Austria

The syntactic probing literature has been largely limited to shallow structures like dependency trees, which are unable to capture the subtle differences in sub-surface syntactic structures that yield semantic nuances. These structures are captured by theories of syntax like generative syntax, but have not been researched in the LLM literature due to the difficulties in probing these complex structures with many silent, covert nodes. Our work presents a method for overcoming this limitation by deploying Hewitt and Manning&#39;s (2019) dependency-trained probe on sentence constructions whose structural representation is identical in a dependency parse, but differs in theoretical syntax. If a pretrained language model has captured the theoretical syntax structure, then the probe&#39;s predicted distances should vary in syntactically-predicted ways. Using this methodology and a novel dataset, we find evidence that LLMs have captured syntactic structures far richer than previously realized, indicating LLMs are able to capture the nuanced meanings that result from sub-surface differences in structural form.

ACL 2025

Evidence of Generative Syntax in LLMs

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

The syntactic probing literature has been largely limited to shallow structures like dependency trees, which are unable to capture the subtle differences in sub-surface syntactic structures that yield semantic nuances. These structures are captured by theories of syntax like generative syntax, but have not been researched in the LLM literature due to the difficulties in probing these complex structures with many silent, covert nodes. Our work presents a method for overcoming this limitation by deploying Hewitt and Manning's (2019) dependency-trained probe on sentence constructions whose structural representation is identical in a dependency parse, but differs in theoretical syntax. If a pretrained language model has captured the theoretical syntax structure, then the probe's predicted distances should vary in syntactically-predicted ways. Using this methodology and a novel dataset, we find evidence that LLMs have captured syntactic structures far richer than previously realized, indicating LLMs are able to capture the nuanced meanings that result from sub-surface differences in structural form.

workshop paper

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

In recent computational psycholinguistics, Merkx and Frank (2021) showed that surprisals from Transformers demonstrate a closer fit to measures of human reading effort than those from Recurrent Neural Networks (RNNs), suggesting that Transformers may capture the cue-based retrieval-like operations in human sentence processing. On the other hand, explicit incorporation of syntactic structures has been shown to improve LMs' predictive power for human cognitive load---for example, Hale et al. (2018) demonstrated that Recurrent Neural Network Grammars (RNNGs), which integrate RNNs with explicit syntactic structures, account for aspects of human brain activity that vanilla RNNs cannot. In this paper, we test the psychometric predictive power of Composition Attention Grammars (CAGs), the integration of Transformers with explicit syntactic structures, to investigate whether they can provide better fit to human gaze durations than vanilla Transformers and RNNGs by capturing cue-based retrieval-like operations on syntactic structures, which could potentially be involved in human sentence processing. The results of our controlled experiment demonstrate that surprisals from CAGs outperformed those from Transformers and RNNGs, suggesting that syntactic attention in CAGs may serve as a mechanistic implementation of human retrieval from syntactically-constructed memory representations.

Investigating Psychometric Predictive Power of Syntactic Attention

While natural language is processed incrementally, it is unclear whether the syntactic structure prediction process is universal across languages or language-specific. This study investigates this question by revisiting parsing strategies of syntactic language models that incrementally predict both the next token and the associated syntactic structure. Unlike previous studies that have focused on a few strategies, we examine a wide range of strategies by introducing different parameterizations of speculation, which quantifies the degree to which a model predicts syntactic structure before encountering the corresponding tokens. The experiments with 10 typologically diverse languages reveal that the optimal strategy differs depending on the language and the beam size.

Is Incremental Structure Prediction Process Universal across Languages?:Revisiting Parsing Strategy through Speculation

In this paper, we evaluate the capacity of current language technologies to understand Basque and Spanish language varieties. We use NLI as a pivot task and introduce a novel, manually-curated parallel dataset in Basque and Spanish and their corresponding variants. Empirical analysis of comprehensive crosslingual and in-context learning experiments with respectively, encoder-only and decoder-based Large Language Models (LLMs), reveals a performance drop when processing linguistic variations, with more pronounced effects observed in Basque. Error analysis indicates that lexical overlap plays no role, suggesting that linguistic variation represents the primary reason for the lower results. All data and code are publicly available under Attribution-NonCommercial 4.0 International license.

Lost in Variation? Evaluating NLI Performance in Basque and Spanish Geographical Variants

In sentences such as John began the book, the complement noun, lexically denoting an entity, is interpreted as an event. This phenomenon is known in linguistics as complement coercion: the event associated with the verb is not overtly expressed but can be recovered from the meanings of other constituents, context and world knowledge. We investigate whether language models (LMs) can exploit sentence structure and compositional meaning to recover plausible events in complement coercion. For the first time, we tested different LMs in Norwegian, a low-resource language with high syntactic variation in coercion constructions across aspectual verbs. Results reveal that LMs struggle with retrieving plausible events and with ranking them above less plausible ones. Moreover, we found that LMs do not exploit the compositional properties of coercion sentences in their predictions.

Compositionality and Event Retrieval in Complement Coercion: A Study of Language Models in a Low-resource Setting

Knowing which words language learners struggle with is crucial for developing personalised education technologies. In this paper, we advocate for the novel task of "dictionary look-up predication" as a means for evaluating the complexity of words in reading tasks. We release the Dictionary Look-Up development dataset (DLU-dev) and the Dialogue Dictionary Look-Up dataset (D-DLU), which is based on chatbot dialogues. We demonstrate that dictionary look-up is a challenging task for LLMs (results are presented for LLaMA, Gemma and Longformer models). We explore finetuning with the ROC* loss function as a more appropriate loss for this task than the commonly used Binary Cross Entropy (BCE). We investigate the transfer between DLU and the related tasks of Complex Word Identification (CWI) and Semantic Error Prediction (SEP); establishing new state-of-the-art results for SEP.

DLU: Dictionary Look-Up Data and Prediction

Recent work has investigated whether extant neural language models (LMs) have an inbuilt inductive bias towards the acquisition of attested typologically-frequent grammatical patterns as opposed to infrequent, unattested, or impossible patterns using artificial languages (White and Cotterell, 2021; Kuribayashi et al., 2024). The use of artificial languages facilitates isolation of specific grammatical properties from other factors such as lexical or real-world knowledge, but also risks oversimplification of the problem.

In this paper, we examine the use of Generalized Categorial Grammars (GCGs) (Wood, 2014) as a general framework to create artificial languages with a wider range of attested word order patterns, including those where the subject intervenes between verb and object (VSO, OSV) and unbounded dependencies in object relative clauses. 
In our experiments, we exemplify our approach by extending White and Cotterell (2021) and report some significant differences from existing results.

GCG-Based Artificial Languages for Evaluating Inductive Biases of Neural Language Models

This study investigates the generalization abilities of discriminative transformers in Natural Language Inference (NLI) tasks, focusing on their tendency to rely on superficial features and dataset biases rather than genuine linguistic understanding. We argue that performance gaps between training and analysis datasets do not necessarily indicate a lack of knowledge but rather a misalignment between the decision boundaries of the classifier head and the representations learned by the encoder. By analyzing the representation space of NLI models on these datasets, we show that, despite poor accuracy based on final predictions, samples from opposing classes often remain linearly separable in the encoder's representation space. This suggests that the encoders possess sufficient knowledge to perform the NLI task effectively, despite the classifier head's challenges.

Beyond Accuracy: Revisiting Out-of-Distribution Generalization in NLI Models

Recent research has increasingly focused on the extent to which large language models (LLMs) exhibit human-like behavior. In this study, we investigate whether the mental lexicon in LLMs resembles that of humans in terms of lexical organization. Using a word association task—a direct and widely used method for probing word meaning and relationships in the human mind—we evaluated the lexical representations of GPT-4 and Llama-3.1. Our findings reveal that LLMs closely emulate human mental lexicons in capturing semantic relatedness but exhibit notable differences in other properties, such as association frequency and dominant lexical patterns (e.g., top associates). Specifically, LLM lexicons demonstrate greater clustering and reduced diversity compared to the human lexicon, with KL divergence analysis confirming significant deviations in word association patterns. Additionally, LLMs fail to fully capture word association responses patterns in different demographic human groups. Among the models, GPT-4 consistently exhibited a slightly higher degree of human-likeness than Llama-3.1. This study highlights both the potential and limitations of LLMs in replicating human mental lexicons, offering valuable insights for applications in natural language processing and cognitive science research involving LLMs.

Human-likeness of LLMs in the Mental Lexicon

Human lexicons contain many different words that speakers can use to refer to the same object, e.g., purple or magenta for the same shade of color. On the one hand, studies on language use have explored how speakers adapt their referring expressions to successfully commu- nicate in context, without focusing on proper- ties of the lexical system. On the other hand, studies in language evolution have discussed how competing pressures for informativeness and simplicity shape lexical systems, without tackling in-context communication. We aim at bridging the gap between these traditions, and explore why a soft mapping between referents and words is a good solution for communication, by taking into account both in-context communication and the structure of the lexicon. We propose a simple measure of informativeness for words and lexical systems, grounded in a visual space, and analyze color naming data for English and Mandarin Chinese. We conclude that optimal lexical systems are those where multiple words can apply to the same referent, conveying different amounts of information. Such systems allow speakers to maximize communication accuracy and minimize the amount of information they convey when communicating about referents in contexts.

Why do objects have many names? A study on word informativeness in language use and lexical systems

Retrieval Augmented Generation (RAG) complements the knowledge of Large Language Models (LLMs) by leveraging external information to enhance response accuracy for queries. This approach is widely applied in several fields by taking its advantage of injecting the most up-to-date information, and researchers are focusing on understanding and improving this aspect to unlock the full potential of RAG in such high-stakes applications. However, despite the potential of RAG to address these needs, the mechanisms behind the confidence levels of its outputs remain underexplored. Our study focuses on the impact of RAG, specifically examining whether RAG increases the confidence of LLM outputs in the medical domain. We conduct this analysis across various configurations and models. We evaluate confidence by treating the model's predicted probability as its output and calculating several evaluation metrics which include calibration error method, entropy, best probability, and accuracy. Experimental results across multiple datasets confirmed that certain models possess the capability to judge for themselves whether an inserted document relates to the correct answer. These results suggest that evaluating models based on their output probabilities determine whether they function as generators in the RAG framework. Our approach allows to evaluate whether the models handle retrieved documents. and improving this aspect to unlock the full potential of RAG in such high-stakes applications. However, despite the potential of RAG to address these needs, the mechanisms behind the confidence levels of its outputs remain underexplored. Our study focuses on the impact of RAG, specifically examining whether RAG increases the confidence of LLM outputs in the medical domain. We conduct this analysis across various configurations and models. We evaluate confidence by treating the model's predicted probability as its output and calculating several evaluation metrics which include calibration error method, entropy, best probability, and accuracy. Experimental results across multiple datasets confirmed that certain models possess the capability to judge for themselves whether an inserted document relates to the correct answer. These results suggest that evaluating models based on their output probabilities determine whether they function as generators in the RAG framework. Our approach allows to evaluate whether the models handle retrieved documents.

Content not yet available

Downloads

Next from ACL 2025

Investigating Psychometric Predictive Power of Syntactic Attention

Similar lecture

Mechanisms vs. Outcomes: Probing for Syntax Fails to Explain Performance on Targeted Syntactic Evaluations

Content not yet available

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2025

Investigating Psychometric Predictive Power of Syntactic Attention

Similar lecture

Mechanisms vs. Outcomes: Probing for Syntax Fails to Explain Performance on Targeted Syntactic Evaluations

Downloads