United States

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.

ACL 2026

Learning Process Interaction Through Simplex ISL Transducers

function decomposition

process interaction

learning

### Welcome to IJCNLP-AACL 2025! 
 It is a great honor to host this joint conference in Mumbai, India, from December 20 to 24, 2025. The joint conferences of IJCNLP and AACL are organized with alternating leadership in the Asia-Pacific region. The event is run by the Asian Federation of Natural Language Processing (AFNLP) in odd years, and by AACL in even years, while it is organized solely by ACL when the annual ACL meeting is held in the region. This year, the conference is primarily organized by AFNLP. 
*Kentaro Inui
MBZUAI, UAE
General Chair, IJCNLP-AACL 2025* 
Read full message and download the Conference Handbook [**here**](https://drive.google.com/file/d/1UTwxkAqSqI-GAoJC3wE1zZt5VP1Y8GX0/view?usp=sharing).

IJCNLP-AACL 2025

The 14th IJCNLP & 4th AACL will be held in Mumbai, India from December 20th to December 24th, 2025.

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.

**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.

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.

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.

Downloads

Next from ACL 2026

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

Similar lecture

Indian Sign Language Recognition and Translation into Odia

.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

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

Similar lecture

Indian Sign Language Recognition and Translation into Odia

Downloads