Logical Formalizations of Commonsense Reasoning: Papers from the 2015 AAAI Spring Symposium

Abduction and Conversational Implicature

(Extended Abstract)
Chiaki Sakama

Katsumi Inoue

Wakayama University
[email protected]

National Institute of Informatics

[email protected]

Introduction

mantically stronger sentence than what is actually uttered.

The clash between Q- and I-implicatures has been studied
in the eld of pragmatics that concerns with the meaning
of sentences in conversation (Horn 1984; Levinson 1987;
Carson 1995). Some researchers have pointed out the utility
of abduction in interpreting speech acts (Hobbs et al. 1993;
McRoy and Hirst 1993; Janc ek 2012). To the best of our
knowledge, however, relations or differences between abduction and conversational implicature have never been formally explored in the literature.
In this abstract we argue a formal account of abduction
and conversational implicature in human dialogues. We rst
formulate two different types of abduction in a propositional
modal logic. We next formulate conversational implicatures
and contrast them with abduction. The results characterize
how hearers use abduction or conversational implicatures to
gure out what speakers have implicated and show how two
commonsense inferences are distinguished.

In conversation or dialogue, people use abduction to understand reasons behind utterances. Suppose that your colleague says I will be at my ofce this weekend. Then
you may surmise that he/she has much work to do. In this
dialogue, the utterance is considered an evidence provided
by the colleague, then you seek reasons to explain the utterance. You abduce that your colleague would have much
work to do if you believe the implication much work
oce weekend . Given an utterance by a speaker, a hearer
could perform two different types of abduction. The rst
one is to produce a hearers belief of a fact that could explain an evidence provided by a speaker. The above example is of this type of abduction. The second one is to produce a hearers belief of a speakers belief which could explain the speakers utterance. In the above example, suppose
that you believe that the colleague believes much work
oce weekend . Then you abduce that the colleague believes that he/she has much work to do. In this case, however, you do not necessarily believe yourself that the colleague has much work to do. In this way, a hearer may use
abduction not only for generating assumption that accounts
for the utterance, but for generating assumption on the belief
state of a speaker who makes the utterance.
In conversation or dialogue, the notion of conversational
implicature (Grice 1975) is known as pragmatic inference
in linguistic phenomena. The principal subject is to investigate the meaning of a sentence more than what is actually said. For instance, if a speaker utters the sentence I
have two children, it normally implicates I do not have
more than two children. This is called a scalar implicature (or Q-implicature) which says that a speaker implicates the negation of a semantically stronger proposition than the one asserted (Levinson 1987). Generally, a
hearer infers from the utterance and the implication by Q-implicature. This is in contrast with
abduction in dialogue, however. In abduction, a hearer infers from the utterance and the implication .
Thus, two inferences appear to reach opposite conclusions
in face of an utterance. In conversational implicature, an alternative implicature, called the I-implicature, implies a se-

Abduction in Dialogue
We use a propositional modal logic of knowledge and belief
that is standard in the literature. A sentence Ba is read as
an agent a believes a sentence , Ka is read as a knows
, and C is read as it is common knowledge that .1 It
holds that Ka Ba , C Ka Ka C, etc. The
logic has the axioms and inference rules of the system KD45
for Ba and S5 for Ka . We assume a dialogue between two
agents, called a speaker and a hearer. Each agent has a (consistent) set of sentences as background knowledge and believes those sentences. When a speaker utters a (consistent)
sentence, a hearer performs abduction to explain reasons behind the utterance.
Denition 1 (objective abduction) Let a be a hearer and b
a speaker. When b utters a sentence ( ), a sentence is
inferred by objective abduction (O-abduction) from by a
if
(1)
Ba Ba ( ) Ba .
In this case, is called an O-explanation of . We write
O-abda (, ) if is an O-explanation of by a.
(1) means that a hearer a believes the utterance , and a believes the implication and disbelieves in his/her

c 2015, Association for the Advancement of Articial

Copyright 
Intelligence (www.aaai.org). All rights reserved.

130

For the precise denition of C, see (Fagin et al. 1995).

Conversational Implicature

background knowledge. In this case, a hearer a infers as

an explanation for the utterance . It is called objective
abduction because abduction is performed based on the objective fact of an utterance. Note that every explanation is
consistent (  ). If , then Ba in (1) is false.
There exist multiple explanations for an utterance in
general, and a hearer selects best explanations to be believed.
We do not address the issue further in this paper.

Conversational implicature (Grice 1975) is a pragmatic inference to an implicit meaning of a sentence that is not actually uttered by a speaker. In his maxims of conversation,
Grice introduces two maxims of quantity:
1. Make your contribution as informative as is required (for
the current purposes of the exchange).
2. Do not make your contribution more informative than is
required.

Denition 2 (subjective abduction) Let a be a hearer and

b a speaker. When b utters a sentence ( ), a sentence
Bb is inferred by subjective abduction (S-abduction) from
by a if
(2)
Ba Bb Ba Bb ( ) Ba Bb .
In this case, Bb is called an S-explanation of . We write
S-abdab (, ) if Bb is an S-explanation of by a.

Based on these two maxims, two principles are introduced

from the speakers viewpoint (Horn 1984; Levinson 1987):2
Q-principle: Say as much as you can.
I-principle: Say no more than you must.
These two principles correspond to the next implicatures
from the hearers viewpoint.

(2) means that a hearer a believes that a speaker b believes

his/her utterance , and a believes that b believes the implication , and a disbelieves that b disbelieves . In
this case, a hearer a infers Bb as an explanation for the
utterance . It is called subjective abduction because abduction is performed based on the hearers subjective view
on the speakers belief state.
In objective abduction, a hearer may believe an Oexplanation which accounts for an utterance by a
speaker. In subjective abduction, on the other hand, a hearer
may believe Bb but does not necessarily believe by
himself/herself. A connection between O-abduction and Sabduction is as follows.

Q-implicature: Imply the negation of a semantically

stronger sentence than what is actually uttered.
I-implicature: Imply a semantically stronger (or more specic) sentence than what is actually uttered.
These two implicatures apparently conict because the
Q-implicatures induce the negation of the very sort of
stronger interpretation that the I-implicatures actually appear to be promoting (Levinson 1987). Various attempts to
resolve the clash are proposed in pragmatics (Horn 1984;
Levinson 1987; Carson 1995). In what follows, we formulate these two implicatures in our logic.
Denition 3 (Q-implicature) Let a be a hearer and b a
speaker. When b utters a sentence ( ), a sentence Bb
is inferred by Q-implicature from by a if
(3)
Ba Bb C ( ) Ba Bb .

Proposition 1 S-abdab (, ) implies O-abda (Bb , Bb ).

and
The
differences
between
S-abdab (, )
O-abda (Bb , Bb ) are twofold. First, a speaker b utters a sentence in S-abdab (, ), while b utters his/her
belief Bb in O-abda (Bb , Bb ). Second, a hearer a
believes that a speaker b believes the implication
in S-abdab (, ), while a believes a weaker implication
Bb Bb (if a speaker b believes then b believes )
in O-abda (Bb , Bb ). One could also use O-abduction
to abduce the belief state of a hearer by O-abda (, Bb )
if Ba Ba (Bb ) Ba Bb . In this case, a
speaker utters a sentence and a hearer abduces the belief
state of the speaker which explains the utterance. All
S-abdab (, ), O-abda (Bb , Bb ) and O-abda (, Bb )
abduce the belief state of a speaker in different ways.
Suppose that a hearer a infers an O-explanation or an
S-explanation Bb in face of a speaker bs utterance . If
it is in fact or Bb , the hearer successfully understands
a reason behind the utterance . On the other hand, if it is
in fact or Bb , the hearer misunderstands a reason behind the utterance. In the introductory example, a colleague
says I will be at my ofce this weekend and you conjecture he/she has much work to do by O-abduction. If you
say You seem to have much work to do in response to the
colleague, and he/she says Not at all, I just come to my
ofce to surf the net, then you realize your incorrect abduction. As such, abduction is nonmonotonic in the sense that
an explanation might be withdrawn if it turns out incorrect.

We write Q-impab (, ) if Bb is inferred by Qimplicature from by a.

(3) is explained as follows. First, a hearer a believes that
a speaker b believes his/her utterance . Otherwise, it would
be meaningless to infer the implicit meaning behind the
utterance. Second, is not just a private belief of
the hearer, but a common knowledge that is shared by the
speaker and the hearer. Third, the hearer disbelieves that
the speaker believes a sentence which is stronger than
. In this case, the hearer infers that the speaker believes
. There exist multiple sentences satisfying (3) in general, and a hearer selects an appropriate one from them in
the context. The I-implicature is dened in a similar way.
Denition 4 (I-implicature) Let a be a hearer and b a
speaker. When b utters a sentence ( ), a sentence Bb
is inferred by I-implicature from by a if
(4)
Ba Bb C ( ) Ba Bb .
We write I-impab (, ) if Bb is inferred by I-implicature
from by a.
In both (3) and (4), conversational implicature is based
on common knowledge, that is, both a speaker and a hearer
2
Q means quantity and I means informativeness. The Iprinciple is called the R-principle (relevance) in (Horn 1984).

131

know the truth of the implication , and each one also

knows that the other party knows the truth of the sentence.
The reason of using common knowledge here is explained as
follows. In Q-implicature (resp. I-implicature), a hearer believes that a speaker implies (resp. ) by an utterance .
In this case, the hearer knows the implication and, at
the same time, the hearer knows that the speaker knows the
same implication. If the hearer does not know whether or not
the speaker knows the implication, then the hearer cannot
infer the intended meaning of the speakers utterance. Conversely, a speaker implies (resp. ) by an utterance in
terms of his/her knowledge of . If the speaker does
not know whether or not the hearer knows the implication,
then the speaker cannot expect the hearers reasoning by Qimplicature (resp. I-implicature). So if the speaker utters ,
he/she knows that the hearer knows the implication .
Thus conversational implicature is in effect if and only if
a speaker and a hearer share the same knowledge and each
one knows that the other party also shares the same knowledge. 3 The difference between two implicatures comes from
a hearers belief of whether a speaker believes a stronger
sentence or not. Given an utterance , if a hearer disbelieves
that a speaker believes a stronger sentence , then the hearer
interprets that the utterance Q-implicates . By contrast,
if a hearer disbelieves that a speaker disbelieves a stronger
sentence , then the hearer interprets that the utterance Iimplicates . Since Bb Bb , the conclusions derived by two implicatures contradict with each other. On the
other hand, it may happen that Ba Bb Ba Bb , so
it is a hearers option to decide which implicature is to be
applied in the context where Ba Bb C ( ) holds.

for the purpose of abduction when one observes wet grass.

On the other hand, it is inappropriate to use ones private
belief to infer conversational implicature. A hearers reasoning based on his/her private belief does not always reect
a speakers intention. Formally, the following relation holds
between S-abduction and I-implication.
Proposition 2 I-impab (, ) implies S-abdab (, ).
With the result of Proposition 1, I-impab (, ) also implies O-abda (Bb , Bb ). As mentioned above, one can
use common knowledge for the purpose of abduction. This
means that a hearer may reach to opposite conclusions by using S-abduction and Q-implicature. Suppose that Ba Bb
C ( ) holds. In this case, if Ba Bb then Bb is
inferred by S-abduction. Else if Ba Bb then Bb is inferred by Q-implicature. For instance, suppose a dialogue
between a mother (a) and her son (b): a: How was your
math exam? b: I could not solve one question. The belief state of mother is represented by Ba Bb C ( ),
where =the son does not solve one question and =the
son does not solve more than one question. Then mother
would infer that her son believes by S-abduction if she
disbelieves that her son disbelieves (i.e., Ba Bb ). Else
if she disbelieves that her son believes (i.e., Ba Bb ),
mother would infer that her son believes =he does not
solve at most one question by Q-implicature.
A hearer believes a speakers utterance (Ba ) in Oabduction, and believes that a speaker believes his/her utterance (Ba Bb ) in S-abduction. What happens if a hearer
believes the falsity of the utterance by a speaker? Suppose
a dialogue in which a speaker (b) utters but a hearer (a)
believes . In this case, it holds that Ba Ba (
) Ba , and the hearer believes if he/she believes
. Likewise, it holds that Ba Bb Ba Bb ( )
Ba Bb . So if a hearer a believes that a speaker b believes
the implication , and the hearer also believes that the
speaker is lying, i.e., the hearer believes that the speaker believes the falsity of his/her utterance , then a believes that
b believes . In this way, when a hearer believes the falsity
of an utterance , the hearer would believe the negation of a
sentence that explains . Moreover, if a hearer believes that
a speaker is lying, then the hearer could infer reasons behind the act (a speaker lies to make a hearer believe by
abduction, but in fact the speaker believes ). In conversational implicature, if a hearer a believes that a speaker b is
lying, it holds that Ba Bb C( ) Ba Bb . So a
hearer does not believe the conclusion Bb of I-implicature
while still believes the conclusion Bb of Q-implicature.
For instance, if a speaker says I have one million dollars
and a hearer does not believe it, the hearer does not infer that
the speaker has at least one million dollars by I-implicature
but still infers that the speaker does not have more than one
million dollars by Q-implicatue. This is an interesting contrast between two implicatures. From a speakers viewpoint,
a speaker will decide what to say by considering the effect
of his/her utterance on the hearers side. A speaker may use
this to mislead a hearer to reach a wrong assumption. Further discussion on this topic will be done in the full version
of this paper.

Abduction vs. Conversational Implicature

Both abduction and conversational implicature infer information behind an utterance. According to our formulation,
an essential difference between the two lies in the use of
implication . In abduction, it is a hearers private belief: a hearer believes in O-abduction while a hearer
believes that a speaker believes in S-abduction.
This is because abduction is a process of private reasoning, and one can reason abductively without knowing the
belief state of the other party. By contrast, in conversational implicature the implication is common knowledge: a
speaker and a hearer share the same implication. This is because conversation aims at communicating information. Examples of implication as common knowledge include valid
sentences (e.g. p q p), elementary arithmetic relations
(e.g. (x n + 1) (x n)), and class-subclass relationships (e.g. dog(x) animal(x)). By contrast, the implication much work oce weekend in the introduction is an example of private belief. Since C Ba and
C Ba Bb , one may use common knowledge for the
purpose of abduction. For instance, if it rains then the grass
is wet (rain wet grass) is considered common knowledge (or common belief), then one may use this knowledge
3

One may consider a weaker denition of implicature using

common belief instead of common knowledge, but we do not
address such alternative denitions here.

132

References
Carson, R. 1995. Quantity maxims and generalized implicature. Lingua 96:213244.
Fagin, R., Halpern, J. Y., Moses, Y. and Vardi M. Y. 1995.
Reasoning about Knowledge. MIT Press.
Grice, H. P. 1975. Logic and conversation. In: P. Cole &
J. Morgan (ed.), Syntax and Semantics, 3: Speech Acts, pp.
4158, Academic Press.
Hobbs, J. R., Stickel, M. E., Appelt, E. E. and Martin, P.
1993. Interpretation as abduction. Articial Intelligence
63:69142.
Horn, L. R. 1984. Toward a new taxonomy for pragmatic inference: Q-based and R-based implicature. In: D. Schiffrin
(ed.), Meaning, Form and Use in Context: Linguistic Application, pp. 1142, Georgetown University Press.
Janc ek, M. 2012. Abductive reasoning for continual dialogue understanding. New Directions in Logic, Language
and Computation, Lecture Notes in Computer Science 7415,
pp. 1631, Springer.
Levinson, S. C. 1987. Minimization and conversational inference. In: J. Verchueren, M. Bertuccelli-Papi (eds.), The
Pragmatic Perspective, pp. 61129, Benjamins Publishing.
McRoy, S. and Hirst, G. 1993. Abductive explanation of dialogue misunderstandings. In: Proceedings of the sixth conference on European chapter of the Association for Computational Linguistics, pp. 277286.

133