Papers by Clara Meister
A Measure-Theoretic Characterization of Tight Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | Language modeling is a core task in natural language processing. |
| Approach: | They propose to characterize leakage onto the set of infinite sequences by a measure-theoretic approach. |
| Outcome: | The proposed language model families are tight, meaning they will not leak . the proposed language models are based on the 'sequence leakage' hypothesis . |
A Formal Perspective on Byte-Pair Encoding (2023.findings-acl)
Copied to clipboard
| Challenge: | Byte-Pair Encoding (BPE) is a popular algorithm used for tokenizing data in NLP, but the underlying optimization problem that BPE seeks to solve has not yet been laid down. |
| Approach: | They propose an algorithm which is a 1/sigma*(1-e(-sigma))-approximation of an optimal merge sequence. |
| Outcome: | The proposed algorithm improves the runtime complexity from O(NM) to O(N log M) and the lower bound of the approximation is approx0.37. |
Language Model Quality Correlates with Psychometric Predictive Power in Multiple Languages (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have found that higher quality language models provide more powerful predictors of human reading behavior, but empirical support for the QP hypothesis is mixed. |
| Approach: | They propose to test the quality–power hypothesis by using surprisal language models to test their ability to predict eye tracking data. |
| Outcome: | The proposed model is based on a set of language models with a 'quality-power' hypothesis. |
Estimating the Entropy of Linguistic Distributions (2022.acl-short)
Copied to clipboard
| Challenge: | Shannon entropy is a quantity of interest for linguists studying the communicative capacity of human language. |
| Approach: | They propose to use entropy estimators to estimate linguistic effects from observed data . they propose to recommend better entropic estimators for future linguistic studies . |
| Outcome: | The proposed estimators are over-estimated due to poor entropy estimators, the authors argue . they recommend the same estimators be used in future linguistic studies. |
Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness in Tokenization (2026.acl-long)
Copied to clipboard
Negar Foroutan, Clara Meister, Debjit Paul, Joel Niklaus, Sina Ahmadi, Antoine Bosselut, Rico Sennrich
| Challenge: | Tokenization is the first step of most NLP pipelines. |
| Approach: | They propose a parity-aware byte pair encoder that maximizes the compression gain of the currently worst-compressed language for cross-lingual parity. |
| Outcome: | a new algorithm reduces tokenization inequality by 89% compared to classical BPE . the proposed algorithm is based on a fair-max rule that maximizes the compression gain of the currently worst-compressed language . |
On the probability–quality paradox in language generation (2022.acl-short)
Copied to clipboard
| Challenge: | a recent study suggests that mode-seeking decoding methods produce unnatural language . a higher probability text is perceived as more human-like, but this is not the case with stochastic decoding . |
| Approach: | They propose to explain why higher-probability text is perceived as more human-like . they argue that human-type language should contain an amount of information close to entropy . |
| Outcome: | The proposed model produces lower-probability text, which is perceived as more human-like . elucidation of this hypothesis supports the hypothesis . |
If beam search is the answer, what was the question? (2020.emnlp-main)
Copied to clipboard
| Challenge: | surprisingly, beam search results on language generation tasks are low-quality . despite its high error rate, beam searches can be used to decode models with high probability . |
| Approach: | They frame beam search as the exact solution to a different decoding objective . they propose a set of decoding objectives that explicitly enforce this property . |
| Outcome: | The proposed method enforces uniform information density in text, a property motivated by cognitive science. |
Determinantal Beam Search (2021.acl-long)
Copied to clipboard
| Challenge: | a new beam search approach allows for a diverse subset selection process . standard beam search does not encode interactions between candidates . |
| Approach: | They propose a beam search reformulation that casts subset selection as the subdeterminant optimization problem. |
| Outcome: | The proposed method offers competitive performance against other diverse set generation strategies while providing a more general approach to optimizing for diversity. |
Revisiting the Optimality of Word Lengths (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing theories of communication cost are based on the length of an utterance, but they are not based upon the frequency of the utterant. |
| Approach: | They propose a method to minimize CCH's cost by comparing language lengths to surprisal's expectation and variance-to-mean ratios. |
| Outcome: | The proposed method does not minimize CCH’s cost, but rather a lower bound, which we term CCH-lower. |
Locally Typical Sampling (2023.tacl-1)
Copied to clipboard
| Challenge: | a discrepancy in probabilistic language generators has puzzled the language generation community for years . |
| Approach: | They propose a method for local typical sampling to enforce a conditional entropy criterion for probabilistic models. |
| Outcome: | The proposed method can provide new insights into why high-probability texts can be dull or repetitive. |
Revisiting the Uniform Information Density Hypothesis (2021.emnlp-main)
Copied to clipboard
| Challenge: | The uniform information density hypothesis posits a preference among language users for utterances structured such that information is distributed uniformly across a signal. |
| Approach: | They propose to test the hypothesis by using reading time and acceptability data to examine the effect of surprisal on language comprehension and acceptabilities. |
| Outcome: | The proposed hypothesis makes predictions about language comprehension and linguistic acceptability . |
How to Compute the Probability of a Word (2024.emnlp-main)
Copied to clipboard
| Challenge: | Language models estimate a probability distribution over strings in a natural language . many recent linguistic studies have been incorrectly computing word probabilities . |
| Approach: | They propose to use the correct method to compute word probabilities . they highlight issues when relying on models that use end-of-word tokenisers . |
| Outcome: | Empirically, correcting the widespread bug affects measured outcomes in sentences and lexical optimisation analyses. |
A Plug-and-Play Method for Controlled Text Generation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for controlling language generation are not able to produce fluent text . current methods require additional models or fine-tuning to ensure specific words are included . |
| Approach: | They propose a plug-and-play decoding method that allows for controlled language generation . they add a shift in the probability distribution over our vocabulary towards semantically similar words . |
| Outcome: | The proposed method outperforms competing methods in human evaluations and does not impact fluency. |
Best-First Beam Search (2020.tacl-1)
Copied to clipboard
| Challenge: | Currently, beam search is the default for decoding structured predictors . however, little work has been done to speed up beam search itself . |
| Approach: | They propose a beam search algorithm that prunes the scoring function to a monotonic sequence length, which allows for safe pruning of hypotheses that cannot be in the final set of hypothecies. |
| Outcome: | The proposed method can be implemented up to 10x faster in practice. |
Naturalistic Causal Probing for Morpho-Syntax (2023.tacl-1)
Copied to clipboard
| Challenge: | Existing methods for probing are limited and lack understanding of their limitations and weaknesses. |
| Approach: | They propose a strategy for input-level intervention on naturalistic sentences . they use morpho-syntactic features of a sentence to intervene on the rest of the sentence . |
| Outcome: | The proposed approach allows for input-level intervention on naturalistic sentences while keeping the rest of the sentence unchanged. |
On Decoding Strategies for Neural Text Generators (2022.tacl-1)
Copied to clipboard
| Challenge: | a recent study suggests that decoding strategies may be more important than the model architecture itself when generating text from probabilistic models. |
| Approach: | They propose to measure changes in attributes of generated text as a function of decoding strategy and task using human and automatic evaluation. |
| Outcome: | The proposed study shows that decoding strategies do not always transfer across tasks . authors show that the differences in attributes are not always consistent across tasks, they say . |
A Cross-Linguistic Pressure for Uniform Information Density in Word Order (2023.tacl-1)
Copied to clipboard
Thomas Hikaru Clark, Clara Meister, Tiago Pimentel, Michael Hahn, Ryan Cotterell, Richard Futrell, Roger Levy
| Challenge: | a recent study has compared real and counterfactual word orders, but one functional pressure has been overlooked . a study of 10 typologically diverse languages shows that real word orders have greater uniformity than reverse word orders . |
| Approach: | They propose to test whether a pressure for UID may have influenced word order patterns cross-linguistically. |
| Outcome: | The proposed model shows that real orders have greater uniformity than reverse orders among SVO languages. |
Mutual Information Alleviates Hallucinations in Abstractive Summarization (2022.emnlp-main)
Copied to clipboard
| Challenge: | Abstractive summarization models exhibit the tendency to hallucinate, i.e., output content not supported by the source document. |
| Approach: | They propose a decoding strategy that optimizes for pointwise mutual information of source and target tokens when models exhibit uncertainty. |
| Outcome: | The proposed method decreases the probability of hallucinated tokens while maintaining the Rouge and BERT-S scores of top-performing decoding strategies. |
Searching for Search Errors in Neural Morphological Inflection (2021.eacl-main)
Copied to clipboard
| Challenge: | Neural sequence-to-sequence models are the predominant choice for language generation tasks. |
| Approach: | They find that on word-level tasks, the empty string is often the global optimum . they suggest that the poor calibration of many neural models may stem from characteristics of a specific subset of tasks rather than general ill-suitedness of such models for language generation. |
| Outcome: | The results suggest that the poor calibration of many neural models may stem from characteristics of a specific subset of tasks rather than general ill-suitedness of such models for language generation. |
A surprisal–duration trade-off across and within the world’s languages (2021.emnlp-main)
Copied to clipboard
| Challenge: | Throughout human evolution, countless languages have evolved, each with unique features. |
| Approach: | They analysed a corpus of 600 languages to find strong evidence for a surprisal–duration trade-off between languages and languages. |
| Outcome: | The proposed model shows that phones are produced faster in languages where they are less surprising and vice versa. |
Language Model Evaluation Beyond Perplexity (2021.acl-long)
Copied to clipboard
| Challenge: | a nascent literature on probing language models has focused on studying linguistic phenomena. |
| Approach: | They propose a framework for evaluating the fit of language models to natural language tendencies. |
| Outcome: | The proposed framework evaluates language models to the tendencies of natural language . it shows that the models learn only a subset of the tendancies considered . |
Conditional Poisson Stochastic Beams (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for beam search are based on a deterministic approach, but the results are not as accurate as those used in SBS. |
| Approach: | They propose a method that turns beam search into a stochastic process by using conditional Poisson sampling design instead of taking the maximizing set at each iteration. |
| Outcome: | The proposed method produces lower variance and more efficient estimators than SBS, even showing improvements in high entropy settings. |
Causal Estimation of Tokenisation Bias (2025.acl-long)
Copied to clipboard
| Challenge: | Modern language models define probabilities over character-strings, but in practice, it does . Ideally, the choice of the tokeniser should not affect the probability assigned to the underlying character- string. |
| Approach: | They quantify a type of tokenisation bias by framing it as a causal effect and estimating it using the regression discontinuity design. |
| Outcome: | The proposed model can estimate tokenisation bias by comparing subwords around arbitrary cutoff points. |
Towards a Similarity-adjusted Surprisal Theory (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have shown that surprisal theory ignores the possibility of similarity between words and treats them as distinct entities. |
| Approach: | They propose a new measure of comprehension effort called information value that accounts for communicative equivalences between possible continuations. |
| Outcome: | The proposed measure of comprehension effort is based on the diversity index of the diversity of communicative units. |
Tokenization and the Noiseless Channel (2023.acl-long)
Copied to clipboard
| Challenge: | Subword tokenization is a key part of most NLP pipelines, but little is known about why some combinations lead to improved downstream model performance. |
| Approach: | They propose that good tokenizers lead to efficient channel usage . they propose that an optimal encoding assigns extremely long codes to low-frequency subwords . |
| Outcome: | The proposed tokenizers have a very strong correlation with BLEU in machine translation . the proposed function can be used to improve model performance in the downstream task . |
The Role of n-gram Smoothing in the Age of Neural Networks (2024.naacl-long)
Copied to clipboard
