A O L M: R D O: N Verview On Anguage Odels Ecent Evelopments and Utlook

A N OVERVIEW ON L ANGUAGE M ODELS :
R ECENT D EVELOPMENTS AND O UTLOOK
Chengwei Wei1 , Yun-Cheng Wang1 , Bin Wang2 , and C.-C. Jay Kuo1
1
University of Southern California, Los Angeles, California, USA
arXiv:2303.05759v1 [cs.CL] 10 Mar 2023
2
National University of Singapore, Singapore
[email protected]
A BSTRACT
Language modeling studies the probability distributions over strings of texts. It is one of the most
fundamental tasks in natural language processing (NLP). It has been widely used in text generation,
speech recognition, machine translation, etc. Conventional language models (CLMs) aim to predict
the probability of linguistic sequences in a causal manner. In contrast, pre-trained language models
(PLMs) cover broader concepts and can be used in both causal sequential modeling and fine-tuning
for downstream applications. PLMs have their own training paradigms (usually self-supervised) and
serve as foundation models in modern NLP systems. This overview paper provides an introduction to
both CLMs and PLMs from five aspects, i.e., linguistic units, structures, training methods, evaluation
methods, and applications. Furthermore, we discuss the relationship between CLMs and PLMs and
shed light on the future directions of language modeling in the pre-trained era.
Keywords Language model, Natural language processing, Pre-trained language model, Conventional language model.
1 Introduction
Language modeling studies the probability distributions over a sequence of words. It is one of the most fundamental
tasks and long-standing research topics in natural language processing (NLP). The developed language models (LMs)
find applications in many computational linguistic problems such as text generation, machine translation, speech
recognition, natural language generation, question-and-answer systems, etc.
There are two major approaches to language modeling: 1) the statistical approach based on a relatively small corpus
set, and 2) the data-driven approach based on a significantly larger corpus set. Conventional language models (CLMs)
predict the probability of linguistic sequences in a causal manner. They can be learned by both language modeling
approaches. The data-driven approach has become mainstream nowadays. It exploits a large number of corpora to train
neural-network models, leading to pre-trained language models (PLMs). PLMs are then fine-tuned with task-specific
datasets and objectives for downstream applications. In this paper, we provide an overview of CLMs and PLMs and
study them from five perspectives: 1) linguistic units, 2) structures, 3) training methods, 4) evaluation methods, and 5)
applications. In the end, we point out several future research directions.
The goal of CLMs is to model the probability distributions over sequences of linguistic units:
P (u1 , u2 , · · · , ut ), (1)
where ui can be either a character, a word, a phrase, or other linguistic units. CLMs attempt to predict the next linguistic
unit in a text sequence given its preceding contexts:
P (ut |u<t ) (2)
CLMs are also called auto-regressive language models since the units are predicted in a causal way. Estimating the
probability of a text sequence as shown in Eq. (1) directly encounters the data sparsity problem. CLMs often estimate
Overview of Language Models
the joint probability of the text sequence by decomposing a text sequence into smaller units. For example, CLMs
leverage the chain rule and the conditional probability to estimate the joint probability in the form of
P (u1 , u2 , · · · , ut ) = P (u1 )P (u2 |u1 )P (u3 |u1 , u2 ) · · · P (ut |u1 , ...ut−1 ). (3)
CLMs are often trained from scratch with a training corpus and, then, predict the probability of text sequences
with respective applications. Representative models include N-grams LMs [1, 2, 3], exponential LMs [4, 5, 6] and
earlier neural LMs [7, 8]. CLMs give a high probability to natural text sequences occurring frequently in the real
world. As a result, they play a fundamental role in text generation, speech recognition [9, 10, 11], and machine
translation [12, 13, 14] until the emergence of PLMs. Nowadays, high-performance PLMs serve as the backbone of
many NLP systems. They are not limited to the causal predictive functionality of CLMs and provide more different
types of LMs.
The differences between CLMs and PLMs can be summarized below.
• Training Methodology. With the development of deep learning, PLMs with neural network structures are
pre-trained by collections of massive unlabeled corpora to learn generic knowledge which is then transferred
to downstream tasks by task-specific fine-tuning.
• Causality Constraint. PLMs do not necessarily follow CLMs in predicting linguistic units as shown in
Eq. (2). For example, bidirectional LMs [15, 16] use both preceding and succeeding contexts to predict the
missing linguistic units via probability estimation:
P (ut |u<t , u>t ). (4)
Bidirectional LMs do not follow the causality constraint and the chain rule in Eq. (3), to access the probability
of a text sequence, which makes it inherently different from CLMs.
• Token Representation. Apart from the differences in the training paradigm and probability modeling, PLMs
adopt a different representation for basic units called tokens. PLMs represent tokens by embedding them in a
high-dimensional continuous space such as word embeddings [17, 18] and sentence embeddings [19, 20, 21].
The new representations offer a flexible and powerful tool that enables PLMs to handle a wide range of tasks.
This overview paper serves two objectives. On one hand, instead of only focusing on recently developed PLMs
[22, 23, 24], we aim to provide a comprehensive overview of the basic concepts of LMs, the transition from CLMs to
PLMs, LM’s recent developments and applications to beginners in the field. On the other hand, we would like to shed
light on future research directions and offer our outlook to experienced engineers and researchers in the NLP field. For
example, we cover large LMs (LLMs) in the survey as there are growing interests in LLMs due to the new services
provided by ChatGPT. Furthermore, we include efficient LMs as an emerging topic since there are increasing concerns
about large model sizes and high training costs of LLMs.
The rest of the paper is organized as below. We introduce several types of LMs that go beyond CLMs in Sec. 2, and
provide an overview of common ways to decompose text sequences into smaller linguistic units in Sec. 3. Sec. 4
introduces different model structures. We discuss the training procedures of LMs in Sec. 5. Common evaluation
methods including, both intrinsic and extrinsic ones, are introduced in Sec. 6. The application of LMs to text generation
is discussed in Sec. 7. We comment on the redundancy problem of LMs and analyze techniques for efficient LMs in
Sec. 8. Promising future research directions are pointed out in Sec. 9. Concluding remarks are given in Sec. 10
2 Types of Language Models

CLMs commonly refer to auto-regressive models that predict the next linguistic units given the preceding context as
shown in Eq. (2). LMs can access the probability of a text sequence using the chain rule. The goal of CLMs is to decode
the probability of text sequences in a causal manner. In this section, we introduce more LMs that go beyond CLMs.
2.1 Structural LM
Instead of predicting linguistic units in a sequential or reversed sequential order, structural LMs [25, 26, 27, 28, 29]
predict linguistic units based on pre-defined linguistic structures such as dependency or constituent parse trees. Structural
LMs utilize the linguistic structure to bring linguistically relevant context closer to the linguistic unit to be predicted.
For example, given a parse tree structure, a structural LM can define the ancestor context A(ut ) of ut as the sequence
from the root node to the parent of ut . For example, the ancestor sequence of word ‘strong’ is {‘binoculars’, ‘saw’,
ROOT} in Fig. 1. Then, the structural LM uses the ancestor context in the tree to predict the next linguistic unit as
P (ut |A(ut )), (5)
2
Figure 1: The example of a dependency parse tree example [28].
Figure 2: The use of different permutations in a natural sentence.
where A(ut ) is the ancestor context of linguistic unit ut . Similar to CLMs, structural LMs are designed to model the
probability of text sequences. Differently, structure LMs decode the sequence probability in the order of their synthetic
structures. It has been successfully applied to sentence completion [27, 28] and speech recognition [25, 26].
2.2 Bidirectional LM
Instead of using the causal contexts to make predictions, bidirectional LMs utilize contexts from both directions as
shown in Eq. (4). The masked LM is one representative bidirectional LM. It masks out linguistic units in a text
sequence and, then, encodes their preceding and succeeding contexts to predict the masked linguistic units. Formally,
the prediction can be defined as the estimation of the following conditional probability
P (um |S̄), (6)
where um is the masked linguistic unit and S̄ is the corrupted text sequence by replacing a certain number of linguistic
units with [MASK] symbols. The goal of bidirectional LMs is to learn the inner dependency between linguistic units
in an unsupervised manner. The trained model can inherit semantics meanings from large-scale unlabeled corpora.
Different from CLMs that aim to model the generation probability of text sequences, pre-trained bidirectional LMs
are used as the backbone that transfers the learned knowledge through further fine-tuning in various downstream
applications.
2.3 Permutation LM
CLMs and masked LMs have their own advantages and disadvantages. A masked LM needs to create artificial tokens
such as [mask], which never occur in downstream tasks while CLMs only condition on preceding context. The
permutation LM [30] is a recently proposed LM that takes advantage of CLMs and masked LMs. Given an input
sequence of linguistic units, permutation LMs randomize the order of input linguistic units and construct different
permutations of the input sequence. Fig. 2 shows an example of different permutations given an input text sequence.
Let Z be the set of all possible permutations. Permutation LMs predict the next linguistic unit, ut , in one permutation,
Z, of the sequence based on
P (ut |uZ
<t ), Z ∈ Z. (7)
3
3 Linguistic Units
To estimate the probability of text sequences, LMs partition text sequences into small linguistic units such as characters,
words, phrases, or sentences. This process is called tokenization. Different languages and models may have different
appropriate tokenization methods. Here, we focus on English and use it as an example. In this section, we examine
typical tokenization methods used in language modeling according to unit sizes.
3.1 Characters
LMs can model text sequences probability based on characters [31, 32, 33, 34, 35]. As compared with other linguistics
units, using characters has a much smaller vocabulary size, leading to a smaller discrete space and model size. On
the other hand, it is challenging to predict the next character. Usually, it requires a long historical context. This
makes the performance of character-level LMs poorer than that of word-level LMs. In addition, the input and output
lengths have to be longer to model the character distribution accurately. This results in higher computational costs,
especially for auto-regressive decoding. Several LM methods use the combination of words and characters to alleviate
the issue [36, 37, 38].
3.2 Words and Subwords
The most natural tokenization for English is to decompose a text sequence into words by white spaces. Many LMs apply
word tokenization. However, there are several issues of naive word tokenization. The first one is the Out-Of-Vocabulary
(OOV) problem. Because an LM has a pre-defined vocabulary size that cannot be arbitrarily large. Less frequent words
and words with character-level errors may not be stored in the pre-defined vocabulary. Thus, they cannot be retrieved
from the dictionary. Although one can extend the vocabulary size to alleviate this problem, it will increase the model
size and still cannot handle all possible words.
LMs beyond the word level still have the OOV problem while a single character is not semantically meaningful by
themselves. Recently, researchers are in favor of decomposing words into subwords if they do not appear in the
dictionary. This offers a flexible and effective solution to the OOV problem [39, 40]. Several subword segmentation
algorithms are developed to boost the performance of LMs. They strike a balance between the good performance of
word-level models and the flexibility of character-level models. Two subword segmentation approaches, statistics-based
and linguistics-based, are presented below.
3.2.1 Statistics-based Subword Tokenizers

