Papers by Tal Linzen
How Much Do Language Models Copy From Their Training Data? Evaluating Linguistic Novelty in Text Generation Using RAVEN (2023.tacl-1)
Copied to clipboard
| Challenge: | Current language models generate high-quality text, but are they copying it or have they learned generalizable linguistic abstractions? |
| Approach: | They propose a suite of analyses for assessing the novelty of generated text . they focus on sequential structure (n-grams) and syntactic structure (syntactical structure). |
| Outcome: | The proposed model-generated text is as novel as the baseline human-generated model- generated text, but it is copied substantially, the authors show . |
Verb Conjugation in Transformers Is Determined by Linear Encodings of Subject Number (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work on LMs' internal vector representations has demonstrated the existence of syntactic information in those vectors, but has not described how LM models use this information. |
| Approach: | They propose a causal intervention paradigm to show that some linguistic features are represented in a linear format. |
| Outcome: | The proposed model can generate verb conjugations 91% correct before and 85% incorrect after the intervention. |
A Systematic Comparison of Syllogistic Reasoning in Humans and Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Psychologists have documented several ways in which humans’ inferences deviate from the rules of logic. |
| Approach: | They focus on syllogisms, which are inferences from two simple premises, and show that larger models are more logical than smaller ones. |
| Outcome: | The results show that language models often mimic human biases, but overcome them in some cases. |
When a sentence does not introduce a discourse entity, Transformer-based models still sometimes refer to it (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing theories on how humans track discourse entities are based on the idea that humans maintain explicit memory representations for each entity that encode all properties of an entity and its relation to other entities. |
| Approach: | They adapt the psycholinguistic assessment of language models paradigm to higher-level linguistic phenomena and introduce an English evaluation suite that targets the knowledge of the interactions between sentential operators and indefinite NPs. |
| Outcome: | The evaluation suite targets the knowledge of the interactions between sentential operators and indefinite NPs and the models are challenged by multiple NP's and their behavior is not systematic. |
Does Putting a Linguist in the Loop Improve NLU Data Collection? (2021.findings-emnlp)
Copied to clipboard
Alicia Parrish, William Huang, Omar Agha, Soo-Hwan Lee, Nikita Nangia, Alexia Warstadt, Karmanya Aggarwal, Emily Allaway, Tal Linzen, Samuel R. Bowman
| Challenge: | Many datasets for training and evaluating natural language understanding (NLU) models contain systematic artifacts that are identified only after data collection is complete. |
| Approach: | They propose to have linguists identify artifacts and gaps in the data and communicate with non-expert crowdworkers to adjust task instructions and incentives. |
| Outcome: | The proposed protocol does not increase accuracy on out-of-domain test sets, and adds a chatroom does not. |
Coloring the Blank Slate: Pre-training Imparts a Hierarchical Inductive Bias to Sequence-to-sequence Models (2022.findings-acl)
Copied to clipboard
| Challenge: | Sequence-to-sequence models fail to generalize in hierarchy-sensitive manner when performing syntactic transformations. |
| Approach: | They evaluate whether seq2seq models generalize hierarchically on two transformations . they use pre-trained models and their multilingual variants to test their generalization . |
| Outcome: | The proposed models generalize hierarchically on two transformations in English and German. |
Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark (2022.tacl-1)
Copied to clipboard
| Challenge: | Knowledge-grounded dialogue systems powered by large language models often generate responses that, while fluent, are not attributable to a relevant source of information. |
| Approach: | They propose to evaluate the validity of 12k dialogue turns generated by neural dialogue systems trained on three knowledge-grounded dialogue corpora and to use them to analyze eight evaluation metrics. |
| Outcome: | The proposed evaluation metrics rely on spurious correlations, do not reliably distinguish attributable abstractive responses from unattributable ones, and perform substantially worse when the knowledge source is longer. |
Studying the Inductive Biases of RNNs with Synthetic Variations of Natural Languages (N19-1)
Copied to clipboard
| Challenge: | Recent studies have identified both strengths and limitations of recurrent neural networks (RNNs) in applied natural language processing tasks. |
| Approach: | They propose a paradigm that addresses typological differences between languages . they create synthetic versions of English and train them to predict agreement features . |
| Outcome: | The proposed model improves on predicting agreement with subject and object, suggesting that RNNs have a recency bias. |
How Can We Accelerate Progress Towards Human-like Linguistic Generalization? (2020.acl-main)
Copied to clipboard
