United States

With the increasing interest in using large language models (LLMs) for planning in natural language, understanding their behaviors becomes an important research question. This work conducts a systematic investigation of LLMs&#39; ability to perform logical reasoning in natural language. We introduce a controlled dataset of hypothetical and disjunctive syllogisms in propositional and modal logic and use it as the testbed for understanding LLM performance. Our results lead to novel insights in predicting LLM behaviors: in addition to the probability of input, logical forms should be considered as important factors. In addition, we show similarities and discrepancies between the logical reasoning performances of humans and LLMs by collecting and comparing behavioral data from both.

NAACL 2025

Logical forms complement probability in understanding language model (and human) performance

alethic modality

lm performance prediction

logical form

logical reasoning

### 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**. 

ACL 2025

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

With the increasing interest in using large language models (LLMs) for planning in natural language, understanding their behaviors becomes an important research question. This work conducts a systematic investigation of LLMs' ability to perform logical reasoning in natural language. We introduce a controlled dataset of hypothetical and disjunctive syllogisms in propositional and modal logic and use it as the testbed for understanding LLM performance. Our results lead to novel insights in predicting LLM behaviors: in addition to the probability of input, logical forms should be considered as important factors. In addition, we show similarities and discrepancies between the logical reasoning performances of humans and LLMs by collecting and comparing behavioral data from both.

workshop paper

### Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Welcome to the 2025 meeting of the Nations of the Americas Chapter of the Association for Computational Linguistics! I am proud to help organize the first NAACL conference to carry the new name of our organization, one that emphasizes inclusion for all of the Americas. I am also pleased to welcome you to Albuquerque, New Mexico, a state whose unique blend of cultural influences will make for an excellent backdrop for NAACL 2025, especially with this year’s special theme on NLP in a Multicultural World. 
**[Continue reading...](https://drive.google.com/file/d/1jX-qGhqVSZZCIrAnJaz798pflu5Irrdn/view?usp=sharing)**

*- Colin Cherry, Google, NAACL 2025 General Chair* 

[![](https://assets.underline.io/markdown_image/1/image/b087f8a4dc5816d6e1a6514e59c59ac3.png)](https://drive.google.com/file/d/1T96GzPqObXrMln2BMByCSXTSizjTg69P/view?usp=sharing)

You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

To access NAACL 2025 event page you are required to register. Please follow [**this link**](https://2025.naacl.org/registration/registration/) to register. Access will depend on your registration type.

Please register!

Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

What is a neural model with minimum architectural complexity that exhibits reasonable language learning capability? To explore such a simple but sufficient neural language model, we revisit a basic reservoir computing (RC) model, Echo State Network (ESN), a restricted class of simple Recurrent Neural Networks. Our experiments showed that ESN with a large hidden state is comparable or superior to Transformer in grammaticality judgment tasks when trained with about 100M words, suggesting that architectures as complex as that of Transformer may not always be necessary for syntactic learning.

Syntactic Learnability of Echo State Neural Language Models at Scale

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 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.

Humanlikeness of LLMs in Mental Lexicon

Emotional support conversation (ESC) aims to alleviate the emotional distress of individuals through effective conversations. Although large language models (LLMs) have obtained remarkable progress on ESC, most of these studies might not define the diagram from the state model perspective, therefore providing a suboptimal solution for long-term satisfaction. To address such an issue, we leverage the Finite State Machine (FSM) on LLMs, and propose a framework called FiSMiness. Our framework allows a single LLM to bootstrap the planning during ESC, and self-reason the seeker's emotion, support strategy and the final response upon each conversational turn. Substantial experiments on ESconv suggest that FiSMiness outperforms many baselines, including direct inference, fine-tuning, self-refine, chain of thought, and external-assisted methods, even those with many more parameters.

FiSMiness: A Finite State Machine Based Paradigm for Emotional Support Conversations

There is widespread of misinformation on futuristic technology and society. To accurately detect such news, the algorithms require up-to-date knowledge. The Large Language Models excel in the NLP but cannot retrieve the ongoing events or innovations. For example, GPT and it's variants are restricted till the knowledge of 2021. We introduce a new methodology for the identification of fake news pertaining to futuristic technology and society. Leveraging the power of Google Knowledge, we enhance the capabilities of the GPT-3.5 language model, thereby elevating its performance in the detection of misinformation. The proposed framework exhibits superior efficacy compared to established baselines with the accuracy of 81.04%. Moreover, we propose a novel dataset consisting of fake news in three languages English, Telugu and Tenglish of around 21000 from various sources.

