United States

Autoregressive language models process text token-by-token, yet the human language system operates over structured multi-word units. We propose that phrase-level representations provide a key granularity at which LLM internal states align with human language processing, and that phrase-segmentation fine-tuning can amplify this alignment. Using Meta-Llama-3.1-8B (base and fine-tuned) and the ReadBrain naturalistic reading dataset (eye tracking and fMRI), we provide three converging lines of evidence. First, phrase-level attention features predict eye movements significantly better than word-level features---a phrase superiority effect confirmed by partial correlation analysis to reflect genuine linguistic boundaries rather than aggregation artifacts. Second, fMRI encoding analyses show that fine-tuning selectively amplifies phrase encoding in left Superior Temporal Gyrus (STG) and pars opercularis at mid-to-high transformer layers, with no improvement for word representations. Third, representational similarity analysis confirms that fine-tuning improves model--brain geometric alignment specifically for phrases. The persistent null for phrase-minus-word residuals suggests that phrases function as holistic, chunk-like units rather than products of separable compositional operations. Together, these results identify phrase-level representation as a linguistic granularity at which LLM attention shows particularly strong correspondence with human language cortex, and demonstrate that targeted training objectives can selectively strengthen model--brain alignment.

ACL 2026

Do Large Language Models Acquire Phrase-Based Processing? Evidence from Eye Movements, fMRI Encoding, and Model-Brain Alignment After Fine-Tuning

fmri encoding

phrase processing

model-brain alignment

representational similarity analysis

eye tracking

large language models

### Welcome to the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL) 
Continuing its mission of expanding and involving the science community of all European countries, EACL had selected the Ukrainian community for the 16th EACL, which took place online due to the COVID pandemic. Unfortunately, the ongoing war made the organisation in Kyiv impossible. Considering the importance of physical interaction among researchers, especially after the restrictions imposed by the COVID pandemic, in addition to an online mode, the conference will be held in Dubrovnik, Croatia, from 2 to 6 of May, 2023. The original aim of strengthening the connection with the Ukrainian community will not change as our program will feature a dedicated session and a workshop to highlight work on Ukrainian language technologies. As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language.

You need to register for the conference in order to access this site. Please visit https://2023.eacl.org/registration for more information.

EACL 2023

As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language.

workshop paper

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields 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.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

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 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

Children’s acquisition of filler-gap dependencies has been argued by some to depend on innate grammatical knowledge, while others suggest that the distributional evidence available in child-directed speech suffices. Unfortunately, the relevant input is difficult to quantify at scale with fine granularity, making this question difficult to resolve. We present a system that identifies three core filler-gap constructions in spoken English corpora -- matrix wh-questions, embedded wh-questions, and relative clauses -- and further identifies the extraction site (i.e., subject vs. object vs. adjunct). Our approach combines constituency and dependency parsing, leveraging their complementary strengths for construction classification and extraction site identification. We validate the system on human-annotated data and find that it scores well across most categories. Applying the system to 57 English CHILDES corpora, we are able to characterize children’s filler-gap input and their filler-gap production trajectories over the course of development, including construction-specific frequencies and extraction-site asymmetries. The resulting fine-grained labels enable future work in both acquisition and computational studies, which we demonstrate with a case study using filtered corpus training with language models.

What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies

Human discourse comprehension includes graded expectations about whether a speaker is likely to lie. If language models capture human-like discourse expectations, they should be sensitive not only to factual consistency but also to lie expectancy as a contextual probability from complex pragmatic cues. We test this idea using discourse scenarios with varying incentives to deceive. Human lie probability is estimated from free continuations, and model lie expectancy is derived from the probability mass assigned to human-produced lie versus truth continuations. Across Qwen3 models, likelihood-derived lie mass aligns strongly with human lie expectancy. The best performance comes from the base checkpoints. By contrast, post-trained and mode-specialized variants show weaker alignment. 
Qualitative analysis suggests a structured error pattern: models tend to overpredict lies when a response directly conflicts with known facts, but underpredict them when lie expectancy depends more on contextual pressures such as politeness, self-protection, or strategic gain. These results suggest that graded lie expectancy is recoverable from model continuation probabilities and can be learned, at least in part, through the ordinary next-token prediction objective.