The statistics-based subword tokenizers generate subword vocabulary purely based on the corpus. The associated
methods are derived from a compression point of view. They work by replacing the commonly appeared character
sequences with a new symbol (word) that does not exist in the current vocabulary. Then, fewer bytes are needed for
information transmission.
Byte Pair Encoding (BPE). BPE [41] is a simple data compression technique that replaces the most common pair of
bytes in a sequence by a single unused byte recursively. It was adopted by [40] to solve the word segmentation problem.
That is, frequent characters or character sequences are merged to generate subwords. BPE is also used by several
advanced PLMs such as GPT-2 [42] and RoBERTa [16] with the following algorithm, called the BPE merge operation.
1. Prepare a training corpus and define the size of the subword vocabulary.
2. Split all words into characters.
3. Generate a new subword by merging a pair of characters or subwords with the highest frequency.
4. Repeat step 3 until the desired vocabulary size is reached.
An illustration of the BPE merge operation conducted on a small dictionary is given in Fig. 3.
WordPiece. [43] WordPiece is another data-driven subword algorithm. The difference between WordPiece and BPE is
that WordPiece merges the pair of A and B if they have the highest score P (AB)/P (A)P (B) (rather than the highest
frequency P (AB)) at each iterative step. For example, WordPiece merges the pair of “u” and “g” in Fig. 3 only if they
have the highest value, P (0 ug 0 )/P (0 u0 )P (0 g 0 ), as compared with other pairs. WordPiece is used as the tokenization
method in BERT [15], DistilBERT [44], and Electra [45].
There are other statistics-based subword tokenizers such as Unigram [46]. Different subword tokenizers and their
performance comparison are studied in [47].
4
Figure 3: Illustration of the BPE merge operation conducted on the dictionary {“hug", “pug", “pun", “bun"}. The
vocabulary is initialized with all characters. Then, a new subword is created by merging the most frequent pair.
3.2.2 Linguistics-based Subword Tokenizers

Linguistics-based subword tokenizers exploit the linguistic knowledge and decompose words into smaller grammatical
units, such as morphemes or syllables. Such subword tokenizers are widely used in machine translation and speech
recognition among different languages [48, 49, 50, 51, 52, 53, 54]. For example, in machine translation, words formed
by compounding, affixation, or inflection, can be conveniently translated by translating the morphemes, respectively.
However, linguistics-based subword tokenizers are not as popular as statistics-based ones due to the complexity and the
rule-based nature of language decomposition.
3.3 Phrases
The semantic meaning of a single word can be ambiguous because of various contexts and set collocations. Since
the linguistic dictionary does not go beyond the word-level, the inter-word dependency is ignored. Phrase-level LMs
replace common and cohesive word sequences by phrases [55, 56, 57, 58]. Phrase-level LMs are suitable for some
applications. For example, it is observed in [57] that short words with fewer syllables in automatic speech recognition
(ASR) are more frequently misrecognized than longer ones. Since phrases provide longer phone sequences than their
constituents, they are more robust to recognition errors for ASR.
3.4 Sentences
LMs with smaller linguistic units (e.g., characters, words, subwords, and phrases) rely on conditional probabilities
to estimate the probability of text sequences as given in Eq. (3). Sentence-level LMs [59, 60, 61, 62, 63] avoid the
use of the chain rule. They generate sentence features and, then, model the sentence probability directly. This is
because modeling the sentence probability directly is more convenient than that in Eq. (3) in encoding the sentence-level
information. It is also easier to encode the inter-sentence information such as the effects of preceding utterances in a
dialog flow.
4 Model Structures
In this section, we conduct a survey on several common structures to model the probability distributions of text sequences.
They are N-gram, maximum entropy, and neural network models. PLMs typically use continuous representations in
probability modeling built upon recurrent neural networks (RNNs) or transformers.
4.1 N-gram Models
An N-gram consists of N consecutive linguistic units from a text sequence. N-gram LMs [1, 2, 3] assume that the
probability of a word depends only on its preceding N-1 linguistic units and it is independent of other contexts. This is
known as the Markov assumption. Thus, instead of using all historical contexts, N-gram LMs only use the previous N-1
linguistic units to predict the current one; namely,
P (wt |w<t ) = P (wt |wt−N +1:t−1 ). (8)
5
N-gram LMs calculate the conditional probability by counting the occurrence time of N-grams given a training corpus
as
C(wt−N +1:t )
P (wt |wt−N +1:t−1 ) = . (9)
C(wt−N +1:t−1 )
N-gram LMs simplify the word probability calculation based on previous N-1 words, but they encounter two sparsity
issues. First, if an N-gram, (wt−N +1:t ), never occurs in the training corpus, the probability for the next word being wt
is zero. Second, if the (N-1)-gram, (wt−N +1:t−1 ), in the denominator never occurs, we cannot calculate the probability
of any word. These sparsity issues can be alleviated by smoothing techniques. A simple smoothing method [64, 65],
called additive smoothing, is to add a small value to the count for every N-gram so as to avoid zero in the numerator
and the denominator in Eq. (9). However, this simple smoothing is still deficient because it assigns the same probability
for N-grams that never occur in the training corpus.
There are more advanced smoothing techniques such as back-off and interpolation [66, 67, 68, 69, 70] that achieve
better probability estimation. In back-off, lower-order N-grams are used for probability estimation if higher-order
N-grams do not occur. For example, if C(wt−3:t−1 ) = 0, we back off to compute P (wt |wt−2:t−1 ). In interpolation,
different N-grams are considered for conditional probability computation. Mathematically, the N-gram probability is
estimated by
P (wt |wt−N +1:t−1 ) = λN P (wt |wt−N +1:t−1 ) + λN −1 P (wt |wt−N :t−1 )
+ λN −2 P (wt |wt−N −1:t−1 ) + ... + λ1 P (wt ), (10)
PN
where λi is the weight for each n-gram and i=1 λi = 1.
4.2 Maximum Entropy Models
Maximum Entropy models (also called the exponential models) [4, 5, 6] estimate the probability of text sequences
using feature functions in the form of
exp(aT f (w, h))
P (w|h) = P T 0
, (11)
w0 exp(a f (w , h))
where f (w, h) is the feature function that generates the feature of word w and its historical context h,
T 0
P
w0 exp(a f (w , h)) is a normalization factor, and a is a parameter vector derived by the Generalized Iterative
Scaling algorithm [71]. The features are usually generated from the N-grams.
4.3 Feed-forward Neural Network (FNN) Models
The discrete nature of the N-gram model is its performance bottleneck even with advanced smoothing techniques.
Neural LMs embrace the continuous embedding space (distributed representation of words) to overcome the data
sparsity problem. Feed-forward Neural Network (FNN) LMs [7, 72, 73, 74] is one of the earlier neural network models.
An FNN LM takes historical contexts as the input, and outputs the probability distribution of words. As shown in Fig. 4,
each word in the preceding context is represented as a vector through a projection layer (i.e., an embedding matrix).
These word vectors are sent to the hidden layer with H hidden units followed by non-linear activation. Then, a softmax
function is used to obtain the posterior probabilities for word candidates, P (wj = i|hj ), which are the probabilities of
words given a specific history predicted by the language model.
An FNN LM uses a fixed window to collect fixed-length contexts. It is essentially a neural version of N-gram LMs. The
FNN LM have several advantages over the N-gram LM by projecting words into continuous space. First, it can handle
unseen N-grams by representing each word as an N-gram with a dense vector space. Second, it is storage-efficient since
it does not need to count and store the transition probability of conventional N-gram models.
4.4 Recurrent Neural Network (RNN) Models
It is clearly insufficient to use the historical context in a fixed-length to predict the next word. In contrast to the limited
historical context used in the N-gram and FNN LMs, Recurrent Neural Network (RNN) LMs [8, 75, 76, 77, 78] can
exploit arbitrarily long histories to predict the next word.
The structure of a vanilla RNN LM is shown in Fig. 5. A word x(i) in position i is first converted into a one-hot
representation x̂(i). Then, the recurrent hidden state, h(i + 1), is computed using the previous hidden state, h(i), and
the one-hot representation, x̂(i), of word x(i) as
h(i + 1) = f (W x̂(i) + U h(i)), (12)
6
Figure 4: The structure of FFN LMs, where hj denotes the preceding contexts wj−n+1 , ..., wj−1 in a fixed-window,
and P , H, and N are the dimensions of the projection, the hidden layer, and the output layer, respectively [72].
where f (·) is a non-linear activation function, W is the weight matrix of the connections from the input layer to the
hidden layer, and U is the connection between the previous and current hidden layers, respectively. By iteratively
computing the hidden states, RNN LMs can encode the historical context of varying length. Finally, the output layer
gives the conditional probability of words y(t) = g(V h(t)), where V is the weight matrix connecting the hidden layer
and the output layer and g(·) is the softmax activation function.
Figure 5: The structure of RNN LMs [79].
In theory, RNN LMs do not need the Markov assumption. They can use all preceding history to predict the next
word. However, the inherent gradient vanishing problem of RNN hampers the learning of the model [80]. Since the
gradient may become very small over a long distance, model weights are actually updated by the nearby context in
practice. Generally, RNN LMs cannot learn the dependency between the current word and its far-away historical
context. Although an attention mechanism can be introduced to RNNs to alleviate this problem [81, 79]. The inherent
sequential nature of RNNs makes them less powerful than transformer-based LMs with a self-attention mechanism.
7
Figure 6: The structure of a transformer [82].
4.5 Transformers
The transformer architecture [82] can capture long-term dependencies and important sequence components by exploiting
a self-attention mechanism. Unlike the recurrent structure of RNNs, a transformer is easy to parallelize in both training
and inference. Its structure is shown in Fig. 6. It consists of an encoder and a decoder. Before being sent to the encoder,
the input textual sequence is first converted to an embedding through an embedding layer plus positional embedding.
Multi-head attention, which is an ensemble of multiple self-attention mechanisms, enables the transformer to capture
more robust and diverse attention between tokens. The other parts in the transformer encoder include feed-forward
layers, residual connections, and normalization layers. The difference between the transformer encoder and decoder is
that the transformer decoder has an additional masked multi-head attention layer. The masking ensures the decoder can
only access preceding tokens of the current one, which makes the decoder auto-regressive.
Based on different purposes, transformers have encoder-only, decoder-only, and encoder-decoder three variants as
shown in Table 1 and Fig. 7. Encoder-only models can access all positions given an input and utilize bi-directional
contexts to predict words. They are suitable for tasks requiring understanding full sentences, such as text classification.
Transformer decoder-only models can only use previous words to predict the current word (namely, auto-regressive
models). They are good at text generation tasks such as story generation. Transformer encoder-decoder models can
access all words in the encoding phase, and words before the current word in the decoding phase. They are suitable for
sequence-to-sequence tasks such as translation and summarization.
8
Table 1: Transformer-based PLMs.
BERT [15]
Encoder-only models
RoBERTa [16]
(Bidirectional)
ELECTRA [45]
PaLM [83]
Decoder-only models
GPT-1, 2 and 3 [84, 42, 85]
(Unidirectional)
Transformer XL [86]
Encoder-Decoder models BART [87]
(Sequence to sequence) T5 [88]
Figure 7: Illustration of different transformer models, where BERT is the encoder-only model, GPT is the decoder-only
model, and BART is the encoder-decoder model [87].
5 Pre-trained Language Models

Pre-trained language models (PLMs) are dominating in the NLP field nowadays. With the development of deep
learning, the training and usage of PLMs have changed a lot as compared with conventional statistical LMs. Before
being applied to real-world tasks, PLMs are first pre-trained on massive collections of corpora so that they learn
universal representations that carry both syntactic and semantic knowledge. After pre-training, PLMs are fine-tuned for
downstream tasks so that the acquired knowledge can be transferred to different tasks. In the following, we first explain
the pre-training objectives in Sec. 5.1 and then talk about how to adapt PLMs to various tasks of interest through
fine-tuning in Sec. 5.2. It is also worthwhile to point out several good survey papers on PLMs, e.g., [22, 23, 24].
5.1 Pre-training
The most commonly used pre-training task is “missing word prediction". There are other pre-training tasks for different
purposes, e.g., next-sentence prediction, which allows an LM to learn sentence relationships.
Word Prediction. Auto-aggressive language LMs are trained to predict the next word using previous words. While
bidirectional LMs mask a subset of words in a sample and learn to predict such masked words using the rest of the
context. For the latter, the most popular objective is the masked language model (MLM) objective as proposed in
BERT [15]. The MLM objective is the cross-entropy loss in predicting masked tokens. It randomly masks out 15% of
9
Figure 8: An illustration of (a) LM pre-training, (b) standard fine-tuning, and (c) prompt-based fine-tuning (or
prompt-tuning) [89].
the input tokens and then predicts the masked tokens. The number of masked tokens is set to 15% based on experimental
verification. If the masking rate is too small, the model only learns from a limited number of masked tokens. On the
other hand, if it is too large, there is not enough context to do reasonable predictions and models cannot learn well.
Other Pre-training Tasks. There are other pre-training tasks to make LMs learn better linguistic knowledge such
as sentence relationships. For example, next sentence prediction is used as the pre-training task in BERT [15]. Next
sentence prediction is formalized as a binary prediction task that decides whether two sentences are two consecutive
sentences or not. In this way, a PLM can be used in downstream tasks that require the understanding of the relationship
between two sentences, such as Question Answering (QA) and Natural Language Inference (NLI). Other pre-training
objectives are adopted by BART [87]. They include token deletion, text infilling, sentence permutation, and document
rotation to corrupt the original sequence for reconstruction. Shuffled tokens are used in T5 [88] to increase the robustness
of the learned representation.
5.2 Fine-Tuning and Prompt-Tuning
PLMs learn non-task-specific language knowledge in the pre-training stage. Fine-tuning performs task-specific
adaptations of the model so that they can be applied to different downstream tasks. The model parameters are updated in
the fine-tuning stage. One can choose to update the entire model, freeze the language model and update the task-specific
heads, or only update certain layers of an LM. PLMs are pre-trained by one or several pre-training objectives and, then,
applied to different downstream tasks. The gap between pre-training tasks and downstream task-specific fine-tuning can
be substantial. Prompt-tuning is used to discover the potential of PLMs by mimicking the pre-training objectives in
the fine-tuning or inference stage. As PLMs get more powerful, they can handle various downstream tasks by seeing
a few examples without any gradient updates or fine-tuning [85]. This is achieved by prompt-based fine-tuning (or
prompt-tuning in short).
A prompt is a human-designed text that is concatenated with the original text as the new model input. Fig. 8 shows an
example of the pre-training task, fine-tuning and prompt-tuning of MLMs. In the pre-training, the MLMs are trained to
predict masked words. Assuming that the downstream task is the sentiment analysis of the movie review. In standard
fine-tuning, we train a new head on the top of a PLM and predict the sentiment labels. The original input appended with
a designed prompt, say, ‘It was’, is sent to the PLM. The PLM has to assign probabilities to designed answers, which
can be ‘great’ or ‘terrible’. If the probability of ‘great’ is higher, then the label of the input will be positive and vice
versa. In this way, prompt-tuning converts a distinct downstream task to the word prediction task to narrow the gap
between the pre-training and fine-tuning stages.
6 Model Evaluation
There are two LM evaluation types: intrinsic evaluation and extrinsic evaluation. The intrinsic evaluation examines the
internal properties of an LM while the extrinsic evaluation studies its performance in downstream tasks.
10
6.1 Intrinsic Evaluation
Auto-regressive LM. LMs estimate the probability of text sequences. A good LM assigns higher probabilities to
natural text sequences and lower ones to unreal or random text sequences. The perplexity is a common evaluation metric
for this purpose. Given a testing text sequence, the perplexity, denoted by P P L, is defined as the inverse probability of
the sequence normalized by the number of words. Mathematically, we have
s
N
1
P P L(W ) = , (13)
(P (w1 w2 ...wN )
where W = w1 w2 ...wN is a testing text sequence. The perplexity can be rewritten in form of
v
u n
uY
N
1
P P L(W ) = t . (14)
i=1
(P (w |w
i 1 ...wi−1 )
A good LM should maximize the text set probability. It is equivalent to minimizing the perplexity. The lower the
perplexity, the better the LM.
Bidirectional Language Model. To calculate the inverse probability in Eq. (13), the auto-regressive LMs can use
a sequence of conditional probabilities. However, this approach does not work for bidirectional LMs (or masked
LMs). Several intrinsic evaluation metrics have been proposed for bidirectional LMs. The pseudo-log-likelihood score
(PLL) [90] is defined as
|W |
X
P LL(W ) = log P (wt |W\t ), (15)
t=1
where log P (wt |W\t ) is the conditional probability of token w in sentence W with all remaining tokens. Instead of
maximizing the joint probability of the entire text sequence, a good bidirectional LM should maximize the probability
of each token in the text sequence given other tokens. Based on PLLs, the pseudo-Perplexity (PPPL) for corpora C is
defined as [91]
1 X
P P L(C) = exp(− P LL(W )). (16)
N
W ∈C
Both PLL and PPPL provide effective means to measure the naturalness of sentences for a bidirectional LM. For
example, it was shown in [91] that PLL and PPPL correlate well with the performance of an LM on downstream tasks,
such as automatic speech recognition and machine translation.
6.2 Extrinsic Evaluation
Any downstream task of LMs can be used for extrinsic evaluation. There are several common downstream tasks selected
as extrinsic evaluation benchmarks. Two popular ones are GLUE (General Language Understanding Evaluation) [92]
and SuperGLUE [93]. GLU is an evaluation benchmark for natural language understanding. It contains single-sentence
tasks, similarity and paraphrase tasks, and inference tasks. SuperGLUE is an enhanced version of GLUE. It includes a
new set of more challenging language understanding tasks, more diverse task formats, improved resources, and a public
leaderboard.
6.3 Relation between Intrinsic and Extrinsic Evaluations
If an LM achieves a lower perplexity, does that mean it can also perform well on downstream tasks? In other words, is
there any correlation between pre-training tasks (based on word prediction) and the downstream tasks? There are many
empirical studies on this question but few theoretical studies.
Empirical Studies. Researchers design experiments to understand what kind of knowledge is learned by an LM
from the pre-training tasks. Examples include [94, 95, 96, 97, 98, 99]. They use part-of-speech tagging, constituent
labeling, and dependency labeling to measure the degree of syntactic knowledge learning, and named entity labeling,
semantic role labeling, and semantic proto-role for testing semantic knowledge. Empirical studies show that pre-training
tasks help LMs learn the linguistic knowledge such as the grammar [97] and the semantic role [99]. However, these
experimental results can only be used as evidence supporting that the word prediction tasks benefit downstream tasks.
They cannot explain the underlying mechanism.
Theoretical Studies. Some researchers attempt to build the connection between LM’s perplexities and its performance
on downstream tasks mathematically. The text classification tasks were studied in [100]. They first hypothesized and
11
verified that text classification tasks can be reformulated as sentence completion tasks. Since the LM pre-training task
is essentially a sentence completion task, it does help the text classification downstream task. Then, they quantified
the connection mathematically and showed that √ the features from LMs that achieve -optimal in log-perplexity can
linearly solve text classification tasks with O( ) error. An underlying generative model was utilized in [101] to show
the relationship between the pre-training tasks and the downstream tasks. Current theoretical studies are limited in the
sense that only a specific downstream task (say, the text classification task) is considered and the proof holds under
certain conditions.
7 Applications in Text Generation
One of the most important applications of LMs is text generation, which aims to generate sequences of words based
on the input data. There are many text generation tasks because of different purposes and inputs. For example,
the automatic speech recognition (ASR) task demands that the input be a speech sequence while the output be the
corresponding text sequence. The machine translation task generates the translated text sequence based on the input
text sequence and the target language. Story Generation is a topic-to-text generation task. In this section, we introduce
common techniques used in text generation and then explain how LMs can be applied in each of the representative
tasks.
7.1 Decoding Methods
Decoding decides the next output linguistic unit to generate text. A good decoding method should generate coherent
continuation given a context. As LMs get more sophisticated, decoding methods have played an increasingly important
role. As shown in Fig. 9, deficient decoding methods lead to bad generated texts even with a powerful LM. There are
two main decoding methods for text generation.
Figure 9: Comparison of texts generated by the powerful GPT-2 large language model (LLM) using Beam search (left)
and pure sampling decoding (right). Beam search yields degenerate repetition (in blue) while pure sampling results in