| Challenge: | a new evaluation paradigm, Pretraining-Agnostic Identically Distributed evaluation, is needed . authors argue that it rewards models that can be trained on massive amounts of data, several orders of magnitude more than a human can expect to be exposed to. |
| Approach: | a position paper describes and critiques the Pretraining-Agnostic Identically Distributed evaluation paradigm . paradigm favors simple, low-bias architectures that can be scaled to process vast amounts of data . authors advocate for supplementing or replacing PAID with paradigms that reward architectures . |
| Outcome: | a new evaluation paradigm favors simple, low-bias architectures that can be scaled to process vast amounts of data. a san francisco-based study finds that the paradigm rewards architectures which generalize as quickly and robustly as humans. |
COGS: A Compositional Generalization Challenge Based on Semantic Interpretation (2020.emnlp-main)
Copied to clipboard
| Challenge: | Natural language is characterized by compositionality: meaning of complex expressions is constructed from the meanings of its constituent parts. |
| Approach: | They propose a semantic parsing dataset based on a fragment of English to assess compositional generalization abilities. |
| Outcome: | The proposed model can generalize meanings in a given sentence in 96–99% of the tests, but generalization accuracy is lower and the generalization sensitivity is higher. |
A Neural Model of Adaptation in Reading (D18-1)
Copied to clipboard
| Challenge: | Several studies suggest that readers do adapt their lexical and syntactic predictions to the current context. |
| Approach: | They propose to add a simple adaptation mechanism to a neural language model to improve predictions of reading times. |
| Outcome: | The proposed model improves predictions of human reading times compared to a non-adaptive model. |
Colorless Green Recurrent Networks Dream Hierarchically (N18-1)
Copied to clipboard
| Challenge: | Recurrent neural networks (RNNs) can induce non-trivial properties of language. |
| Approach: | They investigate whether RNNs can track hierarchical syntactic structure . they include nonsensical sentences where RNN cannot rely on semantic cues . |
| Outcome: | The proposed models can predict long-distance agreement in nonsensical sentences in Italian and English. |
Causal Analysis of Syntactic Agreement Mechanisms in Neural Language Models (2021.acl-long)
Copied to clipboard
| Challenge: | Targeted syntactic evaluations have demonstrated the ability of language models to perform subject-verb agreement given difficult contexts. |
| Approach: | They apply causal mediation analysis to pre-trained neural language models to investigate their models' preferences for grammatical inflections and whether neurons process subject-verb agreement similarly across sentences with different syntactic structures. |
| Outcome: | The proposed model can predict correct token from grammatically minimally different continuations with high accuracy even in difficult contexts. |
How to Plant Trees in Language Models: Data and Architectural Effects on the Emergence of Syntactic Inductive Biases (2023.acl-long)
Copied to clipboard
| Challenge: | a recent study found that pre-training can teach language models to rely on hierarchical syntactic features . aaron ramirez: we find that pretraining on simpler language induces a hierarchic bias . |
| Approach: | They find that pre-training can teach language models to rely on hierarchical syntactic features . authors: this suggests that in cognitively plausible language acquisition settings, models may be more data-efficient . |
| Outcome: | a recent study shows that pre-training can teach language models to rely on hierarchical features . the findings suggest that in plausible language acquisition settings, language models may be more data-efficient than previously thought . |
Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference (P19-1)
Copied to clipboard
| Challenge: | lexical overlap heuristics are effective for frequent example types but break down in more challenging cases. |
| Approach: | They propose to use a set of examples to test whether a sentence entails another . they propose to adopt three fallible syntactic heuristics for statistical NLI models . |
| Outcome: | The proposed model performs poorly on the HANS dataset, suggesting they have adopted these heuristics. |
Does Syntax Need to Grow on Trees? Sources of Hierarchical Inductive Bias in Sequence-to-Sequence Networks (2020.tacl-1)
Copied to clipboard
| Challenge: | Inductive biases can arise from any aspect of the model architecture, study finds . we investigate which architectural factors affect how models generalize . |
| Approach: | They investigate which architectural factors affect generalization behavior of neural network models . they use English question formation and English tense reinflection as test cases . |
| Outcome: | The findings suggest that human-like generalization requires architectural syntactic structure. |
Frequency Effects on Syntactic Rule Learning in Transformers (2021.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models perform well on a variety of linguistic tasks that require symbolic reasoning, raising the question of whether such models implicitly represent abstract symbols and rules. |
| Approach: | They investigate the performance of BERT on English subject–verb agreement by analyzing word frequency and absolute frequency of verb forms. |
| Outcome: | The proposed model generalizes well to subject–verb pairs that never occurred in training, suggesting a degree of rule-governed behavior. |
Syntactic Data Augmentation Increases Robustness to Inference Heuristics (2020.acl-main)
Copied to clipboard
| Challenge: | Pretrained neural models lack sensitivity to word order on controlled challenge sets . augmentation methods that improve accuracy on standard training sets may be a problem . |
| Approach: | They propose to augment standard training sets with syntactically informative examples by applying syntastic transformations to sentences from the MNLI corpus. |
| Outcome: | The proposed method improved BERT’s accuracy on controlled examples that diagnose sensitivity to word order from 0.28 to 0.73 without affecting performance on the MNLI test set. |
SLOG: A Structural Generalization Benchmark for Semantic Parsing (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing compositional generalization benchmarks focus on lexical generalisation, the interpretation of novel lexicals in syntactic structures familiar from training. |
| Approach: | They propose a semantic parsing dataset that extends COGS with 17 structural generalization cases to evaluate how well models generalize to new complex linguistic expressions. |
| Outcome: | The proposed model generalization accuracy is far below the near-perfect accuracy of existing models on COGS, demonstrating the role of SLOG in foregrounding the large discrepancy between models’ lexical and structural generalization capacities. |
What Goes Into a LM Acceptability Judgment? Rethinking the Impact of Frequency and Length (2025.naacl-long)
Copied to clipboard
| Challenge: | Prior work on LM and acceptability judgments treat these effects uniformly across models, making a strong assumption that models require the same degree of adjustment to control for length and unigram frequency effects. |
| Approach: | They propose a linking theory where the optimal level of adjustment is estimated from data via learned parameters for length and unigram frequency. |
| Outcome: | The proposed theory outperforms a commonly used linking theory for acceptability—SLOR—across two families of transformer LMs. |
The Impact of Depth on Compositional Generalization in Transformer Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | In this paper, we test the hypothesis that deeper transformers generalize more compositionally. |
| Approach: | They propose to add layers to transformers to generalize more compositionally . they propose to fine-tune the models so that the total number of parameters is constant . |
| Outcome: | The proposed model generalizes more compositionally than shallower models, but returns diminish . the proposed model can be made shallower without sacrificing performance . |
Quantity doesn’t buy quality syntax with neural language models (D19-1)
Copied to clipboard
| Challenge: | Recurrent neural network language models can learn to predict upcoming words with remarkably low perplexity . but in syntactically complex contexts, they often assign unexpectedly high probabilities to ungrammatical words . |
| Approach: | They investigate whether recurrent neural networks can learn to predict upcoming words with remarkably low perplexity. |
| Outcome: | The proposed models perform worse than GPT and BERT in some constructions than LSTMs in other contexts. |
In-context Learning Generalizes, But Not Always Robustly: The Case of Syntax (2024.naacl-long)
Copied to clipboard
| Challenge: | In-context learning is a common method for teaching large language models new tasks . given labeled examples in the input context, the model learns to perform the task without weight updates. |
| Approach: | They examine whether models guided via ICL infer the underlying structure of the task defined by the context or rely on superficial heuristics that only generalize to identically distributed examples. |
| Outcome: | The proposed model generalizes syntactically or linearly on out-of-distribution examples . the proposed model is able to generalize better on pre-trained models . |
Can You Learn Semantics Through Next-Word Prediction? The Case of Entailment (2024.findings-acl)
Copied to clipboard
| Challenge: | et al. argued that sentence co-occurrence probabilities should reflect entailment . but it is unclear whether probabilities predicted by neural LMs encode enanglement based on their theory . |
| Approach: | They propose a test that decodes entailment relations between natural sentences . they argue that the test that predicts a flipped test does not account for redundancy . |
| Outcome: | The proposed test can decode entailment relations between natural sentences, but not perfectly. |
A Language Model with Limited Memory Capacity Captures Interference in Human Sentence Processing (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Theories of human sentence processing can be divided into two broad categories: expectation-based theories and memory-based ones. |
| Approach: | They propose to integrate expectations and retrieval from working memory into a unified cognitive model that can capture syntactic and semantic interference effects observed in human experiments. |
| Outcome: | The proposed model captures syntactic and semantic interference effects observed in human experiments. |
Between Circuits and Chomsky: Pre-pretraining on Formal Languages Imparts Linguistic Biases (2025.acl-long)
Copied to clipboard
| Challenge: | Pretraining language models on formal language can improve their acquisition of natural language, but it is computationally limited. |
| Approach: | They hypothesize that effective transfer occurs when the formal language captures hierarchical dependencies present in natural language and remains within the computational limitations of the model architecture. |
| Outcome: | The proposed model can achieve lower loss on natural language and better generalization with a 33% smaller token budget than training on a matched amount of natural language. |
Rapid Word Learning Through Meta In-Context Learning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for few-shot word learning are underexplored. |
| Approach: | They propose a method that trains language models to generate new examples of a word’s usage given a few in-context examples, using a special placeholder token to represent the new word. |
| Outcome: | The proposed method trains language models to generate new examples of a word’s usage given a few in-context examples, using a special placeholder token to represent the new word. |
Language Models Struggle to Use Representations Learned In-Context (2026.acl-long)
Copied to clipboard
| Challenge: | a recent study shows that large language models are capable of inducing rich representations of data that are seen in-context . a novel task, adaptive world modeling, shows that even the most performant LLMs cannot reliably leverage novel semantics defined in-constitut. |
| Approach: | They propose to use in-context representations to induce rich representations of data . they also propose to probe models using a novel task to enable flexible deployment . |
| Outcome: | The proposed model can use in-context representations to complete simple downstream tasks. |
Cross-Linguistic Syntactic Evaluation of Word Prediction Models (2020.acl-main)
Copied to clipboard
| Challenge: | A range of studies have concluded that neural word prediction models can distinguish grammatical from ungrammatically sentences with high accuracy. |
| Approach: | They propose to use CLAMS to evaluate LSTM and multilingual BERT models. |
| Outcome: | The proposed model can learn syntax on English, French, German, Hebrew and Russian, and LSTM language models on multilingual and multilingual models. |
Representations of Syntax [MASK] Useful: Effects of Constituency and Dependency Structure in Recursive LSTMs (2020.acl-main)
Copied to clipboard
| Challenge: | a constituency-based network generalizes more robustly than a dependency-based one . a sequential LSTM can learn the appropriate rules governing these dependencies . |
| Approach: | They evaluate whether constituency-based networks introduce biases for syntactic structure . they find that constituency networks generalize more robustly than dependency networks . |
| Outcome: | The proposed model generalizes more robustly than a dependency-based model, the study shows . it shows that data augmentation can improve the robustness of the model on small data sets. |
Targeted Syntactic Evaluation of Language Models (D18-1)
Copied to clipboard
| Challenge: | Recent advances have led to an explosion of neural network-based LM architectures. |
| Approach: | They propose to supplement perplexity with a metric that assesses whether a language model can predict the grammatical sentence more accurately than an ungrammatically-based model. |
| Outcome: | The proposed model performed poorly on many of the constructions. |
Improving Compositional Generalization with Latent Structure and Data Augmentation (2022.naacl-main)
Copied to clipboard
| Challenge: | Generic unstructured neural networks struggle on out-of-distribution compositional generalization. |
| Approach: | They propose a method to recombinate examples from a model called Compositional Structure Learner and add them to a pre-trained sequence-to-sequence model. |
| Outcome: | The proposed model is even stronger than a T5-CSL ensemble on two real world compositional generalization tasks. |
Multilingual Prompting for Improving LLM Generation Diversity (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models lack cultural representation and diversity in their generations . lack of demographic diversity can lead to unfair lack of exposure of artists . |
| Approach: | They propose a prompting method which generates several variations of a base prompt with added cultural and linguistic cues from several cultures, generates responses, and then combines the results. |
| Outcome: | The proposed method outperforms existing diversity-enhancing techniques . it can generate multiple variations of a base prompt with cultural cues from multiple cultures . |
How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech (2023.acl-long)
Copied to clipboard
| Challenge: | LSTMs and Transformers perform well at capturing the surface statistics of child-directed speech, but both model types generalize in a way consistent with an incorrect linear rule than the correct hierarchical rule. |
| Approach: | They train LSTMs and Transformers on text from the CHILDES corpus and evaluate what they learn about English yes/no questions. |
| Outcome: | The proposed models perform well at capturing the surface statistics of child-directed speech, but generalize more consistent with an incorrect linear rule than the correct hierarchical rule. |