an expansion in terms of additional layers of annotation.
For instance, particular cohesive devices establishing reference or substitution can be investigated on
the part-of-speech level. Other types such as cohesive conjunctions can be identified when examining
the part-of-speech as well as the chunk level. For the investigation of ellipsis combined queries into
different layers of annotation can be employed. However, for the analysis of nominal, verbal or clausal
ellipsis the current annotation is too shallow and does not permit a fine-grained differentiation of types
of linguistic devices. Thus, more specific cohesive categories have to be developed and annotated.

In order to narrow the gap between the concept of contact through cohesion and the level of our data,
a structured grid of hypotheses is specified for empirical analysis as a testing ground for

- contrasts in the uses of similar systemic resources

- contrasts in the use of different systemic resources for similar cohesive functions/ purposes

- traces of language contact due to different usages in contact vs. non-contact varieties (categorical
and/ or in terms of frequency).


Examples of such hypotheses are:

!

Hypothesis 1: 3
rd
person singular neuter pronouns vs. masculine and feminine pronouns (frequency
E(nglish)>G(erman) for originals (contrast)), in terms of PoS overall and proportionally within their
word class.

Hypothesis 2: ETrans(lations)>EO(riginals) in non-ambiguous 3
rd
person reference and ETrans-
T(arget)T(ext)>GO(riginals)-S(ource)T(exts) in explicitated 3
rd
person reference through use of fully-
lexical TTequivalent of pronominal source

Hypothesis 3: E>G in cohesive usage of it (because of alternative usage in German of demonstratives
of various sorts and pronominal adverbs), measured both in terms of PoS overall and as proportion of
cohesive vs. non-cohesive usage of it.

Hypothesis 4: EO > GTrans > GO in cohesive usage of it because of interference in GTrans

Hypothesis 5: In terms of the phenomena tested in H1 H4, we predict that in a comparison of
originals and translations (always within one and the same language and register), the translations will
diverge from the originals in the direction of their source language.

Further hypotheses are developed for comparisons of vagueness/ ambiguity of reference and scope.
Differences can be expected here deriving from usage of different lexicogrammatical realizations of
some constant cohesive relationship, or even from different cohesive relationships altogether. An
example would be the contrastive use of a generic full lexical phrase vs. a definite phrase vs. a phrase
pre-modified through a determiner (possessive vs. deixis vs. demonstrative) vs. a phrase headed by a
pro-form (demonstrative vs. pronoun) as tested on aligned ST-TT pairs. The interest would not be in
the phenomenon as such, but in the different kinds of ambiguity and/ or vagueness associated with
each case. In general, we would predict that a) translations are less ambiguous and vague than their
originals in SL-TL configurations (explicitation through translation), but also b) that they diverge from
their original register-identical counterparts in the direction of the respective source language
(interference, shining-through).

A final type of hypothesis will make reference to contrastive register-specificity of cohesive
configurations, and again their behaviour under contrast vs. contact conditions. These configurations
will be operationalized as length of lexical or referential chains, density of chains, number of chains
per text sample, etc. At this stage, we would hypothesize shining-through effects for ST-TT
configurations, and for density of chains only a possibly increasing effect of the translation process as
such. Our main argument will be that the frequency data that can be obtained through work of the type
described here is valid and interesting in itself, and is furthermore only possible through the joining of
efforts from (contrastive) linguistics, translation studies, and computational linguistics.

What remains a task for the immediate future in research attempts of the type discussed is an improved
understanding of the cut-off point between very costly deep (and less reliable) annotation, and more
shallow (and more reliable) annotation, the latter of which leaves a substantial gap between data and
interpretation. We shall also raise the questions of how research architectures can be made more
standardized than hitherto, allowing independent repetition and (dis-)confirmation of findings, and of
how corpora, their processing pipelines and evaluated results can be related to experimental
(processing) studies to pave the way towards more principled explanations of the results obtained.



References:

Alves, F., Pagano A., Neumann S., Steiner E. & Hansen-Schirra S. 2010. Translation Units and
Grammatical Shifts: Towards an Integration of Product- and Process-based Translation Research. In:
Shreve, G.M. & Anglone E. (eds.) Translation and Cognition. Benjamins.109-142

!

Becher, Viktor. 2010. Abandoning the notion of translation-inherent explicitation: against a dogma of
translation studies. In: Across Languages and Cultures 11 (1), pp. 128 (2010)

Biber, Douglas, Susan Conrad, and Randi Reppen, 1998 Corpus Linguistics. Investigating Language
Structure and Use. Cambridge: Cambridge University Press.

Carl, Michael, Arnt Lykke Jakobsen, and Kristian T.H. Jensen 2008 Studying human translation
behavior with user-activity data. In Proceedings of the 5th International Workshop on Natural
Language Processing and Cognitive Science, NLPCS 2008, Barcelona, Spain, June 2008, Bernadette
Sharp and Michael Zock (eds.), 114123.Setbal, Portugal: INSTICC Press.

ulo, Oliver, Silvia Hansen-Schirra, Stella Neumann, and Mihaela Vela 2008 Empirical studies on
language contrast using the English-German comparable and parallel CroCo Corpus. In Proceedings
of the LREC 2008 Workshop "Building and Using Comparable Corpora",Marrakesh, Morrocco, 31
May 2008, 4751.

Doherty, M. 2006. Structural Propensities. Translating Nominal Groups from English into German.
Benjamins, Amsterdam:
Dunn, Michael, Simon J. Greenhill, Stephen C. Levinson & Russell D. Gray. 2011. Evolved structure
of language shows lineage-specific trends in word-order universals in: Nature 473. 2011: 79-82

Fabricius-Hansen and Ramm eds. 2008 Subordination vs. Coordination in sentence and text: a
cross-linguistic perspective. Amsterdam, Philadelphia: John Benjamins, Studies in Language
Companion Series,

Hansen-Schirra, S., Neumann, S. & Steiner, E. 2007. Cohesion and Explicitation in an English-
German Translation Corpus. In: Languages in Contrast 7(2): 241-265.

Miestamo, Matti; Kaius, Sinnemki; and Karlsson, Fred. eds 2008. Language Complexity. Typology,
contact, change. Amsterdam/ Philadelphia: John Benjamins

Siemund, P. & N. Kintana (eds.). 2008. Language contact and contact languages. Amsterdam:
Benjamins (Hamburg Studies in Multilingualism Vol. 7).

Steiner, Erich. 2008. Empirical studies of translations as a mode of language contact - explicitness
of lexicogrammatical encoding as a relevant dimension. in: Siemund, Peter and Kintana, Noemi. eds.
2008. Language contact and contact languages. Amsterdam: John Benjamins (Hamburg Studies in
Multilingualism Vol. 7). pp. 317-346
Teich, E. 2003. Cross-linguistic variation in system and text. A methodology for the investigation of
translations and comparable texts. Berlin, New York: de Gruyter.

Thomason, S. G. 2001. Language Contact. An Introduction. Washington D.C.: Georgetown University
Press.

Vela, Mihaela, Silvia Hansen-Schirra, and Stella Neumann 2007 Querying multi-layer annotation and
alignment in translation corpora. In Proceedings of the Corpus Linguistics Conference CL 2007,
Birmingham, UK, 27-30 July 2007, Matthew Davies, Paul Rayson, Susan Hunston, and Pernilla
Danielsson (eds.). http://ucrel.lancs.ac.uk/publications/CL2007/paper/97_Paper.pdf.
What Can Contrastive Linguistics Tell Us about
Translating Discourse Structure?
Irn Korzen, Morten Gylling
Copenhagen Business School
Dalgas Have 15, DK-2000 Frederiksberg, Denmark
E-mail: [email protected], [email protected]
Abstract
This paper argues that translators can greatly benefit from contrastive studies of discourse structure. Cross-linguistic studies of Italian
and Danish point to significant typological differences in information packaging in the two languages, especially in their use of
deverbalisation. Italian sentences tend to include a larger number of Elementary Discourse Units (EDUs), especially propositions,
than Danish. A higher percentage of these is rhetorically backgrounded by means of non-finite and nominalised predicates. Danish
text structure, on the other hand, is more informationally linear and characterised by a higher number of finite verbs and topic shifts.
These typological differences are transferred into three simple translation rules concerning 1) the number of EDUs, 2) the rhetorical
structure, and 3) the textualisation of rhetorical satellites.
Keywords: discourse structure, information packaging, textualisation, deverbalisation, translation strategies.
1. Introduction
Over the last decades, Contrastive Linguistics and
Translation Studies have experienced a veritable explo-
sion of interest and attention from scholars in different
fields, but the linguistic focus of attention has typically
been confined to lexical and syntactic levels. Contrastive
studies on discourse structure and intersentential rela-
tions, on the other hand, are much less frequent. For
instance, there are extremely few cross-linguistic textual
resources annotated for discourse. According to Webber,
Egg and Kordoni (2010), they are limited to the ones
found in the Copenhagen Dependency Treebanks (CDT),
which cover five different Germanic and Romance lan-
guages: Danish, English, German, Italian, and Spanish.
All CDT texts are annotated for four different linguistic
layers (apart from part-of-speech): syntax, discourse,
anaphora and morphology, see Buch-Kromann et al.
(2010).
The research we shall present in this paper is based partly
on our work with the CDT and partly on other resources,
and we shall focus on two phenomena related to the
information and discourse structures of texts, namely on
informational density, i.e. the amount of information per
sentence, and on text complexity, here defined as the
degree of subordination of the text segments that the
Rhetorical Structure Theory labels as rhetorical satel-
lites (Mann & Thompson, 1987; Mann, Matthiessen &
Thompson, 1992; Matthiessen & Thompson, 1988 and
later work). Like other scholars, such as Asher and Vieu
(2005), we consider these phenomena part of the in-
formation packaging of a text, a term suggested by
Chafe (1976) and later used, especially in connection
with given vs. new entities and definiteness, e.g. by Clark
and Haviland (1977), Prince (1984) and Vallduvi and
Engdahl (1996).
Other cross-linguistic surveys on information packaging
have been conducted e.g. by Fabricius-Hansen (1996;
1999), Ramm and Fabricius-Hansen (2005) and Behrens,
Solfjeld and Fabricius-Hansen (2010), who investigate
English, German and Norwegian, i.e. three Germanic
languages. On information density and explicitness in
English-German translations, see Hansen-Schirra, Neu-
mann and Steiner (2007). Alves et al. (2010) examine
particularly grammatical shifts, e.g. between finite verbs
and nominalisations, in the translation process between
English and German.
In this paper, we compare two languages of different
language families, viz. Danish and Italian, a Scandina-
vian (Germanic subgroup) and Romance language re-
spectively. Our results regarding Danish confirm the ones
obtained by the first mentioned scholars for Norwegian,
whereas their findings on English and German are closer
to our results concerning Italian. On the other hand, the
Italian features presented in the following, are found also
in other Romance languages, for which reason we con-
sider it justified to talk about general typological differ-
ences between Scandinavian and Romance languages,
ceteris paribus, with English and German somewhere in
between.
The paper is structured as follows: In section 2, we ex-
amine an Italian and Danish corpus of argumentative
texts with regard to informational density, measured as
the number of words and Elementary Discourse Units
(EDUs, cf. Carlson and Marcu, 2001) per sentence. In
section 3, we look at text complexity and the textualisa-
tion of rhetorical satellites, and in section 4, we formulate
our findings as a few relatively simple rules for (human
as well as machine) translators that work with Scandi-
navian and Romance languages.
2. Information density
2.1. Sentence length
Differences in discourse structure show themselves in
many ways, one of which is the simple sentence length,
measured as words per sentence
1
. In this context, we used
the parallel Europarl corpus, an open source corpus
compiled by Koehn (2005). Europarl is a very large
multilingual corpus (55 million words) with source and
target texts covering all the official languages of the
European Union. In fact, the corpus was designed to train
and evaluate statistical machine translation, but it can, as
we shall see, also be used for other types of
cross-linguistic studies. The Europarl texts, which are
mainly argumentative (see van Halteren (2008) for a
discussion of this), consist of speeches made by the
members of the European Parliament from 1996 to 2010,
and most of the speeches (88 %) have been tagged with a
language attribute indicating the native language (L1) of
the speaker. We created a Perl script
2
that extracted all

1
We are aware of the many reservations to be made when
conducting linguistic measurements in this way, but subject to
space limitations we cannot go into detail here. However, we
feel that the statistical results cited in this section are convincing
enough to be taken into account and used as a first indication of
profound typological differences between the two languages
analysed.
2
We thank our colleague Daniel Hardt for his help in this
matter.
Danish and Italian L1 text from the entire corpus and
calculated the average sentence length of all texts. In this
context, a sentence is defined as a text segment marked
by a full stop, a question mark, or an exclamation mark.
We then compared the results with those of the texts
translated from one of the two languages into the other
(L2). Thus, in Table 1, Italian L2 texts are translated
from Danish into Italian and Danish L2 texts from
Italian into Danish.

Table 1: Sentence length in L1 and L2 Europarl texts.
We chose Europarl as the empirical basis for a statistical
count because it contains both parallel (L1 L2) texts and
comparable texts, i.e. L1 texts created in different lan-
guages but dealing with similar topics and produced in
similar situations and genres for similar targets. Whereas
parallel texts are clearly best suited for projects aimed e.g.
at improving machine translation (such as the previously
mentioned CDT) because they permit L1L2 text
alignment and evaluation, comparable texts are generally
best suited as the empirical basis for descriptive, typo-
logical comparisons like the present one. In such cases,
parallel texts are inappropriate because the filter of the
translator and his/her translation strategies get in the
way, and L2 texts risk ending up with a text structure too
similar to that of the L1. See McEnery and Wilson (2001)
and Baroni and Bernardini (2006) for discussions in this
regard.
As the upper part of Table 1 shows, there is a consider-
able difference in average sentence length between the
Italian L1 and Danish L1 Europarl texts, a difference
amounting to 10.86 words per sentence or 31.06 %.
However, the lower part of Table 1 confirms the problem
just mentioned regarding translated L2 texts. As far as
sentence length goes, EU translators seem to stick very
much to the structure of the L1 text: the Danish L2 texts
(translated from Italian) are 24.82 % longer than the
Danish L1 texts, while the Italian L2 texts (translated
from Danish) are 35.64 % shorter compared to the Italian
Language Words Sentences Words
/sentence
Italian L1
Danish L1
1,657,592
546,425
47,405
22,668
34.97
24.10
Italian L2
Danish L2
571,115
1,845,951
22,154
57,574
25.78
32.06
L1 texts. When it comes to sentence length, these L2
texts are clearly influenced by the L1 structure.
2.2 Elementary Discourse Units
At this point we shall return to the concept of informa-
tional density and define a little more precisely its
application in our project. In order to determine the
purpose that the more numerous words in the Italian
sentences serve, we then counted the number of Ele-
mentary Discourse Units (EDUs) textualised in each
sentence, using Carlson and Marcus (2001) classifica-
tion. This can be a very time-consuming task, since no
parser has been trained to do this convincingly, and we
therefore randomly selected a limited part of the Europarl
corpus consisting of 7,500 words in each language. We
confined ourselves to texts of 200-600 words, and we
ended up with a subcorpus in each language consisting of
25 texts of an average length of 300 words each. All texts
were manually checked with regard to text type (argu-
mentative), speaker (a certain number of different
speakers were required), and date (so that not all text
were speeches from the same period).
We discovered a very clear tendency towards a higher
number of EDUs in the Italian sentences than in the
Danish ones. A statistical count showed that 27.3 % of the
Italian sentences contained five or more EDUs. By
comparison, only 9.8 % of the Danish sentences con-
tained five or more EDUs.
We also discovered considerable differences in the
number of coordinate vs. subordinate clauses. Finite
coordinate clauses amounted to 27.2 % of all clauses in
the Danish texts, but only to 17.9 % in the Italian texts.
Thus, 82.1 % of the Italian clauses were subordinate as
opposed to 72.8 % of the Danish clauses. This may not
seem a huge discrepancy, but if we examine in detail the
distribution of the subordinate clauses, we encounter
considerable differences, cf. Table 2:

With
connec-
tives
Rela-
tive
clauses
Attri-
bution
Subordi-
nate non-
finite
clauses
IT 22.4 % 40.3 % 13.1 % 24.2 %
DA 25.8 % 40.3 % 22.5 % 11.4 %
Table 2: Distribution of EDUs in subordinate clauses in a
Europarl subcorpus
The use of connectives (or discourse cues in the RST
terminology) and the frequency of relative clauses are
more or less equal in the two languages, whereas Danish
seems to use attribution more often. In our opinion, this
difference should be seen not just as a particular linguis-
tic tendency among Danish parliamentarians, but also as
a stylistic feature used to add particular pragmatic values
to the argument put forward, a point we shall elaborate in
the full version of this paper.
However, the most interesting difference lies in the
distribution of non-finite clauses. As Table 2 shows, these
occur more than twice as often in the Italian texts as in the
Danish ones. Furthermore (not shown in Table 2), Italian
uses the whole range of non-finite verb forms (gerund,
participles, infinitives and normalisations) much more
regularly, whereas Danish mostly confines itself to the
use of infinitives (the gerund does not exist in Danish).
3. Text complexity
The differences in sentence length seen in Table 1 also
have an impact on the distribution of EDUs. Many EDUs
correspond to propositions, and what may be textualised
as one multi-propositional sentence in a Romance lan-
guage may very well correspond to two or more sen-
tences in Scandinavian. In a sequence of propositional
EDUs, P1 + P2, such as the following:
P1: arrive (John, in town); P2: go (John, home)
P1 can be textualised in different ways (possibly with
added adjuncts or other linguistic material), as shown in
the Deverbalisation Scale in Table 3
3
:

P1 textualised as Textualisation P1 + P2
a. an independent
sentence
John arrived late in town.
He went straight home.
b. a main clause,
part of sentence
John arrived late in town
and he went straight
home.
c. a subordinate
finite clause

Since John arrived late in
town, he went straight
home.
d. a subordinate
non-finite
clause
Having arrived late in
town, John went straight
home.
e. a nominalisa-
tion
Upon his arrival in town,
John went straight home
Table 3: Examples of textualisation of EDUs.

3
The scale is based on Hopper and Thompson (1984), Lehmann
(1988), and Korzen (1998; 2007; 2009).
The deverbalisation of P1 increases from (a/b) to (e)
together with its integration and absorption into the
matrix clause. Whereas the finite verb in a main clause,
such as (a/b), has its full (language specific) range of
grammatico-semantic values and the clause its full range
of pragmatic-illocutionary possibilities, these values are
gradually reduced or lost in the textualisations further
down the scale. The verb in the subordinate finite clause
(c) loses its independent tense, mood and illocution; these
values will be determined and/or expressed by the matrix
clause. The non-finite verb in (d) loses all temporal,
modal, and aspectual values and cannot render explicit its
subject (see however note 4), and the nominalisation (e)
is completely integrated in the matrix clause as a second
order entity; its valency complements (here his) are
syntactically reduced to secondary positions or simply
left out.
The further down the scale a proposition is textualised,
the fewer grammatico-semantic and pragmatic features
are expressed by the verb, i.e. the more the proposition is
deverbalised, and the more it is semantically and rhe-
torically subordinated and incorporated into the matrix
clause. In the case of non-finite and nominalised verbs,
(d/e), features such as subject, tense, mood, aspect, and
illocution are entirely interpreted on the basis of the
matrix clause
4
. Therefore, a non-finite or nominalised
structure is entirely pragmatically and semantically
dependent on the matrix clause, and such structures
express a particularly strong rhetorical backgrounding (or
explicit satellite status) of the proposition in question.
Furthermore, the lack of subject generally entails an
inherent topic continuity (a topic shift typically requires a
finite verb with an explicit subject), which means that the
situation or event in question is evaluated and interpreted
as related and less important to the on-going topic than
the situation or event of the matrix clause, textualised
with a finite predicate.
Cross-linguistic surveys show that textualisation at the
levels (d/e) is much more frequent in the Romance lan-
guages than in the Scandinavian ones which show a very

4
We here ignore the subject of the so-called absolute con-
structions consisting of a participle or gerund + a subject
different from the subject of the main verb, e.g. Morto il padre,
Luca part per Roma The father [having] died, Luca left for
Rome, as well as the accusative with infinitive constructions
(Ho visto Luca arrivare I saw Luca arrive). In nominalised
verb forms the subject may appear as a secondary valency
complement, e.g. Larrivo di Luca Lucas arrival.
clear predilection for finite verbs and textualisation at the
levels (a/b/c). These tendencies are not limited to par-
ticular text types or genres, such as the (generally argu-
mentative) Europarl texts. Table 4 indicates the per-
centage of propositions textualised with finite, non-finite,
and nominalised verb forms in a number of comparable
texts belonging to five different text types and genres.
The numbers clearly indicate statistically significant
differences between Italian and Danish text structure
regarding finite and non-finite verb frequency, inde-
pendently of text type or genre.

Verb forms (%)
Fi-
nite
Non-
finite
Nomi-
nalised
a. Legal texts IT 43.9 24.2 31.9
DA 56.4 10.2 33.4
b. Technical
texts
IT 47.5 26.8 25.9
DA 80.7 9.5 9.9
c. News-
groups
IT 61.1 23.1 15.8
DA 75.8 11.5 12.7
d. Websites IT 54 27 19
DA 84 8 8
e. Written
narratives
IT 52.8 44.2 3.0
DA 88.0 12.0 0.01
f. Oral nar-
ratives
IT 72.8 27.1 0.1
DA 93.6 6.4 0
Table 4: Verb forms in different text types
5

As stated above, non-finite and nominalised structures
explicitly express the satellite status of the proposition in
question. Generally but not necessarily this is true
also of subordinate adverbial clauses, such as (c) in Table
3. On the other hand, the structures in (a/b) of Table 3 are
in themselves ambiguous as to mono- or multinuclear
interpretation. However, as is well known, the structure
in (b), the syndetic coordination with the connective and
(and cross-linguistic counterparts), often contains a P1
with satellite status, in Table 3 expressing the cause of P2.
We shall elaborate also on this issue in the full version of
our paper
6
.
4. Perspectives for translation
The differences described above entail a generally higher

5
Precise references will appear in the full version of our paper.
6
Important cross-linguistic studies on and and counterparts are
found e.g. in Ramm and Fabricius-Hansen (2005), Behrens and
Fabricius-Hansen (2010) and Skytte (2000: 652-660).
structural complexity in Italian (and Romance in general)
than in Danish (and Scandinavian in general). Romance
sentences tend to be longer and to include more proposi-
tions, of which a higher number is backgrounded by
means of non-finite and nominalised predicates. This
results in a multi-layered and hierarchical information
structure, characterised by a high degree of topic conti-
nuity, in which the various events are evaluated with
respect to their importance to the on-going topic.
On the other hand, Scandinavian text structure tends to be
more informationally linear and characterised by a higher
degree of topic shifts. Each sentence holds fewer EDUs,
and different events tend to be textualised more chrono-
logically one after the other and with finite verb forms
that permit subject/topic changes.
The results of our study can be transferred into three main
rules concerning translations from a Romance to a
Scandinavian language or vice versa. The rules regard:
the number of EDUs per sentence: ceteris paribus,
there are more EDUs and a higher informational
density in Romance than in Scandinavian sentences;
the textualisation of rhetorical structure: there is a
higher tendency in Romance than in Scandinavian to
distinguish morpho-syntactically between rhetorical
nuclei and satellites;
the textualisation of rhetorical satellites: there is a
tendency to textualise satellites at lower levels of the
deverbalisation scale (cf. Table 3) in Romance than in
Scandinavian.
Naturally, also phenomena such as e.g. the linguistic
register and diamesic dimension (e.g. written vs. spoken
text) come into play. The higher the register, the more
distinct the mentioned cross-linguistic differences. Oral
Italian textualisation and some web variants (such as
newgroups, see Table 4) are characterised by a certain
structural levelling and are therefore closer to typical
Danish textualisation.
5. Conclusion
It is well known that a good translation does not (gener-
ally, at least) follow the source text word for word. But
especially between language families, a good translation
does not often follow the source text sentence for sen-
tence, either. Profound typological differences such as
those regarding informational density and text complex-
ity must be taken into account, and contrastive studies on
discourse structure provide necessary and highly useful
linguistic insights for human as well as machine trans-
lators.
The results of our study presented above and in the full
version of our paper will hopefully provide us with
more precise and detailed knowledge of typological
differences between Romance and Scandinavian dis-
course structure, differences which are of importance
also for syntax (e.g. in the choice of subject type and
voice) and for anaphora (e.g. null-forms vs. pronominal
forms), phenomena that we will develop in future work.
6. Acknowledgements
This work was supported by a grant from The Danish
Council for Independent Research Humanities (FKK).
We thank the two anonymous reviewers for their useful
comments and references.
7. References
F. Alves, A. Pagano, S. Neumann, E. Steiner, and S.
Hansen-Schirra (2010): Translation Units and Gram-
matical Shifts. Towards an Integration of Product- and
Process-based Translation Research. In G.M. Shreve
and E. Angelone (Eds). Translation and Cognition.
John Benjamins, Amsterdam/Philadelphia, pp.
109142.
N. Asher and L. Vieu (2005): Subordinating and Coor-
dinating Discourse Relations. Lingua 115, pp.
591610.
M. Baroni and S. Bernardini (2006): A New Approach to
the Study of Translationese: Machine-learning the
Difference between Original and Translated Text. In
Literary and Linguistic Computing, 21,3, pp. 259274.
B. Behrens and C. Fabricius-Hansen (2010). The Rela-
tion Accompanying Circumstance Across Languages:
Conflict between Linguistic Expression and Discourse
Subordination? In D. Shu and K. Turner (eds.). Con-
trasting Meaning in Languages of the East and West.
Contemporary Studies in Descriptive Linguistics, 14.
Oxford et al.: Peter Lang, pp. 531552.
B. Behrens, K. Solfjeld and C. Fabricius-Hansen (2010):
The Relation Accompanying Circumstance Across
Languages: Conflict between Linguistic Expression
and Discourse Subordination? In D. Shu and K. Turner
(Eds.). Contrasting Meaning in Languages of the East
and West. Contemporary Studies in Descriptive Lin-
guistics, 14. Oxford et al.: Peter Lang, pp. 531552.
M. Buch-Kromann et al. (2010): The Inventory of Lin-
guistic Relations used in the Copenhagen Dependency
Treebanks. Copenhagen Business School,
http://copenhagen-dependency-treebank.googlecode.c
om/svn/trunk/manual/cdt-manual.pdf.
L. Carlson and D. Marcu (2001): Discourse Tagging
Reference Manual. ISI Technical Report, ISI-TR-545.
W.L. Chafe (1976): Givenness, Contrastiveness, Defi-
niteness, Subjects, Topics, and Point of View. In: Li,
Charles N. (Ed.). Subject and Topic. Academic Press,
New York/San Francisco/London, pp. 2555.
H.H. Clark and S.E. Haviland (1977): Comprehension
and the Given-new Contract. In Discourse Production
and Comprehension, R.O. Freedle (Ed.), Hillsdale, NJ:
Erlbau, pp. 140.
C. Fabricius-Hansen (1996): Informational Density: a
Problem for Translation and Translation Theory. Lin-
guistics 34, pp. 521565.
C. Fabricius-Hansen (1999): Information Packaging and
Translation. Aspects of Translational Sentence Split-
ting (German - English/Norwegian). In
Sprach-spezifische Aspekte der Informationsverteilung,
M. Doherty (Ed.). Akademie-Verlag, Berlin, pp.
175213.
H. van Halteren (2008): Source Language Markers in
EUROPARL Translations. In Proceedings of the 22nd
International Conference on Computational Linguis-
tics (Coling 2008). Manchester, August 2008, pp.
937944.
S. Hansen-Schirra, S. Neumann and E. Steiner (2007):
Cohesive Explicitness and Explicitation in an Eng-
lish-German Translation Corpus. Languages in Con-
trast 7(2), pp. 241265.
P. J. Hopper and S. A. Thompson (1984): The Discourse
Basis for Lexical Categories in Universal Grammar.
Language, 60(4), pp. 703752.
P. Koehn (2005): Europarl: A Parallel Corpus for Statis-
tical Machine Translation. MT Summit.
I. Korzen (1998): On the Grammaticalisation of Rhe-
torical Satellites. A Comparative Study on Italian and
Danish. In I. Korzen and M. Herslund (Eds.). Clause
Combining and Text Structure. Studies in Language,
22, Copenhagen, pp. 6586.
I. Korzen (2007): Linguistic Typology, Text Structure
and Appositions. In I. Korzen, M. Lambert, and H.
Vassiliadou. Langues dEurope, lEurope des langues.
Croisement Linguistiques. Scolia, 22, pp. 2142.
I. Korzen (2009): Struttura testuale e anafora evolutiva:
tipologia romanza e tipologia germanica. In I. Korzen
and C. Lavinio (Eds). Lingue, culture e testi
istituzionali. Firenze: Franco Cesati, pp. 3360.
C. Lehmann (1988): Towards a Typology of Clause
Linkage. In J. Haiman and S. A. Thompson (Eds.).
Clause Combining in Grammar and Discourse. John
Benjamins, Amsterdam/Philadelphia, pp. 181225.
W.C. Mann, C. Matthiessen and S.A. Thompson (1992):
Rhetorical Structure Theory and Text Analysis. In W.C.
Mann and S.A. Thompson (Eds). Discourse Descrip-
tion. Diverse Linguistic Analyses of a Fund-raising
text. John Benjamins, Amsterdam/Philadelphia, pp.
3978.
W.C. Mann and S.A. Thompson (1987): Rhetorical
Structure Theory. A Theory of Text Organization. ISI,
Los Angeles, CA, ISI/RS-87-190, pp. 181.
C. Matthiessen and S.A. Thompson (1988): The Struc-
ture of Discourse and Subordination. In J. Haiman
and S.A. Thompson (Eds). Clause Combining in
Grammar and Discourse. John Benjamins, Amster-
dam/Philadelphia, pp. 275329.
T. McEnery and A. Wilson (2001): Corpus Linguistics:
an Introduction. 2
nd
Edition. Edinburgh University
Press, Edinburgh.
E.F. Prince (1984): Topicalization and Left-dislocation: a
Functional Analysis. In Discourses in Reading and
Linguistics, Sheila J. White and Virginia Teller (Eds.).
Annals of the New York Academy of Sciences 433,
Academy of Sciences, New York, pp. 213225.
W. Ramm and C. Fabricius-Hansen (2005): Coordination
and Discourse-structural Salience from a
Cross-linguistic Perspective. SPRIKreports 30.
G.. Skytte (2000): Konnexion og diskursmarkering. In G.
Skytte and I. Korzen. Italienskdansk sprogbrug i
komparativt perspektiv. Reference, konnexion og dis-
kursmarkering, Samfundslitteratur, Copenhagen, pp.
621793.
E. Vallduv and E. Engdahl (1996): The Linguistic Re-
alization of Information Packaging. Linguistics 34, pp.
459519.
B. Webber, M. Egg, and V. Kordoni (2010): Discourse
Structure and Language Technology. Natural Lan-
guage Engineering, 1(1), pp. 149.
1
GSCL 2011 Workshop Contrastive Linguistics Translation Studies Machine Translation what
can we learn from each other?

An analysis of translational complexity in two text types.
Martha Thunes, University of Bergen.
[email protected]

Key words: automatisation of translation, English-Norwegian parallel text, translational complexity,
text types.

Abstract
This paper is based on the study presented in Thunes (2011), where a selection of English-Norwegian
parallel texts have been analysed in order to discuss two primary research questions: firstly, to what
extent is it possible to automatise, or compute, the actual translation relation found in the investigated
parallel texts, and, secondly, is there a difference in the degree of translational complexity between the
two text types, law and fiction, included in the empirical material?
By automatisation I here understand the generation of translations with no human intervention,
and I assume an approach to machine translation based on linguistic information. In the analysed texts
the translations have been produced manually; this is not a study of output produced by machine
translation systems, and the automatisation issue is not discussed with reference to any particular
translation algorithm or system architecture. Rather, it is related to the assumption that there is a
translational relation between the inventories of simple and complex linguistic signs in two languages
which is predictable, and hence computable, from information about source and target language
systems, and about how the systems correspond. Thus, computable translations are linguistically
predictable, i.e. predictable from the linguistic information coded in the source text, together with
given, general information about the two languages and their interrelations. Further, non-computable
translations are correspondences where it is not possible to predict the target expression from the
information encoded in the source expression, together with given, general information about SL and
TL and their interrelations. Non-computable translations require access to additional information
sources, such as various kinds of general or task-specific extra-linguistic information, or task-specific
linguistic information from the context surrounding the source expression.
In order to answer the research questions, a measurement of translational complexity is applied to
the analysed texts. The degree of translational complexity in a given translation task is understood as a
factor determined by the types and amounts of information needed to solve the task, as well as by the
accessibility of these information sources, and the effort required when they are processed.
For the purpose of measuring the complexity of the relation between a source text unit and its
target correspondent, I apply a set of four correspondence types, organised in a hierarchy reflecting
divisions between different linguistic levels, along with a gradual increase in the degree of
translational complexity. In type 1, the least complex type, the corresponding strings are
pragmatically, semantically, and syntactically equivalent, down to the level of the sequence of word
forms. In type 2 correspondences, source and target string are pragmatically and semantically
2
equivalent, and equivalent with respect to syntactic functions, but there is at least one mismatch in the
sequence of constituents or in the use of grammatical form words. Within type 3, source and target
string are pragmatically and semantically equivalent, but there is at least one structural difference
violating syntactic functional equivalence between the strings. In type 4, there is at least one
linguistically non-predictable, semantic discrepancy between source and target string, and pragmatic
equivalence may, or may not, hold. Thus, the type hierarchy is characterised by an increase with
respect to linguistic divergence between source and target string, and by an increase in the need for
information and in the amount of effort required to translate, i.e. an increase in the degree of
translational complexity. Correspondences of types 13 constitute the domain of linguistically
predictable, or computable, translations, whereas type 4 correspondences belong to the non-
predictable, or non-computable, domain, where semantic equivalence is not fulfilled.
This study applies a strictly product-oriented approach to complexity in translation. The four types
of translational correspondences should not be seen as translation methods or strategies, but as
descriptions of correspondence relations between given source text units and their existing
translations. The empirical analysis of translational correspondences does not aim to study what kinds
of knowledge a translator has actually used in order to produce a chosen target expression. Rather, it
focusses on the kinds of information about source text expressions that are needed in order to produce
the translations.
The correspondence type hierarchy can be seen as a fairly general classification model for
translational correspondences. Its main principles were originally defined by Dyvik (1993), and
further articulated in Thunes (1998). The approach chosen for the present study is an adapted version
of the classification model defined by Thunes (1998). The model is also used as a framework for
contrastive language analysis in the studies presented by Hasselgrd (1996), Tucunduva (2007), Silva
(2008), and Azevedo (in progress).
In the present contribution, the empirical method involves extracting translationally corresponding
strings from parallel texts, and assigning one of the types defined by the correspondence hierarchy to
each recorded string pair. The finite clause is chosen as the primary unit of analysis, and the main
syntactic types among the recorded data are matrix sentences, finite subclauses, and lexical phrases
with finite clause(s) as syntactic complement. Since syntactically dependent constructions like finite
subclauses occur as translational units, the data include nested correspondences where a superordinate
string pair contains one or more embedded string pairs. The assignment of correspondence type to
string pairs is an elimination procedure where we start by testing each correspondence for the lowest
type and then move upwards in the hierarchy if the test fails. The analysis is thus an evaluation of the
degree to which linguistic matching relations hold in each string pair. In cases of nested string pairs,
embedded units are treated as opaque items, and the classification of a superordinate correspondence
is done independently of the degree of complexity in embedded string pairs. Otherwise, it is a general
principle that a string pair is assigned the correspondence type of its most complex non-opaque
subpart.
The analysis is applied to running text, omitting no parts of it. Thus, the distribution of the four
types of translational correspondence within a set of data provides a measurement of the degree of
translational complexity in the parallel texts that the data are extracted from. The extraction and
3
classification of string pairs is done manually as it requires a bilingually competent human analyst.
The recorded data cover about 68 000 words, and are compiled from six different text pairs: two of
them are law texts; the remaining four are fiction texts. Comparable amounts of text are included for
each text type, and both directions of translation are covered.
Since the scope of the investigation is limited, the results do not provide a sufficient basis for
generalisations about the degree of translational complexity in the chosen text types and in the
language pair English-Norwegian. Concerning the automatisation issue, the complexity measurement
across the entire collection of data shows that, in terms of string lengths, as little as 44,8% of all
recorded string pairs are classified as computable translational correspondences, i.e. as type 1, 2, or 3,
and non-computable string pairs of type 4 constitute a majority (55,2%) of the compiled data. As
regards the text type issue, the proportion of computable correspondences is on average 50,2% in the
law data, and 39,6% in fiction.
In order to discuss whether it would be fruitful to apply automatic translation to the selected texts,
I have considered the workload potentially involved in correcting assumed machine output, and in this
respect the difference in restrictedness between the two text types is relevant: law text is strongly
norm-governed in a way that fiction text is not. Among the recorded data, I have analysed a set of
phenomena that have been identified as recurrent semantic deviations between translationally
corresponding units, and this shows that within the non-computable correspondences, the frequency of
cases exhibiting only one minimal semantic deviation between source and target string is considerably
higher among the data extracted from the law texts than among those recorded from fiction. Such
cases can be regarded as minimally non-computable string pairs. Among the law data, as much as
45,7% of the correspondences classified as type 4 are minimally non-computable string pairs, whereas
among the fiction data, only 10,5% of the compiled type 4 correspondences are minimal ones. In
minimally non-computable correspondences, I assume that only a small effort would be required in
order to revise an automatically generated target expression according to the standard of manual
translation.
For this reason I tentatively regard the investigated pairs of law texts as representing a text type
where tools for automatic translation may be helpful, if the effort required by post-editing is smaller
than that of manual translation. This is possibly the case in one of the law text pairs, where 60,9% of
the data involve computable translation tasks. In the other pair of law texts the corresponding figure is
merely 38,8%, and the potential helpfulness of automatisation would be even more strongly
determined by the edit cost. That text might be a task for computer-aided translation, rather than for
MT. As regards the investigated fiction texts, it appears likely that post-editing of automatically
generated translations would be laborious and not cost effective, even in the case of one text pair
showing a relatively low degree of translational complexity. In the analysed pairs of fiction texts, there
is a clear tendency that non-computable correspondences exhibit several semantic deviations between
the corresponding strings. Hence, I expect that the workload involved in correcting potential machine
output would be heavy, and I agree with the common view that the translation of fiction is not a task
for MT.
This study is intended to be of relevance to rule-based MT since the chosen analytical framework
relies on assumptions about how translations can be computed on the basis of formal descriptions of
4
source and target language systems and their interrelations. However, I assume that the general issue
of computability underlying this approach likewise applies to statistical machine translation, which is
also dependent on the accessibility of relevant and sufficient information in order to predict correct
target expressions from available translational correspondences.
In my view, the framework applied in this study could be used as a diagnostic tool for the
feasibility of machine translation in relation to specific text types. That is, by applying the method to
limited selections of parallel texts of the same type, it would be possible to estimate to what extent the
target text could be generated automatically. If the proportion of assumed computable
correspondences would exceed a chosen threshold, it might be worthwhile to tune an MT system for
the given language pair to the text type in question. Moreover, in order to estimate the editing distance
between potential machine output and a given target text norm, it would be interesting to identify the
proportion of minimal type 4 correspondences in a given body of parallel texts. Thus, it would be
fruitful to extend the classification model by integrating a fifth correspondence type to be assigned to
minimally non-computable string pairs.

References:
Azevedo, Flvia. In progress. Investigating the problem of codifying linguistic knowledge in two
translations of Shakespeares sonnets: a corpus-based study. Doctoral dissertation. Federal
University of Santa Catarina, Florianpolis.
Aijmer, Karin, Bengt Altenberg, and Mats Johansson (eds). 1996. Languages in Contrast. Papers from
a Symposium on Text-based Cross-linguistic Studies, Lund 45 March 1994. Lund Studies in
English 88. Lund: Lund University Press.
Dyvik, Helge. 1993. Text Pair Mapper. Unpublished manuscript. University of Bergen.
Hasselgrd, Hilde. 1996. Some methodological issues in a contrastive study of word order in English
and Norwegian. In: Aijmer et al. (eds), 1996, 113126.
Johansson, Stig and Signe Oksefjell (eds). 1998. Corpora and Cross-linguistic Research: Theory,
Method, and Case Studies. Language and Computers: Studies in Practical Linguistics 24.
Amsterdam and Atlanta, GA: Rodopi.
Silva, Norma Andrade da. 2008. Anlise da traduo do item lexical evidence para o portugus com
base em um corpus jurdico. Master's thesis. Federal University of Santa Catarina, Florianpolis.
Thunes, Martha. 1998. Classifying translational correspondences. In: Johansson and Oksefjell (eds),
1998, 2550.
Thunes, Martha. 2011. Complexity in Translation. An English-Norwegian Study of Two Text Types.
Doctoral dissertation. University of Bergen.
Tucunduva, Camila de Andrade. 2007. Translating completeness: a corpus-based approach. Master's
thesis. Federal University of Santa Catarina, Florianpolis.

Web-based contrastive comparison of Age-Related Temporal Phrases
in Spanish and French
Sofa N. Galicia-Haro
1
and Alexander Gelbukh
2
1
Faculty oI Sciences UNAM ,Mexico
Depto. de Matematicas, Ciudad Universitaria
04510 Mexico, D. F.
2
Center Ior Computing Research, National Polytechnic Institute, Mexico
Juan de Dios Batiz, esq. con Miguel Othon de Mendizabal
07738 Mexico, D. F.
E-mail:
1
snghIciencias.unam.mx,
2
gelbukhcic.ipn.mx; www.Gelbukh.com
Introduction
Some words or whole sequences oI words in a text are temporal expressions: Ior example,
yesterday, Monday 12, two months, about a year and a half; each reIers to a certain period oI
time. Such words or sequences oI words mainly share a noun or an adverb oI time: yesterday,
month, year. This causes a problem in automatically deciding whether a word or a sequence is a
temporal expression. It is an important part oI many natural language processing applications,
such as question answering, machine translation, inIormation retrieval, inIormation extraction,
text mining, etc., where robust handling oI temporal expressions is necessary.
Automatic recognition oI expressions oI time was introduced in the Named Entity
Recognition task oI the Message Understanding ConIerences
1
where temporal entities were
tagged as 'TIMEX. Since then, researchers have developed temporal annotation schemes; Ior
example, |2| and |6| Ior English, |1| Ior French, and |7| Ior Spanish.
In this work, we analyzed Spanish temporal expressions that were not considered in those
annotation guidelines. These phrases are recognized by an initial adverb: Ior example, around,
still; they end with the noun oI time, e.g. year, and they describe a person`s age. For example:
an a sus 50 aos 'although he is 50 years old, ahora a mis 23 aos 'now I am 23 years old,
alrededor de los 55 aos 'around 55 years old. These phrases are very interesting since the
adverb reinIorce the meaning oI time.
We can observe the relation between the groups oI words in the Iollowing Spanish examples:
1. A sus 30 aos Juan se comporta como nio
2. An a sus 30 aos Juan se comporta como nio
3. Hoy a sus 30 aos Juan se comporta como nio
The sentences describe the same main Iact: John, who is 30 years old, behaves like a child,
but they tell us something else when we introduce a modiIier (an 'still, hoy 'today) in each
one: they argue Ior diIIerent conclusions.
Even at 30 years old, John behaves like a child in spite oI his age he behaves as iI he were
a child
Today, at 30 years old, John behaves like a child today he behaves like a child
The adverbs 'even and 'today make such conclusions obligatory and reinIorce the meaning
oI time in diIIerent Iorms. Both adverbs are related to time duration; one strict reading reIers to
24 hours and the other to a longer period oI time, but they also imply a direct judgment on the
perception oI the speaker, on the behavior oI the subject or both.
Owing to the constructions similarity in Spanish and French we supposed that these phrases
oI interest would be similar in both languages but we Iound that the French translation
2
'encore
ses 38 annes Ior the phrase an a sus 38 aos is not common in French.
In this work, we develop a web-based analysis carried out to compare such Spanish temporal
expressions with age-related temporal phrases in French with the objective oI determining
appropriate annotations Ior marking up text and translations. First, we present the characteristics
oI the Spanish phrases and the method we applied to obtain the materials Ior the comparison.

1
http://timexportal.wikidot.com/timexmuc6
2
http://www.online-translator.com/DeIault.aspx/Text
Then we describe the French phrases obtained with the same method. Finally we present the
comparison oI such phrases and the application oI the results to annotation and machine
translation.
Age-Related Temporal Expressions in Spanish
Usually people's age is described by Spanish temporal expressions including the time noun aos
'years. They can be recognized in the Iollowing ways: 5 aos de edad (5 years old), Jorge, de
52 aos, entrenaba al Vitesse Arnhem (Jorge, a 52-year-old, was training Ior the Vitesse
Arnhem), la nia de 11 aos (the 11-year-old girl), falleci ayer a la edad de 95 aos (died
yesterday at 95 years old)
There are, however, other temporal expressions that describe people's age: Ior example, an a
sus 65 aos, lit 'still at his 65 years, de alrededor de 20 aos, lit. 'oI about 20 years. These
temporal phrases denote a point in the timeline oI a person; it could be a point in the timeline oI
the events related in the sentence or a point in a tangential timeline.
Material Acquisition (MA)
Texts and their exact translation, i.e. parallel corpora, are widely used mainly to train and
evaluate Machine Translation systems. They are also useIul in cross-linguistic analysis by
looking Ior translations oI a given construction into another language. We also intended to use
parallel corpora to make a better comparison oI age-related temporal phrases in Spanish and
French and Ior this purpose we examined Iree texts collections oI the European Commission.
Many oI them were Irom the law domain, however, and they had Iew examples oI the
expressions we were interested in. We decided then to use newspaper texts since they employ
Ireer construction oI sentences.
In previous work, a corpus-based analysis was carried out to determine the context oI such
Spanish temporal expressions Ior their automatic determination. Such a method allows the
manual selection oI examples representing what was considered to be a class: a diIIerent
combination oI an adverb and a preposition beIore the number oI years and then the retrieval oI
web examples Ior that class.
The method consists oI two steps. The Iirst one is the application oI a program to extract the
sentences matching the Iollowing pattern:
AdvTsomethingTimeN
where:
something corresponds to a sequence oI up to six words
3
without punctuation marks, verbs
or conjunctions
TimeN corresponds to ao, aos 'year, years
AdvT adverbs oI time, a collection oI 51 elements Irom a dictionary
4
We applied this step to a text collection compiled Irom a Mexican newspaper and Irom 27054
sentences we manually selected one arbitrary example representing a class, the Iive resulting
classes corresponding to an a, an con, actualmente de, alrededor de, ahora de.
The second step was intended to obtain a more representative group oI phrases since the
newspaper text collection contained a subset oI all possible temporal phrases expressing the age
oI people. We analyzed diverse methods to obtain a more representative group oI phrases and
we chose to look Ior examples on the Internet. This option allowed us to Iind phrases generated
by native speakers more quickly, including the commoner collocations. We realize that
searching the Internet has its drawbacks but we decided to do so on the basis that we did not
know how the results were classiIied |4|.
Many studies Iocused in having a corpus that modeled the whole language. To collect
inIormation Ior annotation and translation oI the phrases we were interested in, however, we
collected only a particular subset oI language that corresponded to them. Thus, the research we
report here reIers to a collection that has been skewed by design.