incoherent gibberish (in red) [102].
Maximization-based decoding. This is the most commonly used decoding objective. Assuming that the model assigns
a higher probability to a higher quality text which is closer to the ground truth written by humans, the maximization-
based decoding strategy searches for tokens with the highest probability as the generated text. Greedy search [103, 104]
chooses the token with the highest probability as the next token in a greedy manner. Beam search [105, 106, 107]
keeps a certain number of most likely tokens at each time step and selects the generated token sequences with the
overall highest probability eventually. It avoids missing reasonable tokens that do not have the highest probability.
Trainable decoding algorithms have been proposed recently. Trainable greedy decoding [108] is a neural-based solution
that works as part of a neural machine translation decoder. It utilizes reinforcement learning to find a translation that
maximizes a decoding objective.
Sampling-based decoding. It chooses the next token from a set of sampled tokens. Because maximization-based
decoding depends highly on the underlying model probabilities and suffers from producing degenerate repetition,
sampling-based decoding increases the diversity of generated texts by random sampling. However, the simple pure
sampling may choose a token with low probability (from an unreliable tail distribution) as the next generated token.
As a result, the generated text could be unrelated to the prefix, leading to incoherent gibberish. Top-k sampling [109]
and Nucleus sampling [102] have recently been proposed to address this problem. Both sample from truncated LM
distributions (i.e., sampling from the most probable tokens). Diverse Beam search [105] is a trainable sampling-based
12
(stochastic) decoding algorithm based on the Beam search. It uses reinforcement learning to determine the beam
diversity parameters for different inputs or tasks.
7.2 Dialogue Systems
A dialogue system aims at simulating human responses when conversing with human users. Recent dialogue systems
such as ChatGPT 1 and LaMDA [110] have attracted a lot of attention in the generative AI field because of their
superior performance as interactive chatbot systems. Dialogue systems can be categorized into task-oriented systems
and open-domain systems. The former is designed for specific tasks such as customer service for online shopping.
The latter is also known as chatbots [111]. Most modern dialogue systems are fine-tuned versions of generative
LMs. Taking ChatGPT as an example, ChatGPT is built based on a generative LM, GPT-3 [85], with over 188 billion
parameters. It is further fine-tuned by supervised learning and reinforcement learning on labeled data.
LMs play an important role in dialogue systems, especially in their natural language understanding (NLU) and natural
language generation (NLG) components [112, 113]. NLU is responsible for understanding and recognizing users’
intent. Nowadays, PLM encoders provide informative representations for NLU while the associated PLM decoders
are responsible for generating an appropriate response. The latter involves constructing the response text, selecting
appropriate words, and determining the correct phrasing and tone. The effectiveness of representations of PLMs was
examined in [114] for dialogue tasks. The evaluation PLM targets included BERT [15] and GPT2 [42]. The few-shot
capability of PLMs in dialogue tasks such as NLU and NLG was evaluated in [115]. Overall, LMs in dialogue systems
play a key role in understanding users’ input and generating appropriate and natural responses.
7.3 Automatic Speech Recognition
Automatic speech recognition (ASR) is a speech-to-text generation task that aims to transform raw audio input into
the corresponding text sequence. The LM plays an essential role in an ASR system. First, it helps solve acoustically
ambiguous utterances. Second, it can lower the computational cost by constraining the search space in a set of words of
higher probability. ASR systems contain an acoustic model and a language model, which are related by
P (word|sound) ∝ P (sound|word)P (word). (17)
The acoustic model is conditioned on phones P (sound|word) while the LM gives the word distribution denoted by
P (word). LMs help search the word hypotheses during recognition. Different types of LMs have been explored in
ASR such as N-gram [116, 117], FFNN [118], RNN [119, 120] and Transformer [121]
7.4 Machine Translation
Machine translation is a text-to-text generation task where the text in the source language is translated into that of the
target language. LMs adopted by machine translation are conditioned on both the source sentence and the previous
partial translation. Recently, transformer-based models has achieved great successes in machine translation [82, 122, 85]
7.5 Detection of Generated texts
As the performance of LMs gets closer to or even outperforms humans, the misuse of LMs, such as fake news and fake
product reviews generation, has become a serious problem. The ability to detect machine-generated texts is important.
There are two types of detection problems: 1) human written vs. machine generated, and 2) inveracious vs. veracious.
Most datasets, e.g., [123, 124, 125], are collected for the first type. Problems of the second type are much harder than
those of the first type [126] since one needs to connect generated text to the fact, which requires a high-level knowledge
reasoning capability.
Two common approaches to detecting machine-generated text are reviewed below. One is to exploit the probability
distribution of LMs [127, 128]. If the probability distribution of a text sequence is closer to that of human-written texts
as compared with known machine-generated texts, the text sequence is classified as human-written. The other is to train
classifiers with supervised learning [129, 124]. It converts the distribution to a supervised binary classification task. For
more details on the detection of machine-generated texts, readers are referred to two survey papers [130, 131].
1
https://openai.com/blog/chatgpt/
13
8 Efficient Models
As recent PLMs get more powerful, their model size, training cost, and demand for training data increase tremendously.
They need high computational resources and energy consumption, limiting their real-world applications. Table 2 shows
the model size, training data, cost, and time of recently developed LMs. This issue is a concern to many people and the
construction of efficient LMs has received attention.
Model Year Model size Training data Training cost Training time
64 TPU chips
BERT-Large 2018 340M 2.5B words $7,000
4 days
512 TPU v3 chips
XLNet-Lagre 2019 340M 32.9B words $245,000
5.5 days
GPT-2 2019 1.5B 8 million web pages $60,000
174 GB 512 GPUs
Megatron-LM 2019 8.3B
deduplicated text 53 minutes
Above $1.3 million
T5 2019 11B
for a single run
Turing-NLG 2020 17.2B
GPT-3 2020 175B 45TB of text data $12 million
Megatron-Turing 2K A100 GPUs
2021 530B 270B
NLG 3 months
Table 2: Comparison of nodel sizes, training data, cost, and time of several large LMs, where blank cells indicate that
the data are not available.
8.1 Data Usage
Pre-training Data Size. A critical question for PLM training is how much data is needed. The effect of the pre-training
data size on the RoBERTa model was studied in [132]. The learning curves of four model performance measures as
a function of the pre-training dataset size are shown in Fig. 10. When the data size ranges between 100M and 1B
words, three learning curves gradually level off and it implies that LMs encode most syntactic and semantic features.
However, a much larger quantity of data is needed for LMs to acquire enough common-sense knowledge and other
skills to achieve better performance on downstream NLU tasks.
Figure 10: The performance curves as functions of the pre-training dataset size, where the classifier probing measures
the quality of the syntactic and semantic features, the minimum description length probing quantifies the accessibility of
these features, the BLiMP curve measures the model’s knowledge of various syntactic phenomena, and the superGLUE
measures the capability of handling NLU tasks [132].
14
Efficient Pre-Training. Several methods have been proposed to use the pre-training data more efficiently. In the
pre-training of masked LMs, a certain percentage of tokens are masked and need to be inferred by context. This
approach incurs a substantial amount of computational cost because the network only learns from a certain percentage
of tokens which are masked. To enhance training efficiency, the work in [45] uses “replaced token detection" (rather
than “masked token prediction") as the pre-training task. As shown in Fig. 11, a generator is trained to perform the
masked LM and predicts the masked tokens. Then, the main model works as a discriminator, called ELECTRA, which
learns to decide the original or replaced tokens. In this way, pre-training tasks are conducted on all tokens instead of a
small subset of masked tokens. Learning from all input positions causes ELECTRA to train much faster than BERT
which adopts masked word prediction. Besides, ELECTRA achieves higher accuracy on downstream tasks when it is
fully trained. Later, a new pre-training task using an energy-based model, which is closely related to ELECTRA, is
proposed in [133].
Figure 11: The structure of ELECTRA (Efficiently Learning an Encoder that Classifier Token Replacements Accu-
rately) [45]
Bridging Pre-training and Downstream Tasks. A typical pre-training task is word prediction, which often has a large
gap with downstream tasks. To mitigate the gap between pre-training and downstream tasks, prompt tuning has been
studied in [84, 42, 89, 134, 135]. As illustrated in Fig. 8, the head is trained to predict the masked words in masked
LMs. For the downstream sentiment analysis task, the head is trained to predict the positive or the negative label in
traditional fine-tuning. A template (e.g., ‘It was’) and its expected text responses (e.g., ‘great’ and ‘terrible’) are used in
prompt tuning. In this way, pre-training and prompt tuning share the same “word prediction" objective.
8.2 Model Size
Besides improving training efficiency, efficient LMs focus on the design of models of smaller sizes. Many methods are
investigated to reduce the model size so that the model can be implemented on mobile or edge devices with limited
computing resources. Model compression is a widely studied topic. Compression methods first train a large LM and then
compress it into a target size. Examples include model pruning [136, 137, 138], knowledge distillation [139, 44, 140],
low rank matrix approximation [78, 141, 142], and parameter sharing [143, 144, 145].
9 Future Research Directions