F2 (FutureFiction): Detection of Fake News on Futuristic Technology

A simile is a powerful figure of speech that makes a comparison between two different things via shared properties, often using words like "like" or "as" to create vivid imagery, convey emotions, and enhance understanding. However, computational research on similes is limited in low-resource languages like Tamil due to the lack of simile datasets. This work introduces a manually annotated Tamil Simile Dataset (TSD) comprising around 1.5k simile sentences drawn from various sources. Our data annotation guidelines ensure that all the simile sentences are annotated with the three components, namely tenor, vehicle, and context. We benchmark our dataset for simile interpretation and simile generation tasks using chosen pre-trained language models (PLMs) and present the results. Our findings highlight the challenges of simile tasks in Tamil, suggesting areas for further improvement. We believe that TSD will drive progress in computational simile processing for Tamil and other low-resource languages, further advancing simile related tasks in Natural Language Processing.

TSD: Towards Computational Processing of Tamil Similes - A Tamil Simile Dataset

Emotion analysis plays a significant role in understanding human behavior and communication, yet research in Tamil language remains limited. This study focuses on building an emotion classifier for Tamil texts using machine learning (ML) and deep learning (DL), along with creating an emotion-annotated Tamil corpus for Ekman's basic emotions. Our dataset combines publicly available data with re-annotation and translations. Along with traditional ML models we investigated the use of Transfer Learning (TL) with state-of-the-art models, such as BERT and Electra based models. Experiments were conducted on unbalanced and balanced datasets using data augmentation techniques. The results indicate that Multinomial Naive Bayes (MNB) and Support Vector Machine (SVM) performed well with TF-IDF and BoW representations, while among Transfer Learning models, LaBSE achieved the highest accuracy (63% balanced, 69% unbalanced), followed by TamilBERT and IndicBERT.

Towards Effective Emotion Analysis in Low-Resource Tamil Texts

Sentiment analysis in code-mixed languages poses significant challenges due to the complex nature of mixed-language text. This study explores sentiment analysis on Tamil code-mixed text using deep learning models such as Long Short-Term Memory (LSTM), hybrid models like Convolutional Neural Network (CNN) + Gated Recurrent Unit (GRU) and LSTM + GRU, along with meta-models including Logistic Regression, Random Forest, and Decision Tree. The LSTM+GRU hybrid model achieved an accuracy of 0.31, while the CNN+GRU hybrid model reached 0.28. The Random Forest meta-model demonstrated exceptional performance on the development set with an accuracy of 0.99. However, its performance dropped significantly on the test set, achieving an accuracy of 0.1333. The study results emphasize the potential of meta-model-based classification for improving performance in NLP tasks.

Bridging Linguistic Complexity: Sentiment Analysis of Tamil Code-Mixed Text Using Meta-Model

The prevalence of misogynistic content online poses significant challenges to ensuring a safe and inclusive digital space for women. This study presents a pipeline to classify online memes as misogynistic or non misogynistic. The pipeline combines contextual image embeddings generated using the Vision Transformer Encoder (ViTE) model with text embeddings extracted from the memes using ModernBERT. These multimodal embeddings were fused and trained using three advanced types of Kolmogorov Artificial Networks (KAN): PyKAN, FastKAN, and Chebyshev KAN. The models were evaluated based on their F1 scores, demonstrating their effectiveness in addressing this issue. This research marks an important step towards reducing offensive online content, promoting safer and more respectful interactions in the digital world.

Misogynistic Meme Detection in Dravidian Languages Using Kolmogorov Arnold-based Networks

We propose a novel dataset for detecting religious hate speech in the context of elections in Karnataka, with a particular focus on Kannada and Kannada-English code-mixed text. The data was collected during the Karnataka state elections and includes 3,000 labeled samples that reflect various forms of online discourse related to religion. This dataset aims to address the growing concern of religious intolerance and hate speech during election periods, it's a dataset of multilingual, code-mixed language. To evaluate the effectiveness of this dataset, we benchmarked it using the latest state-of-the-art algorithms. We achieved accuracy of 78.61%.

Detection of Religious Hate Speech During Elections in Karnataka

Detecting hate speech on social media is increasingly difficult, particularly in low-resource Dravidian languages such as Tamil, Telugu and Malayalam. Traditional approaches primarily rely on text-based classification, often overlooking the multimodal nature of online communication, where speech plays a pivotal role in spreading hate speech. We propose a multimodal hate speech detection model using a late fusion technique that integrates Wav2Vec 2.0 for speech processing and Muril for text analysis. Our model is evaluated on the DravidianLangTech@NAACL 2025 dataset, which contains speech and text data in Telugu, Tamil, and Malayalam scripts. The dataset is categorized into six classes: Non-Hate, Gender Hate, Political Hate, Religious Hate, Religious Defamation, and Personal Defamation. To address class imbalance, we incorporate class weighting and data augmentation techniques. Experimental results demonstrate that the late fusion approach effectively captures patterns of hate speech that may be missed when analyzing a single modality. This highlights the importance of multimodal strategies in enhancing hate speech detection, particularly for low-resource languages.

Next from NAACL 2025

Syntactic Learnability of Echo State Neural Language Models at Scale

Similar lecture

Logical forms complement probability in understanding language model (and human) performance

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES