United States

Naturalistic language comprehension often involves interpretations that go beyond literal meaning. In continuous narratives, literal and non-literal meanings are tightly intertwined, making them difficult to distinguish computationally. Here, we introduce a framework that separates literal sentence representations from human-annotated non-literal interpretations for model-brain alignment. Using fMRI data recorded during passive listening to the Chinese version of The Little Prince, we annotated sentences containing non-literal meaning with human-written interpretations of their implied meaning. We derived sentence representations from LLaMA3.1-8B and evaluated their correspondence with neural activity using whole-brain encoding models. Literal representations aligned strongly with left-lateralized frontotemporal regions, whereas non-literal interpretations showed broader right-hemisphere involvement. Combining the two further improved the performance of the encoding in the bilateral temporal and dorsal frontal cortices, suggesting that naturalistic comprehension engages complementary levels of meaning.

ACL 2026

Non-literal Meaning Representation in the Brain during Naturalistic Listening

non-literal meaning

naturalistic language comprehension

metaphor

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. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<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>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>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.

**Online Registration Form**: https://acl.swoogo.com/acl2026

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.

workshop paper

This paper investigates whether tonotactic learning differs across representations and learning models. We conduct an experiment using the same dataset encoded in three representations: segments, features, and autosegmental representations (ARs). To the extent possible, two learning models are evaluated, the Maximum Entropy (MaxEnt) model and the Bottom-Up Factor Inference Algorithm (BUFIA), to examine how learning outcomes interact with both model type and representations. A follow-up experiment further explores the roles of frequency and complexity thresholds. The results show that (1) AR-based learning gives the strongest overall performance; (2) there is no consistent advantage between segmental and featural representations across learning models; (3) MaxEnt performance improves substantially when frequency information is introduced and lastly (4) the effects of complexity bounds interact with representation type and frequency information. These findings suggest that tonotactic learning benefits from structurally explicit representations. Overall this work highlights the importance of using linguistically meaningful representations into learning.

What Matters in Tonotactic Learning

Korean stem-final conjugations illustrate the interaction between lexical exceptionality and heterogeneous phonological processes. When /p/-, /t/-, and /s/-final stems occur before vowel-initial suffixes, the irregular classes in these paradigms undergo intervocalic lenition, each adopting a distinct alternation strategy. This study explores how learners acquire lexically-specific phonological patterns given multiple possible solutions. Specifically, it employs a lexically scaled MaxEnt model \citep{hughto2019learning,linzen2013lexical} to learn these paradigm-specific alternations. We run simulations under two learning scenarios: (1) when repair strategies occur at equal frequencies and (2) when one strategy significantly outnumbers the others. Results show that the model favors a \emph{least-cost} solution by treating statistically dominant classes of morphemes as the general pattern. We conclude by discussing the model's sensitivity to lexical statistics, predictions for empirical testing, as well as its implications and challenges for language acquisition.

Lexical exceptionality in paradigm-specific learning: modeling stem-final obstruent alternations in Korean verbs and adjectives

A question for computational linguistics has been to what degree do language models encode case information. However, the majority of the work has focused on structural cases (cases which change when the syntactic configuration changes). On the other hand, inherent cases (which are assigned by specific lexical items and do not change if the syntactic configuration changes) have not been investigated. This paper sets out to investigate if German language models distinctly encode inherent dative from structural accusative and nominative. We conducted a linguistic probing investigation where probes are trained on contextual word embeddings of active nominative, accusative, and dative arguments to predict if passivised datives are analysed as a structural nominative. We provide a cased and caseless version of the experiment. Our results suggest that when case information is removed language models can distinguish between inherent dative and structural accusative, regardless of argument position, due to verb information. However, language models cannot distinguish between structural nominative and inherent dative when the dative appears in a position where there is an expected nominative, due to over-relying on surface patterns.