In this section, we describe several promising future research directions in language modeling.
9.1 Integration of LMs and KGs
Knowledge Graph (KG) provides a structured representation of human knowledge [146, 147]. It has been widely used
in many NLP applications, such as question answering [148] and text summarization [149], because of its capability to
represent relationships between entities. There is a growing interest in evaluating the knowledge learned in PLMs [150],
where the relationship between different semantic units is captured in the embedding space and the self-attention layers.
Several ideas are proposed in [151] to leverage KGs for LM training. As a result, the knowledge learned in the models
can be greatly improved. Thus, it is worth careful investigation of integrating KGs with LMs and understanding how
they interact with each other.
It appears that KG can serve as an information database to be queried by LMs. LMs are powerful in natural language
understanding and generation while KGs can organize and store the knowledge information extracted from the training
corpus. In other words, we may decompose knowledge sources into semantic and syntactic two components, which
can be handled by KGs and LMs, respectively. In the training phase, a KG is constructed based on the information
extracted from the training corpus, and an LM can be trained simultaneously. In the inference phase, an LM can serve
15
as an interface between humans and the knowledge database represented in form of KGs. There are advantages to
assigning semantic and syntactic processing tasks to KGs and LMs, respectively. For example, the decoupling facilitates
incremental learning, allows a smaller model size, and improves interpretability. They will be further elaborated on
below.
9.2 Incremental Learning
Incremental learning aims to incorporate new information without re-training existing models entirely. The problem of
catastrophic forgetting associated with neural network models was pointed out in [152]. That is, the information that
has already been learned by a model can be gradually forgotten when training with new information. This problem is
particularly critical to large LMs since the new information keeps arriving. A solution to catastrophic forgetting was
proposed in [153]. It attempts to remember prior important tasks by slowing down learning on weights that are more
relevant to them. However, it is difficult to define important tasks in LMs. Re-training a large LM with both old and
new data is too expensive.
The importance of developing a satisfactory solution to incremental learning for LMs cannot be over-emphasized.
Incremental learning is challenging for neural networks. Yet, it is easy for KGs to add new data to (or remove old data
from) an existing database by adding or removing factual triples [154]. Clearly, the current information in the KGs will
not be overwritten by newly collected data. The information in the database is updated incrementally. To this end, the
integration of KGs and LMs provides an excellent solution that meets the need for incremental learning.
9.3 Lightweight Models
As mentioned in Section 8, PLMs get more powerful at the expense of huge computational resources and energy
consumption. The cost issue has to be faced seriously in the development of large LMs (LLMs). Besides, LLMs are
unfriendly to our environment due to their high carbon footprint. Green Learning (GL) targets learning solutions with
low carbon footprint. The design of lightweight models of smaller sizes and lower computational complexity without
sacrificing performance has received more attention in recent years [155, 156, 157, 158]. The design of green LMs is
an important topic worth serious investigation.
Current PLMs are data-driven models that use neural architectures to learn generic language knowledge from a large
amount of data. Efforts have been made in the development of lightweight LMs. Model compression is one of the
popular approaches to obtaining a small LM. Examples include knowledge distillation or pruning [159]. However, this
methodology appears to be a detour since it trains large models and then shrinks their sizes by compression. Instead, we
may incorporate the linguistic information and the domain knowledge to offer a more direct way to reduce the model
size and the amount of training data.
9.4 Universal versus Domain-Specific Models
A universal LM is developed to handle tasks in the general domain. For example, ChatGPT is a universal dialogue
LM pre-trained on multilingual and general domain corpora. It can converse on open-domain topics in multiple
languages. In contrast, domain-specific LMs [160, 161, 162, 163] are designed to deal with domain-specific tasks, e.g.,
biomedicine, economics, musicology, etc.
A universal LM demands a huge model size, a large number of training examples, and a tremendous amount of
computational resources. Based on the scaling law of neural language models [164], the inference performance scales
as a power-law with the model size, the dataset size, and the amount of computing used for training. So far, the largest
PLM contains 540-billion parameters [83]. Despite the superior performance and the flexibility to adapt to multiple
tasks, we may wonder whether a huge universal LM is cost-effective.
For domain-specific LMs, the amount of training data in need is significantly lower. It was believed that the general
domain PLMs benefit the training of domain-specific LMs. However, it is reported in [161] that domain-specific LMs,
which were pre-trained from scratch on in-domain data, can provide a solid foundation for biomedical NLP. In other
words, training a domain-specific LM may not need a huge amount of general corpora and labeled data. Domain-specific
LMs to be deployed on task-specific scenarios with less training and inference efforts expect to receive more attention
in the future.
9.5 Interpretable Models
Although deep-learning-based LMs are dominating in the NLP field, they are inherently black-box methods without
mathematical transparency. Its interpretability is of concern. Efforts have been made to explain the black-box LMs. As
16
mentioned in 6.3, empirical studies are conducted to understand what PLMs have learned through experimental design.
However, the progress in this direction may offer insights but not a satisfactory and clean answer. Providing theoretical
explanations or establishing explainable LMs is still a challenging and open issue. A direction to interpretability is to
design an interpretable learning model from scratch. For example, we may incorporate KGs with LMs. KGs are proven
to be powerful in many reasoning tasks such as information retrieval [165] and recommendation systems [166]. KGs
can provide a logical path for each prediction so that predictions offered by LMs can be more explainable.
9.6 Machine Generated Text Detection
The most common application of LMs is text generation. As generative LM’s performance gets closer to or even
outperforms humans, these LMs can be used for malicious purposes such as academic dishonesty, spamming, targeted
bot attacks, and fake news/reviews generation. How to determine whether a text is generated by LMs or written by
humans is a big challenge nowadays. A high-performance machine-generated text classifier can only serve as a reference
in real-world applications, since it has false positives (i.e., human-written texts classified as machine-generated) and
false negatives (i.e., machine-generated texts classified as human-written). In addition, people may be even more
interested in detecting veracious and unveracious texts. They care more about whether the text is true or not. Detecting
disinformation could be more difficult than detecting machine/human-generated text without assessing the factuality.
Additionally, the factuality may change as time goes by. It is critical to our society in developing effective tools to
identify malicious usages of generative LMs.
10 Conclusion
A comprehensive overview of CLMs and their successors, PLMs, was presented in this paper and a wide range of topics
was covered. First, different levels of linguistic units were introduced and how linguistic unit prediction is used to train
language models was examined. Second, tokenization methods adopted by language models were discussed. Third,
language model structures and the training paradigm of PLMs were reviewed. Fourth, we studied the evaluation and
applications of language models. Especially, several applications in the context of text generation were detailed. Finally,
several future research directions were pointed out. The need for explainable, reliable, domain-specific, and lightweight
language models was emphasized.
References
[1] Peter F Brown, Vincent J Della Pietra, Peter V Desouza, Jennifer C Lai, and Robert L Mercer. Class-based
n-gram models of natural language. Computational linguistics, 18(4):467–480, 1992.
[2] Marcello Federico. Bayesian estimation methods for n-gram language model adaptation. In Proceeding of Fourth
International Conference on Spoken Language Processing. ICSLP’96, volume 1, pages 240–243. IEEE, 1996.
[3] Thomas R Niesler and Philip C Woodland. A variable-length category-based n-gram language model. In 1996
IEEE International Conference on Acoustics, Speech, and Signal Processing Conference Proceedings, volume 1,
pages 164–167. IEEE, 1996.
[4] Stephen A Della Pietra, Vincent J Della Pietra, Robert L Mercer, and Salim Roukos. Adaptive language modeling
using minimum discriminant estimation. In Speech and Natural Language: Proceedings of a Workshop Held at
Harriman, New York, February 23-26, 1992, 1992.
[5] Adam Berger, Stephen A Della Pietra, and Vincent J Della Pietra. A maximum entropy approach to natural
language processing. Computational linguistics, 22(1):39–71, 1996.
[6] Roni Rosenfeld. A maximum entropy approach to adaptive statistical language modeling. 1996.
[7] Yoshua Bengio, Réjean Ducharme, and Pascal Vincent. A neural probabilistic language model. Advances in
neural information processing systems, 13, 2000.
[8] Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernockỳ, and Sanjeev Khudanpur. Recurrent neural network
based language model. In Interspeech, volume 2, pages 1045–1048. Makuhari, 2010.
[9] Frederick Jelinek, Lalit Bahl, and Robert Mercer. Design of a linguistic statistical decoder for the recognition of
continuous speech. IEEE Transactions on Information Theory, 21(3):250–256, 1975.
[10] Frederick Jelinek. Continuous speech recognition by statistical methods. Proceedings of the IEEE, 64(4):532–556,
1976.
[11] Lalit R Bahl, Frederick Jelinek, and Robert L Mercer. A maximum likelihood approach to continuous speech
recognition. IEEE transactions on pattern analysis and machine intelligence, (2):179–190, 1983.
17
[12] Peter F Brown, John Cocke, Stephen A Della Pietra, Vincent J Della Pietra, Frederick Jelinek, John Lafferty,
Robert L Mercer, and Paul S Roossin. A statistical approach to machine translation. Computational linguistics,
16(2):79–85, 1990.
[13] Franz Josef Och, Nicola Ueffing, and Hermann Ney. An efficient a* search algorithm for statistical machine
translation. In Proceedings of the ACL 2001 Workshop on Data-Driven Methods in Machine Translation, 2001.
[14] Kenji Yamada and Kevin Knight. A decoder for syntax-based statistical mt. In Proceedings of the 40th Annual
meeting of the Association for Computational Linguistics, pages 303–310, 2002.
[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional
transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
[16] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke
Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint
arXiv:1907.11692, 2019.
[17] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke
Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1
(Long Papers), pages 2227–2237, New Orleans, Louisiana, June 2018. Association for Computational Linguistics.
[18] Chengwei Wei, Bin Wang, and C.-C. Jay Kuo. Synwmd: Syntax-aware word mover’s distance for sentence
similarity evaluation. arXiv preprint arXiv:2206.10029, 2022.
[19] Bin Wang and C.-C. Jay Kuo. SBERT-WK: A sentence embedding method by dissecting bert-based word models.
IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2146–2157, 2020.
[20] Tianyu Gao, Xingcheng Yao, and Danqi Chen. SIMCSE: Simple contrastive learning of sentence embeddings.
arXiv preprint arXiv:2104.08821, 2021.
[21] Bin Wang and Haizhou Li. Relational sentence embedding for flexible semantic matching. arXiv preprint
arXiv:2212.08802, 2022.
[22] Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. Pre-trained models for
natural language processing: A survey. Science China Technological Sciences, 63(10):1872–1897, 2020.
[23] Xu Han, Zhengyan Zhang, Ning Ding, Yuxian Gu, Xiao Liu, Yuqi Huo, Jiezhong Qiu, Yuan Yao, Ao Zhang,
Liang Zhang, et al. Pre-trained models: Past, present and future. AI Open, 2:225–250, 2021.
[24] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train,
prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint
arXiv:2107.13586, 2021.
[25] Ciprian Chelba and Frederick Jelinek. Expoiting syntactic structure for language modeling. arXiv preprint
cs/9811022, 1998.
[26] Ciprian Chelba and Frederick Jelinek. Structured language modeling. Computer Speech & Language, 14(4):283–
332, 2000.
[27] Joseph Gubbins and Andreas Vlachos. Dependency language models for sentence completion. In Proceedings of
the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1405–1410, 2013.
[28] Piotr Mirowski and Andreas Vlachos. Dependency recurrent neural language models for sentence completion.
[29] Chengwei Wei, Bin Wang, and C.-C. Jay Kuo. Task-specific dependency-based word embedding methods.
Pattern Recognition Letters, 2022.
[30] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet:
Generalized autoregressive pretraining for language understanding. Advances in neural information processing
systems, 32, 2019.
[31] Ilya Sutskever, James Martens, and Geoffrey E Hinton. Generating text with recurrent neural networks. In ICML,
2011.
[32] Kyuyeon Hwang and Wonyong Sung. Character-level language modeling with hierarchical recurrent neural
networks. In 2017 ieee international conference on acoustics, speech and signal processing (icassp), pages
5720–5724. IEEE, 2017.
[33] Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language models. In
Thirtieth AAAI conference on artificial intelligence, 2016.
18
[34] Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level language modeling
with deeper self-attention. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages
3159–3166, 2019.
[35] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and
Colin Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the
Association for Computational Linguistics, 10:291–306, 2022.
[36] Moonyoung Kang, Tim Ng, and Long Nguyen. Mandarin word-character hybrid-input neural network language
model. In Twelfth Annual Conference of the International Speech Communication Association, 2011.
[37] Yasumasa Miyamoto and Kyunghyun Cho. Gated word-character recurrent language model. arXiv preprint
arXiv:1606.01700, 2016.
[38] Lyan Verwimp, Joris Pelemans, Patrick Wambacq, et al. Character-word lstm language models. arXiv preprint
arXiv:1704.02813, 2017.
[39] Tomáš Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and Jan Cernocky. Subword
language modeling with neural networks. preprint (http://www. fit. vutbr. cz/imikolov/rnnlm/char. pdf), 8(67),
2012.
[40] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword
units. arXiv preprint arXiv:1508.07909, 2015.
[41] Philip Gage. A new algorithm for data compression. C Users Journal, 12(2):23–38, 1994.
[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are
unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
[43] Mike Schuster and Kaisuke Nakajima. Japanese and korean voice search. In 2012 IEEE international conference
on acoustics, speech and signal processing (ICASSP), pages 5149–5152. IEEE, 2012.
[44] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller,
faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.
[45] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders
as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020.
[46] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword
candidates. arXiv preprint arXiv:1804.10959, 2018.
[47] Kaj Bostrom and Greg Durrett. Byte pair encoding is suboptimal for language model pretraining. arXiv preprint
arXiv:2004.03720, 2020.
[48] Daniel Kiecza, Tanja Schultz, and Alex Waibel. Data-driven determination of appropriate dictionary units for
korean lvcsr. In Proceedings of ICASSP, pages 323–327, 1999.
[49] Mathias Creutz and Krista Lagus. Unsupervised morpheme segmentation and morphology induction from text
corpora using Morfessor 1.0. Helsinki University of Technology Helsinki, 2005.
[50] Mathias Creutz, Teemu Hirsimäki, Mikko Kurimo, Antti Puurula, Janne Pylkkönen, Vesa Siivola, Matti Var-
jokallio, Ebru Arisoy, Murat Saraçlar, and Andreas Stolcke. Morph-based speech recognition and modeling of
out-of-vocabulary words across languages. ACM Transactions on Speech and Language Processing (TSLP),
5(1):1–29, 2007.
[51] Tomaž Rotovnik, Mirjam Sepesy Maučec, and Zdravko Kačič. Large vocabulary continuous speech recognition
of an inflected language using stems and endings. Speech communication, 49(6):437–452, 2007.
[52] Ruhi Sarikaya, Mohamed Afify, and Yuqing Gao. Joint morphological-lexical language modeling (jmllm)
for arabic. In 2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP’07,
volume 4, pages IV–181. IEEE, 2007.
[53] Haşim Sak, Murat Saraclar, and Tunga Güngör. Morphology-based and sub-word language modeling for turkish
speech recognition. In 2010 IEEE International Conference on Acoustics, Speech and Signal Processing, pages
5402–5405. IEEE, 2010.
[54] Mijit Ablimit, Graham Neubig, Masato Mimura, Shinsuke Mori, Tatsuya Kawahara, and Askar Hamdulla.
Uyghur morpheme-based language models and asr. In IEEE 10th INTERNATIONAL CONFERENCE ON
SIGNAL PROCESSING PROCEEDINGS, pages 581–584. IEEE, 2010.
[55] Bernhard Suhm and Alex Waibel. Towards better language models for spontaneous speech. 1994.
19
[56] Klaus Ries, Finn Dag Buo, and Alex Waibel. Class phrase models for language modeling. In Proceeding of
Fourth International Conference on Spoken Language Processing. ICSLP’96, volume 1, pages 398–401. IEEE,
1996.
[57] George Saon and Mukund Padmanabhan. Data-driven approach to designing compound words for continuous
speech recognition. IEEE transactions on Speech and audio processing, 9(4):327–332, 2001.
[58] Michael Levit, Sarangarajan Parthasarathy, Shuangyu Chang, Andreas Stolcke, and Benoit Dumoulin. Word-
phrase-entity language models: Getting more mileage out of n-grams. In Fifteenth Annual Conference of the
International Speech Communication Association, 2014.
[59] Ronald Rosenfeld. A whole sentence maximum entropy language model. In 1997 IEEE workshop on automatic
speech recognition and understanding proceedings, pages 230–237. IEEE, 1997.
[60] Stanley F Chen and Ronald Rosenfeld. Efficient sampling and feature selection in whole sentence maximum
entropy language models. In 1999 IEEE International Conference on Acoustics, Speech, and Signal Processing.
Proceedings. ICASSP99 (Cat. No. 99CH36258), volume 1, pages 549–552. IEEE, 1999.
[61] Ronald Rosenfeld, Stanley F Chen, and Xiaojin Zhu. Whole-sentence exponential language models: a vehicle
for linguistic-statistical integration. Computer Speech & Language, 15(1):55–73, 2001.
[62] Daphne Ippolito, David Grangier, Douglas Eck, and Chris Callison-Burch. Toward better storylines with
sentence-level language models. arXiv preprint arXiv:2005.05255, 2020.
[63] Haejun Lee, Drew A Hudson, Kangwook Lee, and Christopher D Manning. Slm: Learning a discourse language
representation with sentence unshuffling. arXiv preprint arXiv:2010.16249, 2020.
[64] George James Lidstone. Note on the general case of the bayes-laplace formula for inductive or a posteriori
probabilities. Transactions of the Faculty of Actuaries, 8(182-192):13, 1920.
[65] William Ernest Johnson. Probability: The deductive and inductive problems. Mind, 41(164):409–423, 1932.
[66] Frederick Jelinek. Interpolated estimation of markov source parameters from sparse data. In Proc. Workshop on
Pattern Recognition in Practice, 1980, 1980.
[67] Slava Katz. Estimation of probabilities from sparse data for the language model component of a speech recognizer.
IEEE transactions on acoustics, speech, and signal processing, 35(3):400–401, 1987.
[68] Kenneth W Church and William A Gale. A comparison of the enhanced good-turing and deleted estimation
methods for estimating probabilities of english bigrams. Computer Speech & Language, 5(1):19–54, 1991.
[69] Reinhard Kneser and Hermann Ney. Improved backing-off for m-gram language modeling. In 1995 international
conference on acoustics, speech, and signal processing, volume 1, pages 181–184. IEEE, 1995.
[70] Stanley F Chen and Joshua Goodman. An empirical study of smoothing techniques for language modeling.
Computer Speech & Language, 13(4):359–394, 1999.
[71] John N Darroch and Douglas Ratcliff. Generalized iterative scaling for log-linear models. The annals of
mathematical statistics, pages 1470–1480, 1972.
[72] Holger Schwenk and Jean-Luc Gauvain. Training neural network language models on very large corpora. In
Proceedings of human language technology conference and conference on empirical methods in natural language
processing, pages 201–208, 2005.
[73] Holger Schwenk. Continuous space language models. Computer Speech & Language, 21(3):492–518, 2007.
[74] Ebru Arisoy, Tara N Sainath, Brian Kingsbury, and Bhuvana Ramabhadran. Deep neural network language
models. In Proceedings of the NAACL-HLT 2012 Workshop: Will We Ever Really Replace the N-gram Model?
On the Future of Language Modeling for HLT, pages 20–28, 2012.
[75] Tomáš Mikolov, Stefan Kombrink, Lukáš Burget, Jan Černockỳ, and Sanjeev Khudanpur. Extensions of recurrent
neural network language model. In 2011 IEEE international conference on acoustics, speech and signal
processing (ICASSP), pages 5528–5531. IEEE, 2011.
[76] Stefan Kombrink, Tomas Mikolov, Martin Karafiát, and Lukás Burget. Recurrent neural network based language
modeling in meeting recognition. In Interspeech, volume 11, pages 2877–2880, 2011.
[77] Martin Sundermeyer, Ralf Schlüter, and Hermann Ney. Lstm neural networks for language modeling. In
Thirteenth annual conference of the international speech communication association, 2012.
[78] Zhilin Yang, Zihang Dai, Ruslan Salakhutdinov, and William W Cohen. Breaking the softmax bottleneck: A
high-rank rnn language model. arXiv preprint arXiv:1711.03953, 2017.
20
[79] Hongli Deng, Lei Zhang, and Lituan Wang. Global context-dependent recurrent neural network language model
with sparse feature learning. Neural Computing and Applications, 31(2):999–1011, 2019.
[80] Sepp Hochreiter. The vanishing gradient problem during learning recurrent neural nets and problem solutions.
International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 6(02):107–116, 1998.
[81] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473, 2014.
[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser,
and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.
[83] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul
Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with
pathways. arXiv preprint arXiv:2204.02311, 2022.
[84] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by
generative pre-training. 2018.
[85] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.
Advances in neural information processing systems, 33:1877–1901, 2020.
[86] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl:
Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019.
[87] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves
Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language
generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019.
[88] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei
Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach.
Learn. Res., 21(140):1–67, 2020.
[89] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot learners. arXiv
preprint arXiv:2012.15723, 2020.
[90] Alex Wang and Kyunghyun Cho. Bert has a mouth, and it must speak: Bert as a markov random field language
model. arXiv preprint arXiv:1902.04094, 2019.
[91] Julian Salazar, Davis Liang, Toan Q Nguyen, and Katrin Kirchhoff. Masked language model scoring. arXiv
[92] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-
task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461,
2018.
[93] Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy,
and Samuel Bowman. Superglue: A stickier benchmark for general-purpose language understanding systems.
Advances in neural information processing systems, 32, 2019.
[94] Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R Thomas McCoy, Najoung Kim, Benjamin
Van Durme, Samuel R Bowman, Dipanjan Das, et al. What do you learn from context? probing for sentence
structure in contextualized word representations. arXiv preprint arXiv:1905.06316, 2019.
[95] Mario Giulianelli, Jack Harding, Florian Mohnert, Dieuwke Hupkes, and Willem Zuidema. Under the hood:
Using diagnostic classifiers to investigate and improve how language models track agreement information. arXiv