| Challenge: | n-gram smoothing techniques were used to overcome overfitting problems in neural language models for decades. |
| Approach: | They propose to convert any n-gram smoothing technique into a regularizer compatible with neural language models. |
| Outcome: | The proposed regularizers outperform label smoothing on language modeling and machine translation. |
Generating Text from Language Models (2023.acl-tutorials)
Copied to clipboard
| Challenge: | a growing percentage of natural language processing tasks focus on the generation of text from probabilistic language models. |
| Approach: | They will provide a centralized discussion of critical considerations when choosing how to generate from a language model. |
| Outcome: | This tutorial will provide a centralized discussion of critical considerations when choosing how to generate from a language model. |
On Homophony and Rényi Entropy (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent theories of language optimality have tried to justify its prevalence, arguing that homophony enables the reuse of efficient wordforms and is thus beneficial for languages. |
| Approach: | They propose a new information-theoretic quantification of a language’s homophony: the sample Rényi entropy. |
| Outcome: | The proposed method is more nuanced than either Piantadosi et al.'s or Trott and Bergen's results. |
Is Sparse Attention more Interpretable? (2021.acl-short)
Copied to clipboard
| Challenge: | Sparse attention has been claimed to increase model interpretability . however, the attention distribution is typically over representations internal to the model rather than the inputs themselves . |
| Approach: | They conduct experiments to understand how sparsity affects our ability to use attention as an explainability tool. |
| Outcome: | The proposed model does not map to a sparse set of influential inputs, but rather to fewer inputs. |
A Natural Bias for Language Generation Models (2023.acl-short)
Copied to clipboard
| Challenge: | a standard probabilistic model for language generation has likely not yet learnt many semantic or syntactic rules of natural language, making it difficult to estimate the probability distribution over next tokens. |
| Approach: | They propose to initialise bias terms in a model's final linear layer with the log-unigram distribution and use it to output the unigram frequency statistics as prior knowledge. |
| Outcome: | The proposed method improves learning efficiency and improves overall performance. |
Generalized Entropy Regularization or: There’s Nothing Special about Label Smoothing (2020.acl-main)
Copied to clipboard
| Challenge: | Prior work has explored regularizing the output distributions of probabilistic models to alleviate overfitting. |
| Approach: | They propose a family of entropy regularizers that have a connection to regularization . they find that label smoothing provably does not allow for sparsity in an output distribution . |
| Outcome: | The proposed method improves the relationship between model entropy and performance on language generation tasks. |
On the Efficacy of Sampling Adapters (2023.acl-long)
Copied to clipboard
| Challenge: | Using sampling adapters can improve the quality of the generated text. |
| Approach: | They propose a framework for understanding sampling adapters and propose 'sampling adapters' they argue that the shift enforced by them can be viewed as a trade-off between precision and recall . |
| Outcome: | The proposed framework can be used to improve the quality of language models by modifying their distributions to improve their precision and recall. |
A Cognitive Regularizer for Language Modeling (2021.acl-long)
Copied to clipboard
| Challenge: | a uniform information density hypothesis is used to explain certain linguistic phenomena . a regularizer that encodes the UID hypothesis can be used for language training . |
| Approach: | They propose to augment the canonical MLE objective with a regularizer that encodes UID . they find that regularization consistently improves perplexity in language models . |
| Outcome: | The proposed hypothesis can be operationalized as an inductive bias for language modeling. |
Analyzing Wrap-Up Effects through an Information-Theoretic Lens (2022.acl-short)
Copied to clipboard
| Challenge: | a lack of studies targeting naturalistic sentence-final reading behavior is likely to explain the lack of data on reading time (RT) data is omitted due to the confounding factors introduced by so-called "wrap-up effects" |
| Approach: | They propose to look for a link between “wrap-up effects” and information theoretic quantities such as word and context information content. |
| Outcome: | The proposed model omits data on words at the end of sentences or clauses to control for the confounding factors introduced by wrap-up effects. |