Graded Expectations: Do Large Language Models Show Human-like Sensitivity to the Likelihood of Deceptive Speech Acts?

Research in sound discrimination demonstrates that listeners exhibit reduced sensitivity to acoustic differences between allophones, as opposed to phonemes. Previous studies indicates that highly predictable, complementary distribution of allophones contributes to this limited sensitivity by providing strong contextual cues. Building on these insights, this study investigates the role of predictability in sound discrimination within a supervised contrastive learning framework. Specifically, we examine how varying levels of predictability affect the ability to distinguish sounds and whether this influence is categorical or gradual. Additionally, we explore the interaction between acoustic distance and predictability, as well as how the presence of other contrasts within a language modulates this process. Our findings indicate that only full predictability leads to a significant decline in discrimination performance, demonstrating a categorical effect. This impairment can be alleviated as acoustic distance increases. Moreover, the presence of additional contrasts sharing the relevant acoustic dimension enhances discriminability, showing the importance of contextual contrasts in speech perception.

Roles of Predictability and Acoustic Distance in Sound Discrimination via Contrastive Learning

Given a listener's native language, some non-native contrasts may be harder to discriminate than others. The computation required to mimic this variable difficulty is not yet known. The present work approaches this question by training small supervised feedforward neural networks to perform Spanish vowel classification and then evaluating model classification of Catalan vowels, thereby approximating Spanish-listeners' cross-linguistic perception of Catalan. Vowels were extracted from Spanish and Catalan audio corpora, respectively. Ultimately, Spanish models exhibited expected misperception of Catalan's /e/-/ɛ/, /o/-/ɔ/, and /ɛ/-/a/ contrasts; Spanish-dominant listeners have difficulty perceiving these contrasts, and Spanish models classified Catalan /ɛ/ as /e/ or /a/, and Catalan /ɔ/ as /o/. This demonstrates that small supervised neural models are capable of making specific, cross-linguistic perceptual predictions given realistic input.

Small Neural Networks as Models of Cross-Linguistic Speech Perception

This study explores the factors that predict hesitations and disfluencies in Hindi spontaneous speech. In particular, we investigate the role of lexical, syntactic, prosodic, and phonological factors. Our results provide corroborating evidence that disfluencies arise due to disruptions at various levels of the planning process, such as lexical retrieval and phonological planning (Goldman-Eisler, 1961; Levelt, 1983; Fraundorf and Watson, 2013), or more broadly due to planning delay (Clark and Fox Tree, 2002). Further, disfluencies increase due to processing load (Bortfeld et al., 2001) and due to strain on working memory (Dammalapati et al., 2021). While previous research (e.g., Bell et al., 2009; Dammalapati et al., 2021) highlights the importance of downstream difficulty on disfluencies, our results suggest that previously encountered difficulties can also lead to disfluencies. Together, these findings highlight the need for increased cross-linguistic research to understand the nature of incrementality and monitoring of the production system across languages.

Determinants of Hesitation and Disfluency in Hindi Spontaneous Speech

In series of recent papers, Anttila, Magri, and their collaborators extend the Greenbergian implicational universals from the categorical to the probabilistic setting and use the resulting implicational universals to argue that ME typologies are less restrictive than the NHG typologies corresponding to the same constraint sets. This paper contributes to this research project. We provide a new formal constraint condition on ME universals, focus on a class of implicational universals that hold ''by transitivity'' in categorical OT and HG as well as in probabilistic NHG, and we use our new formal condition to explain why this class of universals instead fails in probabilistic ME.

MaxEnt fails at reasoning by transitivity

Encoder–decoder neural networks with high-dimensional (e.g. $d$=$300$-–$500$) embeddings and hidden layers can be used to model a variety of morphophonological phenomena as sequence-to-sequence mappings, achieving high accuracy across languages and patterns. We show here that these high-capacity models are overparameterized, at least for the task of morphological inflection, and that simpler and smaller networks can perform near ceiling on the task of inflecting Turkish stems. Moreover these reduced-capacity models encode linguistically relevant information even when they are too small to succeed at the inflectional task.

Empirical lower bounds on encoder-decoder capacity for Turkish inflection

We show that semantic proto-role properties can be used as a tool to measure implicit human perceptions of agency and patiency of entities in human-generated text. First, we demonstrate that silver-generated semantic proto-role property labels are strongly correlated with both human judgment and a probabilistic text-based measure of anthropomorphism. Then, we use our measure to quantify linguistic idiosyncrasies across different AI-related Reddit communities. Our measure shows that subreddits dedicated to discussing AI companionship ascribe higher sentience to "bots" and higher agency to "companies" when compared to other subreddits. This phenomenon reveals not only the unique way in which chatbots are anthropomorphized in such subreddits, but also the users' keen awareness of their power imbalance with the companies that created the chatbots.

Measuring Perceptions of Personhood with Semantic Proto-role Properties

Many kinds of logical systems have been employed in constructing formal languages to model phonological phenomena. A common theme among them is that the systems compile into finite automata. Two questions naturally arise. Can a given phenomenon be described with another logical system? And, if so, what is that description?

To the first question, algebraic techniques are well established through deep connections with logic and automata. To the second, the situation is less clear. Translations from automata are established for first-order and monadic second-order logics under precedence, but these may not translate easily to the simpler systems we often use. Translations for simple cases of restricted propositional logic (strictly local or strictly piecewise languages) are established, but insufficient to describe attested phenomena.

The present work establishes a general way to handle many systems in between. Specifically, we show how to translate between certain kinds of algebraic varieties $\mathbf{V}$ (systems defined by universally satisfied identities) and associated logical systems, then use decomposition to handle classes of the form $\mathbf{V}\ast\mathbf{D}$, where the notion of “symbol” is replaced by “$k$-block”. With this, we handle several (unrestricted) propositional logics, facilitating logical description of natural language.

A Family of Effective Methods for Decompiling Canonical Acceptors, Instantiated for Languages of Dot-Depth One and Tier-Based Extensions

Large language models (LLMs) exhibit remarkable few-shot learning capabilities, yet the role of syntactic structure in demonstration examples remains unexplored. Drawing on psycholinguistic research on structural priming, we investigate whether syntactic patterns in few-shot prompts influence LLM outputs and task performance. We conduct systematic experiments across four model families (Llama, Mistral, Qwen, Gemma) using four syntactic constructions (passive voice, cleft sentences, dative alternation, particle placement). Our results reveal robust syntactic priming effects, with priming strength ranging from 1.3× to 6.4× depending on construction type, indicating that models are substantially more likely to produce constructions matching demonstration syntax. Critically, we find that priming strength shows a positive trend with model size (r = 0.85, p = 0.068), with effects intensifying from 7B to 14B parameter models. We demonstrate that priming is construction-specific rather than reflecting general stylistic preferences, and that priming effects persist across multiple intervening sentences. Analysis across three task types (sentence completion, paraphrase generation, story continuation) reveals that syntactic structure in demonstrations influences output style, and that models produce primed constructions even when the task calls for a different syntactic form. These findings have immediate implications for prompt engineering and reveal that LLMs encode syntactic abstractions beyond surface-level pattern matching. We release our benchmark, SyntaxPrime-ICL, containing controlled examples across multiple constructions for evaluating syntactic priming in few-shot contexts.

Downloads

Next from ACL 2026

What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies

Similar lecture

PiC: A Phrase-in-Context Dataset for Phrase Understanding and Semantic Search

.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 2026

What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies

Similar lecture

PiC: A Phrase-in-Context Dataset for Phrase Understanding and Semantic Search

Downloads