[96] Ian Tenney, Dipanjan Das, and Ellie Pavlick. Bert rediscovers the classical nlp pipeline. arXiv preprint
arXiv:1905.05950, 2019.
[97] Taeuk Kim, Jihun Choi, Daniel Edmiston, and Sang-goo Lee. Are pre-trained language models aware of phrases?
simple but strong baselines for grammar induction. arXiv preprint arXiv:2002.00737, 2020.
[98] John Hewitt and Christopher D Manning. A structural probe for finding syntax in word representations. In
Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–4138, 2019.
[99] Anna Rogers, Olga Kovaleva, and Anna Rumshisky. A primer in bertology: What we know about how bert
works. Transactions of the Association for Computational Linguistics, 8:842–866, 2020.
21
[100] Nikunj Saunshi, Sadhika Malladi, and Sanjeev Arora. A mathematical exploration of why language models help
solve downstream tasks. arXiv preprint arXiv:2010.03648, 2020.
[101] Colin Wei, Sang Michael Xie, and Tengyu Ma. Why do pretrained language models help in downstream tasks?
an analysis of head and prompt tuning. Advances in Neural Information Processing Systems, 34:16158–16170,
2021.
[102] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration.
[103] Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi. Learning discourse-level diversity for neural dialog models
using conditional variational autoencoders. arXiv preprint arXiv:1703.10960, 2017.
[104] Zhen Xu, Bingquan Liu, Baoxun Wang, Cheng-Jie Sun, Xiaolong Wang, Zhuoran Wang, and Chao Qi. Neural
response generation via gan with an approximate embedding layer. In Proceedings of the 2017 conference on
empirical methods in natural language processing, pages 617–626, 2017.
[105] Jiwei Li, Will Monroe, and Dan Jurafsky. A simple, fast diverse decoding algorithm for neural generation. arXiv
[106] Ashwin Vijayakumar, Michael Cogswell, Ramprasaath Selvaraju, Qing Sun, Stefan Lee, David Crandall, and
Dhruv Batra. Diverse beam search for improved description of complex scenes. In Proceedings of the AAAI
Conference on Artificial Intelligence, volume 32, 2018.
[107] Ilia Kulikov, Alexander H Miller, Kyunghyun Cho, and Jason Weston. Importance of search and evaluation
strategies in neural dialogue modeling. arXiv preprint arXiv:1811.00907, 2018.
[108] Jiatao Gu, Kyunghyun Cho, and Victor OK Li. Trainable greedy decoding for neural machine translation. arXiv
[109] Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. arXiv preprint
arXiv:1805.04833, 2018.
[110] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia
Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint
arXiv:2201.08239, 2022.
[111] Jinjie Ni, Tom Young, Vlad Pandelea, Fuzhao Xue, and Erik Cambria. Recent advances in deep learning based
dialogue systems: A systematic survey. Artificial intelligence review, pages 1–101, 2022.
[112] Bin Wang, Chen Zhang, Chengwei Wei, and Haizhou Li. A focused study on sequence length for dialogue
summarization. arXiv preprint arXiv:2209.11910, 2022.
[113] Bin Wang, Chen Zhang, Yan Zhang, Yiming Chen, and Haizhou Li. Analyzing and evaluating faithfulness in
dialogue summarization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language
Processing, pages 4897–4908, Abu Dhabi, United Arab Emirates, December 2022. Association for Computational
Linguistics.
[114] Chien-Sheng Wu and Caiming Xiong. Probing task-oriented dialogue representation from language models. In
Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages
5036–5051, Online, November 2020. Association for Computational Linguistics.
[115] Andrea Madotto, Zihan Liu, Zhaojiang Lin, and Pascale Fung. Language models as few-shot learner for
task-oriented dialogue systems. arXiv preprint arXiv:2008.06239, 2020.
[116] Fred Jelinek, B Merialdo, S Roukos, M Strauss, et al. Self-organized language modeling for speech recognition.
In Readings in speech recognition. Citeseer, 1990.
[117] Manhung Siu and Mari Ostendorf. Variable n-grams and extensions for conversational speech language modeling.
IEEE Transactions on Speech and Audio Processing, 8(1):63–75, 2000.
[118] Ebru Arısoy, Stanley F Chen, Bhuvana Ramabhadran, and Abhinav Sethy. Converting neural network language
models into back-off language models for efficient decoding in automatic speech recognition. IEEE/ACM
Transactions on Audio, Speech, and Language Processing, 22(1):184–192, 2013.
[119] Ebru Arisoy, Abhinav Sethy, Bhuvana Ramabhadran, and Stanley Chen. Bidirectional recurrent neural network
language models for automatic speech recognition. In 2015 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pages 5421–5425. IEEE, 2015.
[120] Zhiheng Huang, Geoffrey Zweig, and Benoit Dumoulin. Cache based recurrent neural network language model
inference for first pass speech recognition. In 2014 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pages 6354–6358. IEEE, 2014.
22
[121] Joonbo Shin, Yoonhyung Lee, and Kyomin Jung. Effective sentence scoring method using bert for speech
recognition. In Asian Conference on Machine Learning, pages 1081–1093. PMLR, 2019.
[122] Qiang Wang, Bei Li, Tong Xiao, Jingbo Zhu, Changliang Li, Derek F Wong, and Lidia S Chao. Learning deep
transformer models for machine translation. arXiv preprint arXiv:1906.01787, 2019.
[123] Max Weiss. Deepfake bot submissions to federal public comment websites cannot be distinguished from human
submissions. Technology Science, 2019.
[124] Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. Authorship attribution for neural text generation. In
Conf. on Empirical Methods in Natural Language Processing (EMNLP), 2020.
[125] Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. Tweepfake: About
detecting deepfake tweets. Plos one, 16(5):e0251415, 2021.
[126] James Thorne and Andreas Vlachos. Automated fact checking: Task formulations, methods and future directions.
[127] Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. Automatic detection of generated
text is easiest when humans are fooled. arXiv preprint arXiv:1911.00650, 2019.
[128] Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. Gltr: Statistical detection and visualization of
generated text. arXiv preprint arXiv:1906.04043, 2019.
[129] Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi.
Defending against neural fake news. Advances in neural information processing systems, 32, 2019.
[130] Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. Automatic detection of machine
generated text: A critical survey. arXiv preprint arXiv:2011.01314, 2020.
[131] Harald Stiff and Fredrik Johansson. Detecting computer-generated disinformation. International Journal of Data
Science and Analytics, 13(4):363–383, 2022.
[132] Yian Zhang, Alex Warstadt, Haau-Sing Li, and Samuel R Bowman. When do you need billions of words of
pretraining data? arXiv preprint arXiv:2011.04946, 2020.
[133] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Pre-training transformers as
energy-based cloze models. arXiv preprint arXiv:2012.08561, 2020.
[134] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language
inference. arXiv preprint arXiv:2001.07676, 2020.
[135] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting
knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980,
2020.
[136] Ziheng Wang, Jeremy Wohlwend, and Tao Lei. Structured pruning of large language models. arXiv preprint
arXiv:1910.04732, 2019.
[137] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention:
Specialized heads do the heavy lifting, the rest can be pruned. arXiv preprint arXiv:1905.09418, 2019.
[138] Demi Guo, Alexander M Rush, and Yoon Kim. Parameter-efficient transfer learning with diff pruning. arXiv
[139] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert:
Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351, 2019.
[140] Iulia Turc, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Well-read students learn better: On the
importance of pre-training compact models. arXiv preprint arXiv:1908.08962, 2019.
[141] Xindian Ma, Peng Zhang, Shuai Zhang, Nan Duan, Yuexian Hou, Ming Zhou, and Dawei Song. A tensorized
transformer for language modeling. Advances in neural information processing systems, 32, 2019.
[142] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu
Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.
[143] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert:
A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019.
[144] Raj Dabre and Atsushi Fujita. Recurrent stacking of layers for compact neural machine translation models. In
Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6292–6299, 2019.
[145] Sascha Rothe, Shashi Narayan, and Aliaksei Severyn. Leveraging pre-trained checkpoints for sequence generation
tasks. Transactions of the Association for Computational Linguistics, 8:264–280, 2020.
23
[146] Yun-Cheng Wang, Xiou Ge, Bin Wang, and C-C Jay Kuo. Kgboost: A classification-based knowledge base
completion method with negative sampling. Pattern Recognition Letters, 157:104–111, 2022.
[147] Xiou Ge, Yun-Cheng Wang, Bin Wang, and C-C Jay Kuo. Compounde: Knowledge graph embedding with
translation, rotation and scaling compound operations. arXiv preprint arXiv:2207.05324, 2022.
[148] Xiao Huang, Jingyuan Zhang, Dingcheng Li, and Ping Li. Knowledge graph embedding based question
answering. In Proceedings of the twelfth ACM international conference on web search and data mining, pages
105–113, 2019.
[149] Luyang Huang, Lingfei Wu, and Lu Wang. Knowledge graph-augmented abstractive summarization with
semantic-driven cloze reward. arXiv preprint arXiv:2005.01159, 2020.
[150] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian
Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019.
[151] Oshin Agarwal, Heming Ge, Siamak Shakeri, and Rami Al-Rfou. Knowledge graph based synthetic corpus
generation for knowledge-enhanced language model pre-training. arXiv preprint arXiv:2010.12688, 2020.
[152] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128–135,
1999.
[153] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran
Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in
neural networks. Proceedings of the national academy of sciences, 114(13):3521–3526, 2017.
[154] Bin Wang, Guangtao Wang, Jing Huang, Jiaxuan You, Jure Leskovec, and C-C Jay Kuo. Inductive learning
on commonsense knowledge graph completion. In 2021 International Joint Conference on Neural Networks
(IJCNN), pages 1–8. IEEE, 2021.
[155] Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. Green ai. Communications of the ACM,
63(12):54–63, 2020.
[156] Jingjing Xu, Wangchunshu Zhou, Zhiyi Fu, Hao Zhou, and Lei Li. A survey on green deep learning. arXiv
[157] C.-C. Jay Kuo and Azad M Madni. Green learning: Introduction, examples and outlook. Journal of Visual
Communication and Image Representation, page 103685, 2022.
[158] Yun-Cheng Wang, Xiou Ge, Bin Wang, and C-C Jay Kuo. GreenKGC: A lightweight knowledge graph
completion method. arXiv preprint arXiv:2208.09137, 2022.
[159] Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joey Gonzalez. Train big,
then compress: Rethinking model size for efficient training and inference of transformers. In International
Conference on Machine Learning, pages 5958–5968. PMLR, 2020.
[160] Denghui Zhang, Zixuan Yuan, Yanchi Liu, Fuzhen Zhuang, Haifeng Chen, and Hui Xiong. E-bert: A phrase and
product knowledge enhanced language model for e-commerce. arXiv preprint arXiv:2009.02835, 2020.
[161] Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng
Gao, and Hoifung Poon. Domain-specific language model pretraining for biomedical natural language processing.
ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1–23, 2021.
[162] Kesong Liu, Jianhui Jiang, and Feifei Lyu. A domain knowledge enhanced pre-trained language model for
vertical search: Case study on medicinal products. In Proceedings of the 29th International Conference on
Computational Linguistics, pages 1014–1023, 2022.
[163] Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. Biogpt: generative
pre-trained transformer for biomedical text generation and mining. Briefings in Bioinformatics, 23(6), 2022.
[164] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray,
Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint
arXiv:2001.08361, 2020.
[165] Laura Dietz, Chenyan Xiong, Jeff Dalton, and Edgar Meij. The second workshop on knowledge graphs and
semantics for text retrieval, analysis, and understanding (kg4ir). In The 41st International ACM SIGIR Conference
on Research & Development in Information Retrieval, pages 1423–1426, 2018.
[166] Yuhao Yang, Chao Huang, Lianghao Xia, and Chenliang Li. Knowledge graph contrastive learning for recom-
mendation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in
Information Retrieval, pages 1434–1443, 2022.
24

A O L M: R D O: N Verview On Anguage Odels Ecent Evelopments and Utlook

Uploaded by

Copyright:

Available Formats

A O L M: R D O: N Verview On Anguage Odels Ecent Evelopments and Utlook

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A O L M: R D O: N Verview On Anguage Odels Ecent Evelopments and Utlook

Uploaded by

Copyright:

Available Formats

A N OVERVIEW ON L ANGUAGE M ODELS :

R ECENT D EVELOPMENTS AND O UTLOOK

2 Types of Language Models

Figure 1: The example of a dependency parse tree example [28].

Figure 2: The use of different permutations in a natural sentence.

P (um |S̄), (6)

3.2 Words and Subwords

3.2.1 Statistics-based Subword Tokenizers

3.2.2 Linguistics-based Subword Tokenizers

4.1 N-gram Models

4.2 Maximum Entropy Models

4.3 Feed-forward Neural Network (FNN) Models

4.4 Recurrent Neural Network (RNN) Models

Figure 5: The structure of RNN LMs [79].

Figure 6: The structure of a transformer [82].

Table 1: Transformer-based PLMs.

5 Pre-trained Language Models

5.2 Fine-Tuning and Prompt-Tuning

6.1 Intrinsic Evaluation

6.2 Extrinsic Evaluation

6.3 Relation between Intrinsic and Extrinsic Evaluations

7 Applications in Text Generation

7.1 Decoding Methods

7.2 Dialogue Systems

7.3 Automatic Speech Recognition

P (word|sound) ∝ P (sound|word)P (word). (17)

7.4 Machine Translation

7.5 Detection of Generated texts

8.1 Data Usage

8.2 Model Size

9 Future Research Directions

9.1 Integration of LMs and KGs

9.2 Incremental Learning

9.3 Lightweight Models

9.4 Universal versus Domain-Specific Models

9.5 Interpretable Models

9.6 Machine Generated Text Detection

You might also like