3
A larger number oI words does not guarantee any relation between the AdvT and the TimeN
4
DRAE, Real Academia Espaola. (1995): Diccionario de la Real Academia Espaola, 21 edicion (CD-ROM),
Espasa, Calpe.
2
The main idea oI obtaining more examples Irom the Internet was based on obtaining a Iew
examples Irom the newspaper texts (corresponding to the Iive classes mentioned above),
simpliIying them (eliminating determinants, adjectives, etc.) and searching Ior variants by
including Google`s asterisk Iacility |3|. For example: Ior the phrase an con sus jvenes 48 aos
the string when simpliIied becomes 'an con ao and the search is 'an con * aos using the
Google search engine tool limited to the Spanish language where the asterisk substitutes Ior the
eliminated words. Google returns hits where there is a string oI words initiated by 'an con
and then a sequence oI words, ending with 'aos Ior example: y el bachillerato en Lleida,
an con diecisis aos entr a trabajar de chico
The process was repeated several times until no new repeated phrases were obtained,
determining the sequences oI words that appeared with greater Irequency. AIter this compilation
oI examples, we manually selected 18 cases: Ior example, ahora a los NUM aos, actualmente
de unos NUM aos, where NUM treats numbers represented by digits or letters. We Iound that
some oI the 18 classes obtained Irom the Internet seem to preserve their meaning independently
oI the context and others require some Iorm oI words in context to denote the age oI a person.
The quantity oI pages automatically obtained was limited to 50, i.e. to obtain 500 snippets. For
each oI the 18 classes we manually analyzed the number oI examples that corresponded to a
person`s age.
Age-Related Temporal Expressions in French
We applied the previous method to 10234 sentences obtained Irom the Europarl corpus |5| that
we called EuropAns. The sentences were retrieved in response to a query on the noun ans
'years, since people's age is described by French temporal expressions including that noun.
To analyze the temporal phrases expressing age similarly to the Spanish phrases detailed
above we applied the Iirst step to the EuropAns. We obtained 257 sentences matching the
AdvTsomethingTimeN pattern where Adv corresponds to 57 elements. From them we
manually selected Iive classes to process the MA second step by launching the Iollowing
queries: autour * ans, actuellement * ans, encore * ans, environ * ans, maintenant * ans.
We applied the second step to access the Google search engine tuned to the French language.
The results obtained Irom the Internet produced 3717 examples that corresponded to 24 cases:
Ior example, g(e,s,s) dautour de NUM ans, encore maintenant NUM ans. For each oI the
24 cases we manually analyzed the number oI examples that corresponded to a person`s age.
Comparison
We considered the results obtained Irom the Internet comparable since they derived Irom
similarly qualiIied authors using similar registers. To compare the age phrases, we considered
the Iollowing elements: (1) adverbs, (2) surIace structure oI the phrase between adverb and
noun years, and (3) adjective placement. For this comparison we divided the results into Iour
groups corresponding to the adverbs: ahora/maintenant 'now, alrededor/autour-environ
'around, actualmente/actuellement 'at present, an/encore 'still. In general, we classiIied
our compared examples as identical, diIIerent in some respects, and having no equivalent.
We matched the above-described groups in Spanish and French by considering Iirst the
adverb, then the age meaning and Iinally the percentage oI phrases with age-related meaning.
For example:

Type of phrase # ex/
% age
Type of phrase # ex/
% age
environ ge de NUM ans 1/100
ge(e,s,es) d`autour de NUM ans 2/100
ge(e,s,es) d`environ de NUM ans 37/84
ge(e,s,es) d`environ NUM ans 24/80
ge(e,s,es) environ de NUM ans 38/74
autour de NUM ans 430/56
alrededor de los NUM aos 355/84
environ de NUM ans 158/30
autour de mes NUM ans 3/67
environ NUM ans 285/19
3
The group oI phrases initiated by the 'around adverb is the only French group including one
case where the structure oI the phrase between the adverb and the quantity oI years is a
prepositional phrase with a possessive adjective related to the person whose age is described
(autour de mes NUM ans). This French case has no equivalent in Spanish in this group. The
environ NUM ans case has no Spanish equivalent in this group in addition to being the least
productive.
The Iirst case is classiIied as diIIerent in some respects. These phrases have a de prepositional
phrase in both languages governed by the adjective in Iour French counterparts. The other three
phrases have the prepositional phrase governed by the adverb, being more similar to the Spanish
phrase but less productive oI age-and-person phrases.
Annotation and Translation
Generally, temporal expressions are annotated as entire constituents, typically noun phrases
(e.g. 38 years). For example:
environ 10 secondes about 10 seconds` is annotated in |1| as:
TIMEX3 tid"t1" type"DURATION" value"P10S" mod"APPROX"~environ 10
secondes/TIMEX3~
We suggest annotating the phrases we analyzed here with the TLINK tag. TLINK is a
temporal link that can represent the relation between two temporal elements: TIMEX3-
TIMEX3. The phrases we considered Ior this annotation were the more productive ones with
appropriate contexts extracted Irom the examples.
The results classiIied as diIIerent in some respects were the cases that we analyzed in detail to
obtain alignment templates Ior their correct translation. These Spanish and French temporal
expressions Iorm a contiguous sequence given an appropriate context, and can be translated into
the entire sequence as a multi-word unit. In this work we suggest translating them to take into
account a diIIerent syntactic Iorm and ordering Iunction words` arguments.