An LLM Investigation into Inherent and Structural Case Representation: a German Case Study

Context-dependent morphophonological rules -- consonant gradation, vowel harmony, possessive suffix assimilation -- lack a compositional formalization: cascades of finite-state transducers (FSTs) require multiplicative state explosion to encode rule interactions, and neural models provide no formal account of the underlying processes. We formalize these rules as coKleisli arrows of the Writer comonad (DeletionSet x Zipper), where extend lifts a local rule to a global transformation and coKleisli composition chains rules with strict associativity. This resolves length-changing operations -- notably consonant deletion -- algebraically, restoring compositionality that FST cascades sacrifice. Thirteen coKleisli arrows replace 874 FST continuation classes at the morphophonological rule level (67:1 reduction via orthogonal composition), and the same abstraction unifies character-level, morpheme-level, and sentence-level processing. Evaluated on Universal Dependencies (UD) Finnish-TDT, the system achieves 83.92% universal part-of-speech (UPOS) accuracy with rule-only disambiguation (94.66% with a lightweight suffix tagger), and the coKleisli pipeline enables bidirectional morphology -- a MorphGenerator reuses the same arrows for generation.

Comonadic Morphophonology: A Compositional Framework for Context-Dependent Morphological Rules in Finnish

Models of morphophonological learning have focused primarily on concatenative processes, leaving the challenges of non-concatenative morphology largely unaddressed. Reduplication, the systematic copying operation (e.g., Ilokano pluralization [kal-kald\'IN}] `goats'), is particularly revealing because successful learning requires the joint inference of prosodic templates that govern copying, underlying representations (URs) of stems and other affixes, and the phonological grammar. In this paper, we present a learner that tackles this challenge by allowing reduplication to be learned alongside general morphophonemic alternations, a combination that, to our knowledge, has not been directly modeled in prior computational work. We show that the learner successfully captures the attested typology of reduplication–phonology interaction.

Learning reduplicative templates as hidden structures: the case of reduplication-phonology interactions

This paper investigates the learnability of interacting phonological processes by restricting the hypothesis space to a subregular class of functions. Interacting processes can be modeled as function composition, where the output of one function serves as the input to another. We focus specifically on interactions between two simplex Input Strictly Local (ISL$_2$) functions, a proper subclass of the ISL function class. We propose a decomposition algorithm that reconstructs both the individual component processes and their relative ordering by exploiting structural properties of simplex ISL$_2$ transducers and their compositions. This work provides an initial step toward understanding how learners can infer not only single phonological processes, but structured interactions between processes.

Learning Process Interaction Through Simplex ISL Transducers

Few studies have examined how large language models (LLMs) handle negative polarity items (NPIs) in languages other than English. This study investigates NPI licensing in Korean, an under-explored language that is subject to a well established syntactic constraint. Compared to human judgments, Korean-trained models exhibit clear difficulties in detecting this requirement, underscoring a persistent gap between human and model performance on syntax-sensitive phenomena. The consistently poor performance of all three GPT language models examined suggests that the syntactic properties of NPI licensing are not adequately captured. This study thus contributes to the growing literature on linguistic dependencies in language models, while extending evaluation to underrepresented languages. (We will be able to present our work in person.)

Beyond linearity and beyond English: Language model performance on Korean NPI licensing

Zipf's law of abbreviation, a negative correlation between word length and frequency, has been shown to apply in many languages, and is widely regarded as a linguistic universal. However, it has also been shown that word length correlates strongly with surprisal, or average information content. The relative strengths of these correlations have been used as an argument that lexicons are somehow optimized for efficient communication: Words are short because they are frequent or highly informative, implying a causal relationship between word length and frequency or average surprisal. While previous work implies the existence of these causal relationships, they have not explicitly been tested for. In this paper, we apply causal discovery algorithms to lexical data from 12 languages, and find that there is no universal causal relationship between word length, frequency, and average surprisal. Instead, languages vary in the causal structure of their lexicons, suggesting that while Zipf's law describes a universal correlation across languages, that correlation is not the result of a universal causal process.

Word lengths are not optimized for efficient communication

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.

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

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.

Downloads

Next from ACL 2026

What Matters in Tonotactic Learning

Similar lecture

LLM-based Literal Example Generation for Japanese Multiword Expressions

.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 Matters in Tonotactic Learning

Similar lecture

LLM-based Literal Example Generation for Japanese Multiword Expressions

Downloads