We conclude that variety in the structure oI temporal expressions necessitates analysis oI
diIIerent combinations oI classes oI words. Our study provides insights into the cross-lingual
behavior oI the temporal structure oI age expressions particularly in the type
AdvTsomethingTimeN as realized in Spanish and French.
We made an empirical veriIication oI a substantial degree oI parallelism between the
realization oI such age expressions in Spanish and French but showed the diIIerences in their
Irequency, structure and variety.
References
1. Bittar, Andre. ISO-TimeML Annotation Guidelines Ior French. Version 1.0
http://www.linguist.univ-paris-diderot.Ir/~abittar/docs/FR-ISO-TimeML-Guidelines.pdI,
(2010)
2. Ferro, Lisa, Laurie Gerber, Inderjeet Mani, Beth Sundheim and George Wilson. TIDES 2005
Standard Ior the Annotation oI Temporal Expressions, MITRE Corporation (2005).
3. Gelbukh, A. and I.A. Bolshakov. Internet, a true Iriend oI translator: the Google wildcard
operator. International Journal oI Translation 18(12), 4148 Bahri Publications (2006)
4. KilgarriII, A. Googleology is Bad Science. Computational Linguistics 33, 147151 MIT
Press (2007)
5. Salmon-Alt, Susanne, Eckhard Bick, Laurent Romary and Jean-Marie Pierrel. La
FREEBANK: Vers une base libre de corpus annotes. In: Proceedings oI TALN 2004. Fes,
Morocco. http://corp.hum.sdu.dk/tgrepeyeIr.html
6. Sauri, R., J. Littman, B. Knippen, R. Gaizauskas, A. Setzer and J. Pustejovsky. TimeML
Annotation Guidelines Version 1.2.1 (2006) http://www.timeml.org/site/publications/
timeMLdocs/annguide1.2.1.pdI
7. Sauri, Roser, Estela Saquete and James Pustejovsky. Annotating Time Expressions in
Spanish. TimeML Annotation Guidelines. Version TempEval-2010.
4
A Contrastive Study on Abstract Anaphors
in German and English
Stefanie Dipper
1
, Christine Rieger
2
, Melanie Seiss
2
& Heike Zinsmeister
2
Ruhr University Bochum
1
& Konstanz University
2
Introduction Abstract anaphors denote anaphoric relations between some
anaphoric expression and an antecedent that refers to an abstract object like
an event or a fact. In the classical example by Byron (2002), the pronoun it
(underlined in (1a)) refers to an event: the migration of penguins to Fiji. In
the alternative sequence, (1b), the demonstrative pronoun that refers to the fact
that penguins migrate to Fiji in the fall.
(1) a. Each Fall, penguins migrate to Fiji. It happens just before the eggs hatch.
b. Each Fall, penguins migrate to Fiji. Thats why Im going there next month.
The automatic resolution of abstract anaphors still poses a problem to language
processing systems.
We pursue a contrastive, corpus-based approach to investigate the proper-
ties that characterize dierent instantiations of abstract anaphora in English and
German. In the long run, we envisage to derive features from the corpus anno-
tation that will serve us to tackle the automatic resolution of abstract anaphors.
In this paper we investigate what kind of anaphoric elements are employed in
the two languages to refer to abstract objects. The range of possible realizations
includes pronouns, lexical NPs (e.g. this issue, this situation, etc.), adverbials
(e.g. likewise).
We present results of a comparative corpus study on the realization of ab-
stract anaphora in a parallel bi-directional corpus of English and German. Be-
sides comparing the cross-linguistic realizations, we also look into the dierences
between original text and translated text in both languages.
Most annotation projects that analyze abstract anaphora restrict themselves
to pronominal markables (e.g. Byron (2003), Hedberg et al. (2007), M uller
(2007), Dipper and Zinsmeister (2011)); some also annotate full NP markables
(e.g. Vieira et al. (2002), Pradhan et al. (2007), Poesio and Artstein (2008)).
Multilingual corpora have been annotated in Recasens (2008), Navarretta and
Olsen (2008), and Vieira et al. (2002). A recent overview of projects annotating
abstract anaphora is provided by Dipper and Zinsmeister (2010).
Annotation of parallel texts has been performed in Vieira et al. (2002). The
present work deals with the annotation of the full range of abstract anaphors
(including full NPs and anaphoric adverbs) in a parallel corpus.
Corpus For our study, we extracted about 100 German and English medium
sized turns each (contributions by German and English speakers; average length
about 20 sentences), along with their sentence-aligned translations, from the
Europarl Corpus (Release v3, 19962006, Koehn (2005)). For cross-lingual an-
notation of the German and English turns, we used two MMAX2 annotation
windows, put side by side on the screen.
We started out with a well-dened set of markables in the original language
and collected all variants of translations on the side of the target language
(the translation of the original language). In the rst round of annotation, we
chose original texts from German, because in German there isin contrast to
Englisha pronoun that is unambiguously used as an abstract anaphor: the
uninected singular demonstrative pronoun dies (this). In addition to this, we
dened as markables the (ambiguous) demonstrative pronoun das (that) and
the (ambiguous) third person neuter pronoun es (it). The target language was
English.
For the second round of annotation we considered the reversed translation
direction: English original texts and their German translations. We extended
our set of markables and included the adverbs as, so and likewise, because these
adverbs frequently served as translations of German anaphors in the rst round.
We will apply this method of bootstrapping back and forth to extend the set
of markables iteratively. This approach allows for a fast and ecient way of
extracting anaphors in both languages.
Results The German part of the corpus features 223 abstract anaphors203
of which could be aligned with English text instances. On average, we identied
2.37 abstract anaphors per turn (with the basic set of markables).
The English part of our corpus contains 77 turns. It features 234 abstract
anaphors. This corresponds to 3.03 abstract anaphors per turn (with the ex-
tended set of markables).
We used our annotations to test the hypothesis that English avoids the use of
pronominal abstract anaphors. The results from the German-to-English (DE-to-
EN) annotations seem to support this hypothesis. 35% of German pronominal
anaphors (71 out of 203) were not translated as pronouns in English.
We identied the following main strategies to avoid pronominal anaphors in
the translation of German to English:
there is no corresponding material, e.g. a dierent verb or a dierent argu-
ment frame is employed, see Ex. (2)
1
use of full NPs rather than pronouns (all these things, these measures, this
objective, this situation, this thread . . . ).
use of adverbials or conjunctions (likewise, so, as)
1
In the examples, the lines prexed with DE contain the German original text,
the EN lines the ocial English translation, and the DE-LIT lines a literal
translation of (parts of) the German original.
(2) DE: Europa ist nie fertig! Aber das Projekt muss entschlossen, gemeinschaftsorientiert
und vision ar zur politischen Union weiterentwickelt werden. Wenn dies nicht geschieht,
verlieren wir das Vertrauen der B urger.
EN: Europe will never be nished, but we must press on with the project for
political union with determination and vision, and on a Community basis. If we
do not, the public will lose condence in us.
DE-LIT: . . . If this does not happen, the public will lose condence in us.
Following Passonneau (1989) and Navarretta (2008), we hypothesized that
English prefers demonstrative pronouns to personal pronouns in abstract anaphora
in comparison to other languages. Our ndings are that both German demonstra-
tive and personal pronouns tend to be translated as demonstratives in English,
as in Ex. (3).
(3) DE: Sie selbst haben gesagt: Vertrauen ist herzustellen. Tun Sie es!
EN: You said yourself that trust had to be built up. Do that!
DE-LIT: Do it!
The German pronoun es it is often not represented in the English transla-
tion. 43% of es-anaphors do not receive a pronominal translation vs. only 32%
of the demonstrative anaphors are not translated into a pronoun. Furthermore,
comparing the frequencies of anaphoric pronouns and selected anaphoric adverbs
in the English turns, the annotations show that 73% of the instances are realized
by demonstrative pronouns.
So far we discussed results from comparing original German texts (GO)
and their English translations (ET). To be able to really interpret the results
of this contrastive analysis, it is important to show that there is no signicant
dierence between the English translated texts (ET) and English original texts
(EO). For the purpose of this abstract, we tested for signicant dierences at
dierent levels of abstraction. At the coarse-grained level, we found no signicant
dierence between the proportions of pronominal subjects and objects in the
EO and ET text. Likewise, on a more ne-grained level, there was no signicant
dierence between the proportions of specially marked constructions in the EO
and ET texts, such as topicalization in Ex. (4)-EN. These ndings seem to allow
us to using the translated texts in comparing German and English usage of
abstract anaphors.
(4) DE: Wir k onnen es nicht andern.
EN: That is something we cannot change.
DE-LIT: We cannot change it.
Conclusion Although the main target of our research is to detect cross-linguistic
and contrastive features for automated anaphora resolution, we believe that these
features are also important for eective, high-standard machine translation. For
both applications it is necessary to consider features in detail, such as paying
attention to grammatical function and syntactic position.
References
Donna K. Byron. Resolving pronominal reference to abstract entities. In Pro-
ceedings of the ACL-02 conference, pages 8087, 2002.
Donna K. Byron. Annotation of pronouns and their antecedents: A comparison
of two domains, 2003. Technical Report, University of Rochester.
Stefanie Dipper and Heike Zinsmeister. Towards a standard for annotating ab-
stract anaphora. In Proceedings of the LREC 2010 workhop on Language
Resource and Language Technology Standards, pages 5459, Valletta, Malta,
2010.
Stefanie Dipper and Heike Zinsmeister. Annotating abstract anaphora. Language
Resources and Evaluation, Online First, 2011. doi: DOI 10.1007/s10579-011-
9160-1.
Nancy Hedberg, Jeanette K. Gundel, and Ron Zacharski. Directly and indirectly
anaphoric demonstrative and personal pronouns in newspaper articles. In
Proceedings of DAARC-2007, pages 3136, 2007.
Philipp Koehn. Europarl: A parallel corpus for statistical machine translation.
In Proceedings of MT Summit, 2005.
Christoph M uller. Resolving it, this, and that in unrestricted multi-party di-
alog. In Proceedings of ACL-07 conference, pages 816823, 2007. URL
http://www.aclweb.org/anthology/P07-1103.
Costanza Navarretta. Pronominal types and abstract reference in the Danish and
Italian DAD corpora. In Proceedings of the Second Workshop on Anaphora
Resolution, pages 6371, 2008.
Costanza Navarretta and Sussi Olsen. Annotating abstract pronominal anaphora
in the DAD project. In Proceedings of LREC-08, 2008.
Rebecca J. Passonneau. Getting at discourse referents. In Proceedings of the
27th Annual Meeting of the Association for Computational Linguistics, 1989.
Massimo Poesio and Ron Artstein. Anaphoric annotation in the ARRAU corpus.
In Proceedings of LREC-08, 2008.
Sameer Pradhan, Lance Ramshaw, Ralph Weischedel, Jessica MacBride, and
Linnea Micciulla. Unrestricted coreference: Identifying entities and events in
OntoNotes. In Proceedings of the IEEE-ICSC, 2007.
Marta Recasens. Discourse deixis and coreference: Evidence from AnCora. In
Proceedings of the Second Workshop on Anaphora Resolution, pages 7382,
2008.
Renata Vieira, Susanne Salmon-Alt, and Caroline Gasperin. Coreference and
anaphoric relations of demonstrative noun phrases in a multilingual corpus.
In Proceedings of DAARC-2002, 2002.
A Corpus-based Contrastive Analysis for Defining Minimal
Semantics of Inter-sentential Dependencies for Machine
Translation
Thomas Mever, Andrei Popescu-Belis, Jeevanthi Livanapathirana
Idiap Research Institute, Martigny, Switzerland
Bruno Cartoni
University oI Geneva, Switzerland
Abstract
Inter-sentential dependencies such as discourse connectives or pronouns have an impact on
the translation oI these items. These dependencies have classically been analyzed within
complex theoretical Irameworks, oIten monolingual ones, and the resulting Iine-grained
descriptions, although relevant to translation, are likely beyond reach oI statistical machine
translation systems. Instead, we propose an approach to search Ior a minimal, Ieature-based
characterization oI translation divergencies due to inter-sentential dependencies, in the case oI
discourse connectives and pronouns, based on contrastive analyses perIormed on the Europarl
corpus. In addition, we show how to automatically assign labels to connectives and pronouns,
and how to use them Ior statistical machine translation.
1. 1he Aeed for Inter-sentential Information in Machine 1ranslation
Long-range dependencies are a well known challenge Ior machine translation (MT) systems,
especially Ior statistical ones. The correct translation oI lexical items such as pronouns oIten
depends on the correct identiIication oI their antecedent. Similarly, the correct translation oI
multi-Iunctional discourse connectives depends on the correct identiIication oI the rhetorical
relation which they convey between two clauses. However, especially when translating
between closely related languages, the Iull disambiguation oI such lexical items is sometimes
unnecessary Ior a correct translation. The question that arises is thus how to Iind the most
suitable level oI representation Ior such dependencies, as a trade-oII between linguistic
accuracy and computational tractability, with the direct aim oI improving MT output.
This paper presents a method Ior Iinding the minimal semantic/discourse inIormation that
must be assigned to two types oI lexical items, namely connectives and pronouns, in order to
avoid translation mistakes by statistical MT systems. The method starts Irom contrastive
analyses oI a Irequently used parallel corpus, Europarl (Koehn, 2005), in order to deIine and
annotate the minimal semantic/discourse inIormation necessary Ior MT. The paper Iirst
describes our analyses and manual annotation methods Ior disambiguating connectives
(Section 2.1) and pronouns (Section 2.2), in the context oI English/French MT. Section 3
outlines methods Ior automatically perIorming these disambiguation tasks, while Section 4
explains how the automatically labeled linguistic items can be integrated into a statistical MT
system. Section 5 concludes the paper and outlines Iuture work.
2. Contrastive Analysis of 1wo 1ypes of Inter-sentential Dependencies
2.1 Discourse Connectives
Discourse connectives are generally considered as indicators oI discourse structure, relating
two sentences or propositions and making explicit the rhetorical relation between them.
Explicit discourse connectives such as because, but, however, since, while, etc., are Irequent
lexical items and are used to mark rhetorical relations such as Cause or Contrast between
units oI discourse. Several theoretical Irameworks have been proposed Ior connectives
(mainly starting Irom English ones), such as the Rhetorical Structure Theory (RST) (Mann
and Thompson, 1988), or the Segmented Discourse Representation Theory (SDRT) (Asher,
1993). In such theories, more than one hundred possible rhetorical relations have been
identiIied, and complex semantic and logical representations have been used to characterize
discourse structure. In a more empirically oriented eIIort, the Penn Discourse Treebank
(PDTB) (Prasad et al., 2008) contains manual annotations oI discourse connectives with a
large set oI labels: Ior example, the connective while was annotated with 17 possible senses
beyond its Ior main meanings, which are Comparison, Contrast, Concession and Opposition
(Miltsakaki et al., 2005).
While a Iine-grained characterization provides the necessary theoretical level oI linguistic
description oI discourse structure, it may prove to be intractable to Iully automatic processing.
Nevertheless, the disambiguation oI at least the main senses oI discourse connectives is
generally required Ior their translation
1
, to avoid the rendering oI a wrong sense in translation.
For instance, in the Iollowing example, the French connective alors que in its contrastive
usage is wrongly translated to the English connective so, which signals a causal meaning
instead
2
.
FR: Oui, bien entendu, sauf que le developpement ne se negocie pas, alors que le
commerce, lui, se negocie.
EA: *Yes, of course, but development cannot be negotiated, so that trade can.
To disambiguate connectives Ior MT, parallel corpora with sense-labeled connectives are
required Ior training and test. As the PDTB data is in English only, we perIormed manual
annotation on the Europarl corpus. The annotation method, called translation spotting,
requires annotators to consider bilingual sentence pairs, and annotate each connective in the
source language with its translation in the target language (Meyer et al., 2011). A contrastive
analysis showed that these translations can be: a target language connective (in principle
signaling the same sense(s) as the source language one), reIormulations with diIIerent
syntactical constructs, or no connective at all. The indications gained with this method are
then used in a second step to manually derive and cluster the minimal semantic and theory-
independent labels needed to generate correct translations oI a connective.
We exempliIy this procedure here Ior the English connective while. From the Europarl corpus
Ior English-French, we extracted 499 sentences containing the connective while. In 198 cases
(43) the annotators spotted 'no translation' or reIormulations oI the connective
3
. In the
remaining 301 sentences (57), the annotators identiIied the corresponding French
connectives. As a second step, the French connectives (signaling the same rhetorical
relation(s) as while itselI) were manually clustered under the minimally necessary sense labels
to disambiguate the connective while in order to translate it correctly Irom EN to FR. The
most Irequent French connective clusters and the derived sense labels are the Iollowing:
alors que (18) Contrast/Temporal
si / mme si / bien que / sil est vrai que (25) Concession
tandis que / mais (9 ) Contrast
tant que (2) Temporal/Causal
pendant (1) Temporal/Duration
puisque (1) Temporal/Causal
lorsque (0.8) Temporal/Punctual
1
The only exception is the case when the ambiguity oI a connective is conserved in translation.
2
Source sentence Irom Europarl, translated by Moses (Koehn et al., 2007) trained on Europarl.
3
These are valid translation problems and will be reconsidered Ior clustering in Iuture work.
Compared to the PDTB sense hierarchy Ior example, the clustered senses Ior while are as
detailed as the PDTB ones on hierarchy level 2, but less detailed than the deepest PDTB level
3. For the temporal meaning oI while, however, even more diIIerentiation than PDTB level 3
is needed in order to be able to generate the correct translations.
2.2 Pronouns
The resolution oI pronouns can be seen as a similar issue to that oI resolving connectives in
terms oI Iinding a minimal set oI Ieatures to disambiguate a pronoun Ior translation. In many
cases, depending on the language pair, pronouns can be translated unequivocally, such as the
English pronoun he generally rendered by il in French. However, the French pronouns il and
elle may both be translated into it in English iI their antecedent, i.e. the noun they reIer to, is
not human. However, iI the antecedent is human, they are in general translated respectively as
he and she. Vice versa, the translation oI the English pronoun it into French requires
knowledge about the gender oI its antecedent in the target text. ThereIore, whereas the
disambiguation oI connectives can be done on the source text only, prior to MT, the
translation oI pronouns requires inIormation about the translation oI neighboring Iragments.
A close comparison oI the English and French pronoun systems shows that the complete list
oI Ieatures characterizing pronoun choice is in theory very large. However, we only aim here
to Iind the minimal set oI Ieatures which will allow a statistical MT system to avoid
generating mistaken pronouns, taking also into consideration the pronoun generated by the
system without these Ieatures. For instance, in the Iollowing example Irom Europarl, the
pronoun generated by Moses is correct in every respect except the gender; thereIore,
knowledge about the required gender would help correcting il into elle.
EN: The European Commission must make good these omissions as soon as possible. It
must also cooperate with the Member States...
FR: *La Commission europenne doit reparer ces omissions des que possible. Il doit
egalement cooperer avec les Etats membres ...
3. Automated Disambiguation for Machine 1ranslation
To improve the output oI MT, we propose automatic methods that attempt to disambiguate, or
at least set additional constraints, on the translation oI connectives and pronouns. These
methods can either be used as direct input to MT, or to prepare training data Ior it. For
instance, using surIace Ieatures such as part-oI-speech tags or syntactical and dependency
parses, we have built classiIiers (Meyer et al., 2011) Ior the senses oI the English connectives
since (Temporal, Causal, or Temporal/Causal) and while (Temporal/Causal, Temporal/
Punctual, Temporal/ Durative, Contrast/Temporal, Contrast, or Concession), as well as Ior the
French connective alors que (Temporal, Contrast, Temporal/Contrast).
since while alors que
Baseline (most Irequent sense) 51.6 44.8 46.9
SVM classiIier 85.7 60.9 54.2
Table 1: Accuracies oI sense disambiguation Ior the connectives since (700 sentences), while (300) and
alors que (400). For comparison, the baseline is the majority class in each training set, i.e. respectively
Cause, Concession, and Contrast.
ClassiIiers were also built Ior pronoun disambiguation, considering in addition to Ieatures
Irom the source text also Ieatures Irom a candidate translation, such as inIormation about the
preceding noun phrases, the candidate Moses translation oI the pronoun computed Irom the
GIZA word alignment, and various ways to determine gender constraints Ior the
translation oI English it into French Irom the gender oI the preceding nouns (e.g., majority,
most recent, etc.). Although this method bears similarities with that oI LeNagard and Koehn
(2010), we do not attempt to identiIy explicitly the antecedent, in the target language, oI the
pronoun under consideration, but train classiIiers to use the optimal combination oI Ieatures
to inIer the correct gender. OI course, this approach cannot pretend to be Iully accurate, but
compares Iavorably to state-oI-the-art accuracy oI automatic pronoun resolution.
The accuracy oI the classiIier, a decision tree trained using the C4.5 algorithm, is 61 using
ten-Iold cross-validation on a set oI 393 sentences Irom Europarl annotated with the correct
pronoun. The task was to correct the Moses candidate translation oI English it into French (il,
elle, le, la, l, lui, celui-ci, celle-la, ce, c) using automatic alignment and automatically
extracted surIace Ieatures. II the alignment is manually corrected, then the accuracy reaches
64. This small increase shows that alignment is not the main issue, also because it cannot
deal with cases when the MT system omitted the pronoun in translation. However, when the
gender prediction is manually corrected, the accuracy reaches 88, which shows that, as
expected, gender is the main Ieature required Ior correct translation oI it into French.
4. Integration into Statistical M1
We experimented on three ways to propagate the above-mentioned discourse inIormation
annotated to connectives into the MT processing chain. The integration oI annotated pronouns
proceeds diIIerently, as a way to post-edit candidate pronouns generated by MT.
The Iirst method to integrate the minimal sets oI labels Ior discourse connectives is to tag
their occurrences directly in the phrase table oI an already trained statistical MT system.
During the training stage, a phrase table is generated with all phrase pairs Iound by the word
alignment, with their lexical probability and Irequency scores. We tagged three senses oI the
connective while, namely Temporal (1), Contrast (2) and Concession (3) in the phrase table
oI a trained Moses MT system Ior EN-FR. The most Irequent French translations were: (1)
pendant que, (tout) en J-ant, (2) alors que, tandis que, (3) bien que. Each phrase containing
while was automatically checked iI it is Iollowed by a corresponding translation. II Iound, the
word Iorm while was annotated with while-1, while-2 or while-3, and, in addition, the lexical
probability score was set to one (all other occurrences were leIt untagged). Translations tests
with a set oI 20 sentences already led to noticeably better translations (i.e. automatically
generated translations closer to the reIerence translations, especially in terms oI the
connective) which were also conIirmed by a rise in the BLEU score oI 0.8 absolute.
A second method that we explored is the opposite oI Iorcing the system to use the tagged
connectives. They are instead automatically tagged in a large corpus which is used Ior SMT
training, where all connectives Iollowed by their tags and their corresponding translation in
the parallel corpus can be learned by the system. Every occurrence has thereby to be tagged
by the disambiguation tool using the classiIier model. A third and similar approach to this
method is to directly use the manually annotated discourse connectives aIter the sense
clustering. This has the advantage that the hand-annotated resources are correct (gold
standard) as opposed to the automated tagging, which is well below 100 accuracy and may
thereIore propagate a certain error rate in the whole translation process. We built and trained
SMT systems able to handle the same manually or automatically tagged data. As a basis Ior
comparison, two other systems were trained on the same two corpora, by discarding all labels
(resulting in 4 SMT systems). When comparing the manually tagged system to its untagged
counterpart, the tagged system got closer to the reIerence translations oI a test set oI 35
sentences in 21 cases versus 14 cases only Ior the untagged system (the counts were done
based on manual checking oI the connective translation and the surrounding words and
syntax). Even the automatically tagged system, tested on 62 sentences, perIormed noticeably
better in 14 cases compared to its untagged counterpart.
For pronouns, we evaluated the eIIect on translation oI replacing every candidate translation
oI the English it, in the MT output to French, by the translation proposed by our classiIier, as
a Iorm oI post-editing. By deIinition, this method is only applicable to sentences where a
pronoun was indeed generated by MT (about 95 oI the sentences). We perIormed Iive
diIIerent runs, training on 353 sentences and testing on 40. In the Iully automatic setup, this
resulted, on average, in improving pronoun choice Irom incorrect to correct in 10.8 sentences
(27), but also in turning 6.6 (16) correct pronouns into incorrect ones. The global result is
thus an improvement oI about 10 oI the overall pronoun accuracy. In these experiments, our
classiIier did not change the pronoun proposed by MT in 22.6 sentences (56), oI which 27
were correct and 29 were incorrect.
5. Conclusion and Future Work
Integrating discourse inIormation into statistical MT systems remains a challenging task, but
one which has the potential to improve over the current sentence-by-sentence MT paradigm.
The contrastive corpus analyses and the translation-oriented, multilingual annotation methods
have shown to positively aIIect the output oI current statistical MT systems. We will Iurther
investigate the automated disambiguation methods Ior pronouns and connectives as well as
Ior verbal tenses. The perIormance and error rate oI the disambiguation tools is crucial in
order to generate annotated resources which are as error-Iree as possible in order to not
negatively inIluence the SMT training and testing on these resources.
References
Nicholas Asher. 1993. Reference to Abstract Obfects in Discourse. Kluwer Academic Publisher,
Dordrecht, NL.
Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola
Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar,
Alexandra Constantin, and Evan Herbs. 2007. Moses: Open source toolkit Ior statistical machine
translation. In Proceedings of ACL 2007 (45th Annual Meeting of the ACL), Demonstration Session,
pages 177180, Prague, Czech Republic.
Philipp Koehn. 2005. Europarl: A parallel corpus Ior statistical machine translation. In Proceedings of
MT Summit X, pp. 7986, Phuket, Thailand.
William C. Mann and Sandra A. Thompson. 1988. Rhetorical Structure Theory: towards a Iunctional
theory oI text organization. Text, 8(3):243281.
Thomas Meyer, Andrei Popescu-Belis, Sandrine ZuIIerey, and Bruno Cartoni. 2011. Multilingual
Annotation and Disambiguation oI Discourse Connectives Ior Machine Translation. Proceedings of
SIGDIAL 2011 (12th annual SIGdial Meeting on Discourse and Dialogue), pp. 194203, Portland, OR.
Eleni Miltsakaki, Nikhil Dinesh, Rashmi Prasad, Aravind Joshi, and Bonnie Webber. 2005.
Experiments on sense annotations and sense disambiguation oI discourse connectives. In Proceedings
of the TLT 2005 (4th Workshop on Treebanks and Linguistic Theories), Barcelona, Spain.
Ronan Le Nagard and Philipp Koehn. 2010. Aiding Pronoun Translation with Co-ReIerence
Resolution. In Proceedings of the Joint 5th Workshop on Statistical Machine Translation and Metrics
MATR, pp. 258267, Uppsala, Sweden.
Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, and Bonnie
Webber. 2008. The Penn Discourse Treebank 2.0. In Proceedings of LREC 2008 (6th International
Conference on Language Resources and Evaluation), pp. 29612968, Marrakech, Morocco.
Formalising translation behaviour with parallel treebanks
Oliver !ulo, Silvia Hansen-Schirra
Johannes Gutenberg-Universitt Mainz
culo|[email protected]
1 Introduction
Statistical machine translation, in a simplified view, is based on extracting translation probabilities from
parallel corpora. However, these corpora are used rather uncritically. Factors like translation direction or text
type/register have largely been neglected. The factors mentioned here have a major influence on the
production of human translation. As e.g. (Koehn & Schroeder 2007) show, the influence of register is such
that training SMT models, both the monolingual language model as well as the translation model, on
domain-adapted data has a positive effect on the correctness of translations. But, in order to account for these
factors and to derive them automatically, high-quality annotated resources are necessary.
It is probably to a large extent the role of translation studies and contrastive linguistics to study and describe
divergences in local structures (cf. Hawkins 1986; Knig & Gast 2007 for English and German) as well as
the role of non-local factors such as text type and their effect on translation (cf. Hansen-Schirra et al.
forthcoming for English-German translations). In addition to this, the present contribution introduces the
formalisation of factors like register and translation direction as well as local divergences in the dependency
structure of a sentence. We try to link these factors to certain translation phenomena, and suggest first steps
for the implementation of an algorithm in order to facilitate the adoption of domain, register and translation
pair knowledge in MT.
2 Registerial influences
Register has an effect on various dimensions of a text. For instance, register has an effect on word order.
Table 1 shows that the number of subjects in sentence-initial position varies depending on the register.
1
German speeches adhere more frequently to the canonical word order SVO than German shareholders' letters
and even more frequently compared to German fictional texts. When translating from English, which has a
rather fixed SVO word order, into German these register-specific word order patterns should be taken into
account otherwise atypical word order frequencies might cause interference effects in the target texts, which
might in turn haven in impact on the target language.
GO_FICTION 42,16 %
GO_SHARE 45,87 %
GO_SPEECH 54,54 %
Table 1: Percentage of subjects in sentence-initial position
Figure 1 shows typical shifts in syntactic functions for the language pair English-German and the influence
of the register. Shifts from subject to object are, for instance, more common for political essays and
instructional texts whereas they are less frequent for fictional texts and this holds true irrespectively of the
translation direction.
1 All results and examples presented in this article are all taken from the English-German CroCo Corpus (cf. Hansen-
Schirra et al. forthcoming).
3 Typological influences
Figure 1, however, also shows a typical shift which is triggered by typological differences: the non-canonical
subject positions in German often cause objects being placed in sentence-initial position. These sentence-
initial German objects are translated with English subjects. This means that the word order is kept in the
translation, but as the direct object cannot be kept in sentence-initial position in English an active-passive
shift is applied (cf. the translation procedure modulation by Vinay & Darbelnet 1958). This phenomenon is
illustrated through the following examples taken from the subcorpus of shareholders' letters:
Wichtige Erfolge [DIRECT OBJECT] knnen wir bereits verzeichnen, weitere werden folgen.
Some i mportant [SUBJECT] successes have already been achieved, others will follow.
Einzelheiten [DIRECT OBJECT] knnen Sie diesem Bericht entnehmen.
Additional details [SUBJECT] are contained in this report.
Consequently, for the translation direction from German into English the object-to-subject shift should be
more frequent than translating from English into German. This typologically driven pattern can be seen from
figure 1 where object-to-subject shifts are more typical of the translation direction German-English
irrespectively from the register.
A similar translation behaviour can be detected when looking at part-of-speech shifts (cf. the translation
procedure transposition by Vinay & Darbelnet 1958). Figure 2 shows that shifts from verbal to nominal
word classes are typical of the translation direction English-German (e.g. adverb-adjective, verb-adjective,
verb-noun) while the opposite shifts from nominal to verbal constructions are less frequent (e.g. adjective-
adverb, noun-adverb, noun-verb). This conforms to the typological differences between English and German,
the latter being more nominal and content-oriented (cf. House 1997). Nevertheless, figure 2 also shows
register-specific translation behaviour e.g. when it comes to high-frequency nominalisation patterns in
English-German instructions:
If vertical bars appear on the display after adjusting [VERB] the focus , press ...
Falls nach dem Einstellen [NOUN] des Fokus vertikale Streifen auf dem Display erscheinen , drcken Sie ...
When replacing [VERB] the lamp be sure to turn off power and unplug the power cord.
Schalten Sie vor dem Austausch [NOUN] der Lampe das Gert aus und entfernen Sie das Netzkabel.
Figure 1: Shifts in syntactic functions
!
"
# !
# "
$ !
$ "
% !
% "
& !
& "
" !
' ( ) * + , - . /
' ( ) * + , 0 ) *
, - . / + ' ( ) *
, 0 ) * + ' ( ) *
4 Local divergencies in dependency structure
Dependency treelet translation (e.g. Ding and Palmer 2004; Quirk, Menezes, and Cherry 2005) is one branch
of syntactically informed SMT. (Ding and Palmer 2004) present methods which cope with various changes in
the dependency structures, e.g. when head and dependent switches take place or when a dependent is
removed and appears as dependent of another head. The account of (Quirk, Menezes, and Cherry 2005) deals
with various re-ordering phenomena, e.g. to deal with post- rather than pre-position of modifiers, a typical
difference between French and English.
While register as well as typological features may have an influence on the frequency of such phenomena,
these phenomena are at first local (effect-wise). We want to add further descriptions of such local phenomena
which we have observed while annotating parallel dependency structures and aim to link them to certain
translation properties. In the following, we will present a selection of the phenomena we encountered.
Previously, two types of alignment phenomena with regard to translation shifts were defined in the course of
the CroCo project: empty links, corresponding to 1:0-alignments, and crossing lines, mostly corresponding to
shifts in grammatical function. These alignment phenomena were defined on the flat, top-level only
annotation of grammatical functions in CroCo. However, when looking at alignments in dependency
annotations, these concepts must be adapted to fit the added dimension of depth in the trees.
We annotated and aligned a sample of the CroCo corpus (around 4,000 sentences from 8 registers) in
dependency fashion. For this, we used the tree editor TrEd
2
. The figures 3 to 5 show examples of this
annotation and alignment. Due to the display settings in TrEd, the trees from the original sentences are on the
right, the trees of the translated sentences on the left.
Figure 3: An added leave: sich is inserted as additional complement in
German due to changed valency properties
2 http://ufal.mff.cuni.cz/~pajas/tred/ (last visited June 14
th
2011)
Figure 2: Shifts in part-of-speech
! " ! ! #
$ " ! ! #
% " ! ! #
& " ! ! #
' " ! ! #
( " ! ! #
) " ! ! #
* " ! ! #
+ " ! ! #
, " ! ! #
$ ! " ! ! #
- % . / - 0 0 1 2
- % . / 3 4 5 6 4 7 8
- % . / 4 8 0 6 9
- % . / : 7 : 0 5 4
- % . / 0 ; 1 9 -
- % . / 0 : - - 5 ;
- % . / 6 7 <
- % . / = - >
One typical phenomenon in translation is that of a syntactic complement being added or deleted. In the case
of an additional complement, this triggers an added leaf. We see such an example in figure 3, where the
English verb be has been translated with a reflexive variant sich befinden. The reflexive pronoun sich has no
alignment and triggers a new node plus a new incoming edge. In the other translation direction, i.e. in cases
where a complement is dropped, we speak of a dropped leaf.
Sometimes, the addition or deletion of nodes rather corresponds to the fact that a phrasal expression in one
language matches a single word in the other language. We see such a case in figure 4. Here, the expression in
the face of corresponds to the German adverb angesichts. A number of nodes in the English tree is collapsed
into a single node in the German tree. This also happens in the very frequent cases of compound nouns
which are written as one word in German, but as separate words in English (e.g. multi word expression
Mehrwortausdruck). In cases in which several nodes are collapsed into one, we speak of collapsed nodes, in
the opposite case we speak of expanded nodes. Note that in the case of figure 4, for the expressions in the
face of and angesichts we get treelet pairs in which the head is to be filled. In terms of formalisation of the
structures, this correspond to type-B trees (root unlexicalised trees) as defined by (Ding and Palmer 2004).
Figure 4: An expanded node: The phrasal expression in the face of
becomes the adverbial angesichts in German, a typical divergency for
the two languages
However, changes in the dependency structure not only take place at the terminal nodes of the tree, but
sometimes within the tree, changing the path from the root to the affected terminal node(s). In the case of the
example in figure 5, the auxiliary verb did is present in the English original (He never gave her gifts the way
the old man did) as verbal substitution, but this verbal branch is cut in the German translation (Der hat ihr
nie was geschenkt, so wie der alte Mann). This is due to a typical contrast between English and German, the
latter rather construing cohesion through elliptical constructions (cf. Hawkins 1986). With respect to
dependency annotation, we obtain cut branches, or, when translated from German to English, inserted
branches. Besides enlarging the number of nodes and edges and prolonging the path from the root to the
daughter tree of the inserted branch, these inserted branches usually have additional effects. In the example
from figure 5, the fact that the verbal branch is cut triggers a shift of the German Mann to a modal adverbial,
as opposed to its English counterpart which is the subject of the auxiliary.
Figure 5: The node did plus in- and outgoing edges was cut from the tree in
the German translation. The former dependent of did is shifted from a
subject to a modal adverbial.
5 Discussion
The present contribution has shown the influence of register on translation shifts, as well as of various
factors like typological differences between languages or valence divergencies on the dependency structure
of translated sentences. The former findings go along with the findings in SMT that training language and
translation models on domain-adapted data will improve the performance of the model. The latter findings
present a selection of shifts in dependency structures which can be included as tree configurations dealt with
in dependency-based SMT
While formal accounts on translation divergencies are available (e.g. Vinay & Darbelnet 1958, Catford
1965), algorithmic formalisations such as (Dorr 1994) are needed in order to facilitate adaption by the MT
community. The examples and results presented here have shown that domain and register knowledge as well
as patterns typical of the translation direction can be quantified and categorised according to the independent
variables involved (language, register, translation direction, etc.). Using parallel treebanks and exploiting
them quantitatively and qualitatively has paved the way for the formalisation of human translation behaviour.
The implementation of algorithms which represent this translation knowledge for MT will be our next steps
in future research.
6 References
Catford, John C. 1965. A linguistic theory of translation. an essay in applied linguistics. Oxford: Oxford
University Press.
Ding, Yuan, and Martha Palmer. 2004. Automatic learning of parallel dependency treelet pairs. In The first
international joint conference on natural language processing (IJCNLP-04).
Hansen-Schirra, Silvia, Stella Neumann, and Erich Steiner. forthcoming. Cross-linguistic Corpora for the
Study of Translations. Insights from the Language Pair English-German. Berlin: De Gruyter.
Hawkins, John A. 1986. A comparative typology of English and German. Unifying the contrasts. London:
Croom Helm.
House, Juliane. 1997. Miverstehen in interkulturellen Begegnungen. In Wie lernt man Sprachen - wie lehrt
man Sprachen, 154-169.
Koehn, Philipp, and Josh Schroeder. 2007. Experiments in domain adaptation for Statistical Machine
Translation. In ACL Workshop on Machine Translation 2007.
Quirk, Christopher, Arul Menezes, and Colin Cherry. 2005. Dependency treelet translation: syntactically
informed phrasal SMT. In Proceedings of the 43rd annual meeting of the ACL, ed. Ann Arbor, 271-
79.
Vinay, Jean-Paul, and Jean Darbelnet. 1958. Stylistique compare du franais et de l!anglais. Mthode de
translation. Paris: Didier.
!"#$% '$$()')*+ ,(-.(-' /(- -'.#+ +*0*1(.2*$) (/ $*3 1'$4
%5'%* .'#-" #$ 67
1 Introduction
The present paper introduces the PRESEMT MT system whose most innovative feature is that new
language pairs can easily and rapidly be set up. To this end, PRESEMT uses publically available re-
sources and tools as much as possible. These comprise corpora of various types and corpus annotation
tools such as statistical taggers and chunkers.
With the advent of statistical machine translation (SMT) corpora have started to play an important
role in machine translation (MT). However, whereas large monolingual corpora are mostly available,
e.g. through the world wide web, large bilingual corpora are much harder to obtain. The PRESEMT
MT system uses a mix of small bilingual corpora and large monolingual corpora to overcome this bot-
tleneck. The small bilingual corpora (several hundred sentences) will not be used to extract statistical
parameters (they would be too small), but to automatically (!) extract a bilingual phrase structural
mapping, a kind of contrastive (or synchronous) grammar of two languages.
Recent years have seen a rapid increase in the availability of publically available tools for corpus
annotation. Statistical taggers and chunkers are available for many languages. These tools will be in-
tegrated to do a source language and target language preprocessing which builds the input for the syn-
chronous grammar.
Bilingual dictionaries are available to a large extent for many language pairs, some of them open
source, for some publishers are willing to make them available, at least for research purposes. How-
ever, they are normally not tagged in appropriate and systematic ways and thus need processing be-
fore being integrated into the system.
Another set of tools that build a substantial part of the PRESEMT MT system are statistical mod-
ules that make the choices on the target language level. Those tools are mainly language models de-
rived from huge corpora of several billion words. On the basis of these algorithms choices about word
readings, placement of articles etc. are made.
2 Problems faced in MT
MT faces two problems, a cost problem and a quality problem. MT systems have high development
costs and the translation quality is poor for certain phenomena. Rule-based MT (RBMT) and statistic-
al MT (SMT) face these two problems in different ways:

RBMT:
Cost problem: manually written linguistic resources are expensive.
Quality problem: Word translation disambiguation is rather poor.
Strength:
RBMT has the theoretical means to account for rich morphology and non-local phenomena.
SMT:
Cost problem: large bilingual corpora are difficult to obtain.
Quality problem: rich morphology and non-local linguistic phenomena are a problem.
Strength:
Word translation disambiguation is accounted for rather well.
Hybrid MT systems have been proposed in order to benefit from the strengths of both systems while
overcoming their weaknesses. PRESEMT is also a hybrid system. It uses linguistic knowledge in the
preprocessing stage and by applying a synchronous grammar (which has been automatically derived
from a bilingual corpus). It also uses statistical knowledge to a considerable extent by processing the
target language. It also attempts to minimize the cost and the quality problem in the following way:
2.1 Minimising the cost problem
Rule-based components are not manually written but automatically generated.
Publically available taggers and chunkers are used as much as possible.
There is no dependency on large bilingual corpora, instead, a small parallel corpus suffices. In
addition, easily available large monolingual corpora are used.
Available bilingual dictionaries are used. No specific tagging is needed in the lexical entries.
The tagging information is taken from other resources such as the tagged monolingual corpora.
2.2 Minimising the quality problem
Linguistic representations are used to account for rich morphology and non-local phenomena.
Large monolingual corpora are used to account for word translation disambiguation and TL
structure and morphology.
3 Outline of the translation process
3.1 Preprocessing
Shallow parsers are adopted to annotate a small parallel bilingual corpus, which then is aligned on
word, chunk, and clause level, based on information found in a bilingual dictionary. From the bilin-
gual corpus, cross-linguistic information is automatically extracted which serves as rule base for a
synchronous grammar parser. The actual translation process consists of two phases:
3.2 Structure Selection module
The incoming source language (SL) sentence is tagged, chunked and clause-chunked by shallow pars-
ers. The resulting flat structure is fed into a synchronous grammar parser that builds up an SL tree
structure and a set of corresponding target language (TL) structures.
3.3 Translation equivalent selection module
Information extracted out of large monolingual corpora is used to do lemma disambiguation, to select
the best TL structure and to determine morphological features such as case, person, number and gend-
er, and to generate the appropriate TL tokens.
4 Translation pair English German
For the translation direction English (EN) German (DE) which is a first prototype the following re-
sources and tools have been used:
Publically available resources and tools:
A small bilingual corpus comprising 300 parallel sentences
A large monolingual German corpus comprising 7 million tokens (Again this is only the first
prototype. There are German corpora available that comprise several billion words).
TreeTagger for English and German, RFTagger for German morphology information
Additional resources and tools:
A bilingual dictionary German English with about 800 000 entries
A rule-based clause chunker
1

Slightly extended Earley chart parser for parsing synchronous grammars
Derived resources:
synchronous grammar productions
language models
token generation table

1
Clause chunkers are needed if structural divergences between SL and TL depend on clause type or clause boundaries.
4.1 Bilingual corpus English German
The bilingual corpus is the cross-linguistic heart of the MT system. The bilingual corpus is used to
define structural mappings from SL to TL. It should be representative in the sense that it should con-
tain the grammatical constructions that are expected to occur in the sentences to be translated.
The bilingual corpus can be text-type specific. So for example, the bilingual corpus used in
DE-EN is taken from a website of the EU (!""#$%%&'()#*+&'%*,-%./0&11)21%324&56&2+!"78. It out-
lines the history of the EU. It does not contain any direct questions or 1. or 2. person pronouns. It is
appropriate to take such a corpus if the MT system focuses on descriptive text. In this case it is even
advisable not to add direct questions to the bilingual corpus since it would unnecessarily complicate
the generation of appropriate TL structures.
4.2 Synchronous grammar generation
The limited size of the bilingual corpus does not allow for statistical methods for deriving synchron-
ous grammars as in Chiang 2007. Also, unlike other tree-to-tree translation approaches (Eisner 2003,
Cowan et al. 2006, Zhang et al 2007), the system proposed here does not use deep syntactic
processing of the corpus data. Instead, shallow parsers are adopted to annotate the small parallel bilin-
gual corpus. Before deriving synchronous grammar productions, the bilingual corpus is aligned on
chunk and tag level using the phrase aligner developed by Tambouratzis et al 2011. The alignments
are then converted into productions. In order to extend the sentence patterns covered by the corpus
alignments several strategies are employed:
The sentential chunk and tag alignments are broken down into the smallest self-contained align-
ments. These are then converted into productions. Consider the following schematic example in which
the alignment of B can also be taken as an abstract representation of the translation of English simple
verbs into German separable prefix verbs which is a non-local phenomenon that poses problems for
statistical MT:
2

SL: A B A C D

TL: A B A C B D
The self-contained alignments are converted into productions. Since the alignments of A and D are
self-contained one-to-one alignments they are turned into unary productions. Only the complex
alignment of B affords a more complex production. Here, some linguistic insight is fed into the pro-
duction generation. The chunks intervening between the split chunk B are replaced by a clause node
CL. Thus the derived production covers more sentence types than the one found in the corpus. The
format of the productions is: SL rule TL rule. CL is also introduced as mother node for the pro-
ductions that express chunk alignments. For each production a recursive and a non-recursive variant is
generated. In the following only the recursive variants are listed.

CL1 A2 CL3 CL1 A2 CL3
CL1 C2 CL3 CL1 C2 CL3
CL1 D2 CL1 D2
CL1 B2 CL3 CL1 B2 CL3 B2
Another way to extend the coverage of the productions beyond the patterns found in the corpus is
to define equivalence classes of tags and to multiply templates according to those equivalence classes.
For example all finite tags form an equivalence class. Thus if a production has been generated for the
finite tag 3.Pl.Pres, the corresponding productions for all other person, number and tense specifica-
tions are automatically generated. Another equivalence class consists of different noun tags for names
and regular nouns in singular and plural form. And the current system also treats NP and PP chunks as
mother nodes in productions as equivalence class.
3


2
A corresponding natural language sentences would be:
They accepted it immediately.
Sie nahmen es sofort an.
they accepted it immediately SEPPREF
3
The TL tags and chunks corresponding to the SL tags and chunks are also specified in the equivalence classes.
In order to cut down the number of TL structures produced by the productions, ambiguities in tag
alignments are not spelled out in different productions but represented as a local TL tag disjunction.
Tag alignment ambiguities arise if SL and TL taggers assign tags with different granularity. E.g. the
English TreeTagger assigns IN to both prepositions and subordinate conjunctions whereas the Ger-
man TreeTagger assigns APPR and APPRART to prepositions and KOUS to subordinate conjunc-
tions. The following recursive production accounts for three tag alignments, namely SLT1 TLT1,
SLT1 TLT2 and SLT1 TLT3. The disjunctive TL tags are separated by the pipe symbol.
A1 SLT12 A3 A1 TLT1|TLT2|TLT32 A3
Which TL tag is suitable in a given translation is determined by a lookup in the token generation table
which contains for each TL lemma also all possible tags. If this lookup is not decisive then the lan-
guage models will further disambiguate.
5 Evaluation
The preliminary version of the PRESEMT system has been evaluated using a test set of 50 sentences.
In a first run, the PRESEMT prototype has achieved 65% of the NIST scores of Google translate
(http://translate.google.de/#) and 40% of the BLEU scores of Google translate. The Moses
(http://www.statmt.org/moses/?n=Public.Demos) scores are in between the PRESEMT scores and the
Google scores.
A manual evaluation of the translations has shown that the lemma disambiguation often produces sub-
optimal results. Therefore, it is expected that replacing the simple 3-gram models with more sophisti-
cated language models will considerably improve the test scores. This is planned for the future.
6 Extending to new language pairs
Since even the modules that employ non-statistical methods such as the synchronous grammar parser
use only automatically derived resources it is relatively easy to extend the PRESEMT system to new
language pairs. Work on DE-EN has already started. Other language pairs that are planned to be in-
cluded in the near future are: Greek, Norwegian and Czech as SL and English and German as TL.
References
Koehn, Philipp, Abhishek Arun, and Hieu Hoang. 2008. Towards better Machine Translation quality for the
German English Language pairs. Proceedings of the Third Workshop on Statistical Machine Translation,
Association for Computational Linguistics. 139-142.
Chiang, David. 2007. Hierarchical phrase-based translation. Computational Linguistics 33(2):201228.
Cowan, Brooke, Ivona Kucerova, and Michael Collins. 2006. A discriminative model for tree-to-tree translation.
Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing. 232-241
Earley, Jay 1968. An efficient context-free parsing algorithm. Ph.D. thesis, Carnegie Mellon University,
Pittsburg, PA.
Eisner, Jason 2003. Learning non-isomorphic tree mappings for machine translation. Proceedings of the 41st
Annual Meeting of the Association for Computational Linguistics.
Schmid, Helmut. 1994. Probabilistic Part-of-Speech Tagging Using Decision Trees. Proceedings of the Interna-
tional Conference on New Methods in Language Processing. 44-49.
Schmid, Helmut and Florian Laws. 2008. Estimation of Conditional Probabilities with Decision Trees and an
Application to Fine-Grained POS Tagging. COLING 2008, Manchester, Great Britain.
Tambouratzis, George; Sofianopoulos, Sokratis; Vassiliou, Marina; Simistira, Fotini; Tsimboukakis, Nikos
2011. A resource-light phrase scheme for language-portable MT. Proceedings of the 15th International Con-
ferecne of the European Association for Machine Translation. 185-192.
Zhang, Min, Hongfei Jiang, Ai Ti Aw, Jun Sun, Sheng Li, and Chew Lim Tan. 2007. A Tree-to-Tree Align-
ment-based Model for Statistical Machine Translation. MT-Summit-07. 535-542.
Phrase Table Support
for Human Translation
Gerhard Kremer Matthias Hartung Sebastian Pad o Stefan Riezler
Institute of Computational Linguistics (ICL), University of Heidelberg
Selecting an appropriate translation for a word in context is a difcult task for
humans, and the support offered by bilingual dictionaries is unprincipled and spotty.
This paper reports on an ongoing study that aims at testing how non-professional
translators can prot directly from the data in phrase tables generated from large
parallel corpora for the purpose of machine translation. We present the design of an
experiment in which human translators were asked to translate adjectivenoun pairs
in context while being supported with different types of information extracted from
phrase tables. Our hypothesis is that bigram information from phrase tables will lead
to faster and more accurate translation.
1 Introduction
Translating a sentence adequately from one language into another is a difcult task for humans.
One of its most demanding subtasks is to select, for each source word, the best out of many
possible alternative translations. This subtask is known, in particular in computational contexts, as
lexical choice or lexical selection (Wu and Palmer, 1994). Bilingual lexicons which are commonly
used by human translators contain by no means all information that is necessary for adequate
lexical choice, which is often determined to a large degree by context. Often, dictionaries merely
list a small number of translation alternatives, or a small set of particularly prototypical contexts
is provided. The provided translations are neither exhaustive, nor do they provide distinguishing
information on which contexts they require.
In this study, we ask whether the shortcomings of traditional dictionaries can be evaded by using
a data structure used in most current machine translation (MT) systems, namely phrase tables (cf.
Koehn, 2010b). Phrase tables are merely bilingual lists of corresponding word sequences observed
in parallel corpora, and thus provide a compact representation of the translation information
inherent in a corpus, complemented with statistical information about the correspondences (e. g.,
frequencies or association measures). Phrase tables can potentially provide both smaller and
larger contexts surrounding a particular target word (i. e., context size can be adapted to specic
needs of a translator), but they are not prepared for easy interpretation by human translators.
1
We aim to investigate how phrase tables can be presented to translators for faster and better
translation. We approached this question through an experiment in which users had to solve a
translation task. They were presented with different types of phrase table information, and we
compare the efcacy of different modes of presenting the information.
To keep the experiment manageable, the current study focuses on one particular construction,
namely the translation of adjectives in attributive position (preceding a noun). Adjectives are
known to be highly context-adaptive in that they express different meanings depending on the noun
they modify (Sapir, 1944; Justeson and Katz, 1995). Second, adjectives tend to take on gurative
or idiomatic interpretations, again depending on the semantics of the noun in context (Miller,
1998). Lexical choice is therefore nontrivial, and context-dependent translations are seldom given
systematically in dictionaries. For example, consider the adjective heavy. In noun contexts like
use, trafc, and investment, its canonical translation as German schwer is inappropriate. It might
be translated as intensiv(e Nutzung), stark(er Verkehr), and gro(e Investition).
After presenting related work, section 3 describes in more detail the experimental setup that
we have designed, the data, and our hypotheses.
2 Related Work
Interactive MT systems aim to aid human translators by embedding MT systems into the human
translation process. Several types of assistance by MT systems have been presented: Translation
memories provide translations of phrases recurring during a project, but they have to be provided
by the translator the rst time they appear, and they are typically restricted to a document, a
project, or a domain (cf. Zanettin, 2002; Freigang, 1998). In the TransType system of Langlais
et al. (2000), the machine translation component makes sentence completion predictions based on
the decoders search graph. The interactive tool is able to deal with human translations that diverge
from the MT systems suggestions by computing an approximate match in the search graph and
using this as trigger for new predictions (Barrachina et al., 2008). Other types of assistance
integrate the phrase tables of the MT systems more directly: Koehn and Haddow (2009) and
Koehn (2010a) deploy a phrase-based MT system to display word or phrase translation options
alongside the input words, ranked according to the decoders cost model. Finally, full-sentence
translations can be supplied for post-editing by the user.
While the above cited previous work could show a signicant increase in productivity and
quality for machine-assisted translation, especially for less qualied translators, the presented
experiments allow only for a weak correlation between translation times and translation quality.
This is due to the varying complexity of test examples and the varying degree of expertise of
human translators. In our experiments we tried to control the variable of translation complexity
by restricting the task to adjectivenoun pairs of roughly the same ambiguity rate and providing
machine assistance for these pairs only. Furthermore, the human translators in our experiments
were all native speakers of the target language with a similar educational background (regarding
experience in this projects source language English). The goal of our pilot experiment is to
provide a basis for re-interpretation of results by using a clear and simple experimental design
which allows us to analyse the contribution of each variable.
2
Table 1: Partitions of the set of 30 adjective stimuli presented to each participant for the factors
variability and support. Factor context: Each adjective was embedded into 1 out of 4
sentences, where each adjective occurs with a different adjacent noun.
Variability Class Translation Support Condition Noun Context
none adjective unigrams adjectivenoun bigrams
high 5 5 5

4
low 5 5 5
3 Experimental Setup
We conducted series of two experiments. In the rst one (the main experiment), participants
performed a translation task with different kinds of supporting information. In order to test the
impact of presenting phrase tables on translation speed, we measured several time points during
each of the participants translation tasks, using time gain/loss
1
as a measure for the usefulness of
machine-aided human translation (as discussed in Gow, 2003).
The goal of the second experiment is to complement the time aspect with a measure of the
translations quality.
2
For this purpose, we collected human judgements for all translations from
experiment 1 on a simple 3-point scale. We do not believe (at least not a priori) that this step can
be automated through the use of MT evaluation strategies like BLEU (see Papineni et al., 2002) or
edit distance (discussed in Gow, 2003), given the restricted phenomenon and the semantic nature
of the distinctions that we focus on.
In this abstract, we concentrate on describing the setup of the rst experiment. Participants were
asked to translate an attributive adjective in sentential context, given one of our set of translation
support types. With German participants, we investigated translation from English into German,
the participants native language. This is the preferred type of translation direction in professional
human translation as the translators experience of commonly used words in a particular semantic
context is more extensive in the native language. In this experiment we assumed three factors to
interact with translation speed and accuracy (cf. table 1): variability class (2 levels), translation
support (3 conditions), and noun context (4 sentences per adjective, each sentence with a different
adjacent noun).
Variability classes. Stimuli for the translation experiment have been collected by examining
the most frequent adjectives from the British National Corpus (BNC), many of which are polyse-
mous, i. e., showing high context-dependent variability in translation (cf. section 1). An analysis
with 200 high-frequent adjectives in the BNC showed a highly signicant correlation (Spearmans
=0.5121) between corpus frequency and variability in translation (operationalised as the number
of unique translations in the EUROPARL (Koehn, 2005) v6 phrase table). We divided adjectives
1
The response times might vary a lot depending on differing translation habits of the participants: One person might
select the rst cognitively available lexical items, while other persons might take their time to consider alternatives.
We plan to modify experiment instructions to equalise translation strategies among participants.
2
Note that there have been ongoing debates on how translation quality can be assessed objectively (cf. House, 1998).
3
into two classes: one set that shows a particularly high variability in unique translations, and one
set with a relatively low translation variability.
Hypothesis. Highly variable adjectives are more difcult to translate, but translators will prot
more from the presentation of phrase table information.
Adjectives and Contexts. For each variability class, we selected 15 adjectives according to
the phrase table. For each English adjective, we randomly sampled four full sentences from the
BNC (with the adjective in attributive position). In order to minimise variation in translation times,
we have restricted the length of sentences to a dened range of number of words and characters (so
that reading times are comparable). Note that our setup results in a domain difference between the
sentences to be translated (sampled from the BNC) and the phrase table (drawn from EUROPARL)
a standard situation for translation.
Hypothesis. We expect speed differences among adjectives and noun contexts, but will treat them
as random effects (cf. section 4).
Translation Support. Finally, we provided three kinds of translation support to participants:
(a) no support, (b) the list of translations for the adjective unigram from the phrase table, and
(c) the list of translations for the adjectivenoun bigram from the phrase table. We presented three
distinct candidate translations as support, ranked by their order in the n-best list produced by the
moses
3
MT system (trained and tuned on EUROPARL v6) that decoded each target sentence.
Hypothesis. Presenting unigram translations leads to faster and more appropriate translations.
Bigram phrases will produce the most appropriate translations, even if translating in this condition
might be slower due to the need to read through more complex translation suggestions.
4 Procedure and Evaluation
Participants were all required to be native German speakers with at least a working knowledge
of English (they were asked to specify their level of prociency). Each participant was asked to
translate all 30 adjectives, but each adjective in only one sentence context, in order to avoid faster
translation of previously seen target adjectives. Of the 30 adjectives, each set of 10 was presented
in one of the three translation support conditions. In summary, 85 persons participated.
Note that the design of our experiment does not conform to classic psychological or psycholin-
guistic expectations: it is not balanced and, with the exception of translational variability, we
did not control any variables regarding the sentence stimuli. Instead, our materials mirror the
distribution in the corpus. This is a conscious decision that we have taken because (a) there is a
very large number of potentially inuential factors which are very difcult to control; and (b) we
are interested in testing our hypothesis under practical rather than idealised conditions.
In order to quantify the inuence of the individual factors and test the hypotheses formulated
above, we analysed the reading times with a mixed effects model (see, e. g., Baayen et al., 2008).
We treated the variability class and the translation support as xed effects, and the identity of
adjective, context, and participant as random effects. On the quality judgement data we (measured
the inter-rater correlation and) performed an analysis of variance for our experiment conditions.
3
URL http://www.statmt.org/moses
4
References
Baayen, R. H., Davidson, D. J., and Bates, D. M. (2008). Mixed effects modeling with crossed random
effects for subjects and items. Journal of Memory and Language, 59:390412.
Barrachina, S., Bender, O., Casacuberta, F., Civera, J., Cubel, E., Kadivi, S., Lagarda, A., Ney, H.,
Thomas, J., Vidal, E., and Vilar, J.-M. (2008). Statistical approaches to computer-assisted translation.
Computational Linguistics, 35(1):328.
Freigang, K.-H. (1998). Machine-aided translation. In Baker, M., editor, Routledge Encyclopedia of
Translation Studies, pages 134139. Routledge, New York.
Gow, F. (2003). Metrics for evaluating translation memory software. Masters thesis, University of Ottawa.
House, J. (1998). Quality of translation. In Baker, M., editor, Routledge Encyclopedia of Translation
Studies, pages 197200. Routledge, New York.
Justeson, J. S. and Katz, S. M. (1995). Principled disambiguation. Discriminating adjective senses with
modied nouns. Computational Linguistics, 21:127.
Koehn, P. (2005). Europarl: A parallel corpus for statistical machine translation. In Proceedings of the
Tenth Machine Translation Summit, pages 7986. Asia-Pacic Association for Machine Translation
(AAMT).
Koehn, P. (2010a). Enabling monolingual translators: Post-editing vs. options. In Proceedings of the
Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the
ACL, Los Angeles, CA.
Koehn, P. (2010b). Statistical Machine Translation. Cambridge University Press.
Koehn, P. and Haddow, B. (2009). Interactive assistance to human translators using statistical machine
translation methods. In Proceedings of Machine Translation Summit XII, Ottawa, Ontario, Canada.
Langlais, P., Foster, G., and Lapalme, G. (2000). TransType: A computer-aided translation typing system.
In Proceedings of ANLP-NAACL Workshop on Embedded Machine Translation Systems, Seattle, WA.
Miller, K. J. (1998). Modiers in WordNet. In Fellbaum, C., editor, WordNet. An Electronic Lexical
Database, pages 4767. MIT Press.
Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J. (2002). BLEU: A method for automatic evaluation
of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computa-
tional Linguistics, ACL 02, pages 311318, Stroudsburg, PA, USA. Association for Computational
Linguistics.
Sapir, E. (1944). Grading. A study in semantics. Philosophy of Sciences, 11:83116.
Wu, Z. and Palmer, M. (1994). Verbs semantics and lexical selection. In Proceedings of the 32nd annual
meeting on Association for Computational Linguistics, pages 133138.
Zanettin, F. (2002). Corpora in translation practice. In Yuste-Rodrigo, E., editor, Proceedings of the
Workshop Language Resources for Translation Work and Research at the International Conference on
Language Resources and Evaluation (LREC), pages 1014, Las Palmas de Gran Canaria. LREC.
5
Inside the Monitor ModeI: Processes of DefauIt and
ChaIIenged TransIation Production
Michael Carl and Barbara Dragsted
ISV, Copenhagen Business School, Dalgas Have 15, DK-2000 Frederiksberg
Abstract
It has been the subject oI debate in the translation process literature whether human translation is a sequential
and iterative process oI comprehension-transIer-production or whether and to what extent comprehension
and production activities may occur in parallel. Tirkkonen-Condit (2005) suggests a 'monitor model
according to which translators start with a literal default rendering procedure where a monitor interrupts the
deIault procedure when problems occurs. This paper suggests an extension oI the monitor model in which
comprehension and production are processed in parallel by the deIault procedure. Deviations Irom this
deIault behaviour are triggered through text production problems and involve conscious decision-making
processes, related to text comprehension or to text production problems.
In an experiment we compare text copying with translation activities under the assumption that text copying
is a prototypical literal deIault rendering procedure. Both tasks, translation and text copying, require
decoding, retrieval and encoding oI textual segments, but translation additionally requires transIer into
another language. Comparing user behaviour obtained in copying and translation experiments, we observe
surprisingly many similarities between the two diIIerent activities. Copyists deviate Irom the deIault literal
text reproduction into more eIIortIul text understanding, and much oI the translators` behaviour resembles
that oI copyists. We discuss how extended source text (ST) comprehension is triggered through production
problems, during translation as well as during text copying.
The stratificationaI modeI of transIation production
According to the eye-mind hypothesis (Just & Carpenter, 1984)
1
there is a strong correlation between where
one is looking and what one is thinking about. The eye-mind hypothesis is controversial; on the one hand it
is well known that the observation oI a correlation between two events does not imply a causal relation, and
hence no conclusion can be made regarding the existence or the direction oI cause and eIIect only Irom the
observation that two events, e.g. gaze location and mental processes, correlate. On the other hand, the strong
correlation assumption between gaze and mind has be questioned, e.g. by (John R. Anderson, Dan Bothell,
and Scott Douglass, 2004) who Iind that longer gaze durations do not correlate with greater problems oI
memory retrieval.
In Translation Process Research it has oIten been stated that gaze location reIlects the Iocus oI attention oI
the translator (e.g. Hyrskykari, 2006). That is, when the gaze Iocusses on the ST the mind is involved in ST
comprehension processes, and when the gaze is directed at the TT, the mind is involved in text production
processes. Longer gaze durations on the ST or TT reIlect bigger comprehension or production problems
respectively (Pavolovic and Jensen, 2009). These assumptions Iit well with a stratiIicational process model
oI translation, which states that at any one time the translator either reads (understands) the ST, transIers it
into the target language, or types the translation.
Craciunescu et al. (2004), Ior instance, claim that 'the Iirst stage in human translation is complete
comprehension oI the source language text. Only aIter this complete (i.e. deep) comprehension is achieved
can the translation be produced. Similarly Gile (2005) suggests a stratiIicational translation process model, in
which a translator iteratively reads a piece oI the ST and then produces its translation. First the translator
creates a 'Meaning Hypothesis Ior a ST chunk (i.e. a Translation Unit) which is consistent with the 'context
and the linguistic and extra linguistic knowledge oI the translator (p. 107). Subsequently, a translation is
produced.
Also Angelone (2010) supports that translators process in cycles oI comprehension-transIer-production and
that 'uncertainties oI translators can be attributed to any oI the comprehension, transIer, or production
phases. He claims that 'non-articulated indicators, such a pauses and eye-Iixations, give us no real clue as to
how and where to allocate the uncertainty |p.23|
! 'there is no appreciable lag between what is Iixated and what is processed"
The monitor modeI
Some scholars challenge this view, stating that translation processes can also be based on a shallow
understanding and that ST understanding and TT production can occur in parallel. According to Ruiz et al.
(2008) 'the translator engages in partial reIormulation while reading Ior the purpose oI translating the
source text. They assume that in parallel processing 'code-to-code links between the SL and TL |are
involved| at least at the lexical and syntactic level oI processing. Similarly, Mossop (2003) claims the
existence oI 'direct linkages in the mind between SL and TL lexicogrammatical material, independent oI
meaning`, and that a translator 'automatically produces TL lexical and syntactic material based on the
incoming SL Iorms.
In a study comparing reading behaviour Ior diIIerent purposes, Jakobsen & Jensen (2008) investigate (among
other things) the diIIerence between test persons reading a text Ior comprehension and reading a similar text
in preparation Ior translating. Their study showed that reading purpose has a 'clear eIIect on eye movements
and gaze behaviour and they suggest 'that a Iair amount oI pre-translation probably enters into the reading
oI a text as soon as it is taken to be the source text Ior translation |p.116|.
Although it is unclear what is exactly meant by 'pre-translation, such Iindings are obviously in contrast
with the eye-mind hypothesis when assuming a stratiIicational model oI translation. Reading with 'a Iair
amount oI pre-translation implies certainly diIIerent mental activities than reading Ior understanding, but
the eyes remain in both cases on the ST. Since it may be diIIicult (iI not impossible) to disentangle which
parts oI the gaze behaviour are to be linked to text understanding and which correspond to pre-translation,
either the eye-mind hypothesis has to be weakened or the stratiIicational model oI translation has to be
reconsidered.
We assume, with Tirkkonen-Condit, that 'literal translation is a deIault rendering procedure, which goes on
until it is interrupted by a monitor that alerts about a problem in the outcome. The monitor`s Iunction is to
trigger oII conscious decision-making to solve the problem (Tirkkonen-Condit 2005: 407-408). In our
interpretation oI the model, the literal deIault rendering procedure implies parallel, tightly interconnected text
production and comprehension processes: while the mind is engaged in the production oI a piece oI text, the
eyes search Ior relevant textual passages to gather the required inIormation needed to continue the text
production Ilow. When this deIault procedure is interrupted by the monitor, can we observe gaze patterns on
the ST or on the TT which indicate comprehension- or production-related translation problems. Note,
however, that these decision-making processes are triggered problems related to production activities.
Similarly, Gile (2005) reports that deeper understanding oI the ST may emerge through problems in TT
production, rather than when Iirst reading a ST passage. He points out that the translation practice indicates
processing Irom a production-based perspective:
OItentimes, the translator does not test Meaning Hypothesis until aIter verbalising it in the target
language (...) Frequently, he or she only realizes there is a problem when trying to read the Iirst
target-language version (...) in other words, when already in the reIormulation phase.
A clear-cut allocation oI 'uncertainties to one oI the stratiIicational processes then becomes diIIicult, since
such processes do not normally exist independently in the translator`s mind. Not only is it inIeasible (or
impossible) to distinguish between comprehension and pre-translation activities during reading Ior
translation, but also the borders between ST understanding and TT production problems become blurred.
Observations from transIation experiments
We investigated patterns oI typing behaviour Irom text copying and translation experiments. Our
investigation is based on empirical data obtained in 10 copying sessions and 15 translation sessions. The
experiments were recorded using the Translog 2006 soItware (Jakobsen and Schou, 1999), which logs
keystrokes and gaze movements during a reading, translation or text production task.
We take it that copying (i.e. re-typing) a text may be processed in a much more shallow/parallel manner than
translation since: 1) apart Irom a lexical encoding and decoding (John, 1999), text copying does not, in
theory, require any deep ST (or TT) understanding; 2) copying can proceed in parallel to a maximal degree,
since no revision
2
and no lexical or structural transIer is required. Typing patterns and speed would thus
essentially depend on the typing skills oI the copyist. Comparing copying behaviour and translation
2 some revision may be going on, Ior instance correction oI typos, but these activities are oI a diIIerent kind than most
oI those in translation revisions.
behaviour would reveal the additional eIIort oI translation.
We observe that most oI the text is copied smoothly and straight-Iorwardly, with only little look-ahead in the
ST. But we also observe that the copying activity may trigger extended reading activities in the ST context
when a text passage is unclear. That is, word meaning seems to be processed also during text copying. As
predicted in the monitor model (Tirkkonen-Condit, 2005), the copying pause occurred when typing the
unclear expression, rather than when reading it the Iirst time. Lack oI comprehension is discovered (or at
least actions are taken) only during (re)production oI the text, rather than during Iirst reading. We then
compared typing in a copying task with typing in a translation task and observed basically the same patterns.
While much oI the typing activity Ior translation resembles text copying in L2, the gaze is on average
slightly Iurther ahead in the ST during translation than when copying.
We also looked at passages oI conscious, eIIortIul text production, which uncovers more entangled relations
between comprehension and production. Similar to text copying, translation examples clearly show that
diIIiculties occur when reIormulating (render and address) the translation rather than when reading the ST.
The translation oI a phrase may already start beIore the translator knows how to go on with the translation.
The initial translation guess is not always appropriate, and sometimes the beginning oI the phrase must be
revised. In some cases the ST context has to be re-consulted and in other cases the produced TT is
reconsidered.
Inter-key time spent during unchallenged production was approximately the same in copying and in
translation during periods oI coherent typing. In addition we observed parallel and alternating reading and
typing behaviour, where reading and writing activities occurred respectively simultaneously or sequentially.
There are more pauses during translation than during copying, indicating more alternating processes in
translation.
Looking at gaze activities we Iound that the number oI ST word Iixations during parallel unchallenged
translation activities equalled approximately those oI unchallenged copying while there were more ST
Iixations during alternating translation activity.
ConcIusions
Two types oI translation behaviour can be distinguished:
!" Much oI the translation draIting is unproblematic and approximately within the time limits predicted
Ior text copying by Johns` (1999) TYPIST model
3
. Translators look only a Iew words ahead into the
ST Irom the position where they are currently translating. In an alternating mode, ST decoding adds
to the typing time, while in a parallel mode decoding and encoding run in parallel. Many oI the
smaller translation problems, such as multi-word translations or local reordering, may be solved by
looking only a Iew words ahead. We suspect that the degree oI parallel activity depends on
experience and typing skills oI the translator. A touch typist would more likely exhibit parallel
processing behaviour, while a translator with less developed typing skills would show alternating
translation patterns.
2. At some points in the translation, extensive reading behaviour can be observed, signalling more
serious translation problems. Depending on the type oI problem, it may be necessary Ior the
translator to re-scan the ST or the TT. In both cases, the increased reading activity seems to be
triggered by a TT production problem rather than by a ST comprehension problem. That is, we
observed that the ST was understood, and meaning hypotheses were generated only to the extent
required to keep on producing target text. II, Ior whatever reason, TT production cannot go on
smoothly, and the typing Ilow is interrupted, the missing inIormation needs to be retrieved. This may
lead to the re-reading oI a ST passage with a view to veriIication or reinterpretation, and/or the
revision oI the produced TT.
#n a stratiIicational comprehension-transIer-production theory oI translation, this behaviour is diIIicult to
explain. Ruiz et al. (2008) point out that 'comprehension Ior translation does not diIIer Irom normal
monolingual comprehension since comprehension and reIormulation occur at diIIerent stages . But iI the
ST would Iirst have to be completely understood beIore a translator could start translating it, why would the
translation purpose have an impact on the ST reading behaviour? Instead, we assume that 'Meaning
Hypotheses are constructed to the extent and at the moment they are needed to continue the task at hand.
DiIIerent meaning hypotheses are required Ior diIIerent kinds oI activities, e.g. a technician reading a car
3 This conclusion is based on our translation material Irom English into Danish, two relatively close languages with
similar word order.
repair manual needs a diIIerent kind oI understanding than a translator translating the same text into another
language. The reading purpose thus determines what kind and depth oI meaning representation is required.
During translation and text copying, the ST meaning is oIten not elaborated and tested until the writing
process which leads to the surprising conclusion that comprehension does not precede, but rather Iollow
text production.
An extended version oI this paper will appear in the journal 'Translation: Computation, Corpora, Cognition.
References
Anderson, John R.; Dan Bothell, and Scott Douglass, (2004) Eye Movements Do Not ReIlect Retrieval Processes
Limits oI the Eye-Mind Hypothesis, PSYCHOLOGICAL SCIENCE , Volume 15Number 4, 2004,
Angelone, Erik. 2010. 'Uncertainty, uncertainty management and metacognitive problem solving in the translation
task. In Translation and Cognition, Shreve, Gregory M. and Erik Angelone (eds.), 1740.
Carl, Michael and Martin Kay (2011) Gazing and Typing Activities during Translation: A Comparative Study oI
Translation Units oI ProIessional and Student Translators, Accepted Ior publication in META
Carl, M. and Jakobsen, A. L. (2010). Towards statistical modelling oI translators` activity data. International Journal oI
Speech Technology, 12(4), 124-146.
Craciunescu, Olivia; Constanza Gerding-Salas, Susan Stringer-O'KeeIIe (2004), Machine Translation and Computer-
Assisted Translation: a New Way oI Translating? Translation Journal, Volume 8, No. 3 , July 2004
http://translationjournal.net/journal/29computers.htm
Gile, Daniel (2005) La Traduction. La comprendre, l'apprendre. Paris: Presses Universitaires de France
Dragsted, B. (2010). 'Coordination oI reading and writing processes in translation: An eye on uncharted territory. In
G. M. Shreve, and E. Angelone (eds), Translation and Cognition, pp. 4162, Amsterdam/Philadephia, Benjamins
Hyrskykari, Aulikki 2006, Utilizing eye movements: Overcoming inaccuracy while tracking the Iocus oI attention
during reading , Attention aware systems - Special issue: Attention aware systems
Jakobsen, Arnt Lykke (1999): ,Logging target text production with Translog'. In Hansen, Gyde, Hrsg.: Probing the
Process in Translation: Methods and Results. (Copenhagen Studies in Language 24). Kopenhagen:
SamIundslitteratur. 920.
Jakobsen, A. L. and Schou, L. 1999. Logging target text production with Translog, In Copenhagen Studies in Language,
volume 24, SamIundslitteratur, Copenhagen, 9-20
Jakobsen, A. J. and Jensen, K. T. H. (2008) Eye movement behaviour across Iour diIIerent types oI reading task, In
GpIerich et al. (eds) Looking at Eyes. Eye Tracking Studies oI Reading and Translation Processing, CSL 36,
103-124.
John, Bonnie E.. Typist: a theory oI perIormance in skilled typing. Hum.-Comput. Interact., 11(4):321355, 1996.
Just, M.A., & Carpenter, P.A. (1984). Using eye Iixations to study reading comprehension. In D.E. Kieras & M.A. Just
(Eds.), New methods in reading comprehension research (pp. 151182). Hillsdale, NJ: Erlbaum.
Kromann, M. T. (2003). The Danish Dependency Treebank and the DTAG treebank tool. In Proceedings oI the Second
Workshop on Treebanks and Linguistic Theories (TLT 2003), 14-15, November, Vxj. 217220.
Mossop, Brian. An Alternative to Deverbalization`. Technical report, York University, 2003.
http://www.yorku.ca/brmossop/Deverbalization.htm
Pavlovic, Natasa, and Kristian T. H. Jensen. (2009). Eye tracking translation directionality. In Translation Research
Projects 2, eds. Anthony Pym and Alexander Perekrestenko. Tarragona: Intercultural Studies Group, p. 93-109.
Available at: http://isg.urv.es/publicity/isg/publications/trp22009/index.htm.
Perrin, Daniel. 2003. Progression analysis (PA): investigating writing strategies at the workplace. Pragmatics, 35:907
921.
Ruiz, C.; N Paredes, P Macizo, M T Bajo , (2008), Activation oI lexical and syntactic target language properties in
translation. Acta Psychologica, Volume: 128, Issue: 3, Pages: 490-500
Staub, A. and Rayner, K. 2007. Eye movements and on-line comprehension processes. In Gaskell, G. (ed.), The OxIord
Handbook oI Psycholinguistics. OxIord: OxIord University Press. 327-342.
Tirkkonen-Condit,Sonja (2005) The Monitor Model Revisited: Evidence Irom Process Research, META, Volume 50,
numero 2, avril 2005, p. 405-414