United States

Developing specialised language models for low-resource domains typically involves a trade-off between two specialisation strategies: adapting a general-purpose model through continued pretraining or retraining a model from scratch. While adapting preserves the model&#39;s linguistic knowledge, retraining benefits from the flexibility of an in-domain tokeniser -- a potentially significant advantage when handling rare languages. This study investigates the impact of tokenisation, specialisation strategy, and pretraining data availability using classical scholarship -- a multilingual, code-switching and highly domain-specific field -- as a case study. Through extensive experiments, we assess whether domain-specific tokenisation improves model performance, whether character-based models provide a viable alternative to subword-based models, and which specialisation strategy is optimal given the constraints of limited pretraining data. Contrary to prior findings, our results show that in-domain tokenisation does not necessarily enhance performance. Most notably, adaptation consistently outperforms retraining, even with limited data, confirming its efficiency as the preferred strategy for resource-constrained domains. These insights provide valuable guidelines for developing specialised models in fields with limited textual resources.

NAACL 2025

Don&#39;t stop pretraining! Efficiently building specialised language models in resource-constrained settings.

ancient greek

low-resource nlp

language models

### 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)

Developing specialised language models for low-resource domains typically involves a trade-off between two specialisation strategies: adapting a general-purpose model through continued pretraining or retraining a model from scratch. While adapting preserves the model's linguistic knowledge, retraining benefits from the flexibility of an in-domain tokeniser -- a potentially significant advantage when handling rare languages. This study investigates the impact of tokenisation, specialisation strategy, and pretraining data availability using classical scholarship -- a multilingual, code-switching and highly domain-specific field -- as a case study. Through extensive experiments, we assess whether domain-specific tokenisation improves model performance, whether character-based models provide a viable alternative to subword-based models, and which specialisation strategy is optimal given the constraints of limited pretraining data. Contrary to prior findings, our results show that in-domain tokenisation does not necessarily enhance performance. Most notably, adaptation consistently outperforms retraining, even with limited data, confirming its efficiency as the preferred strategy for resource-constrained domains. These insights provide valuable guidelines for developing specialised models in fields with limited textual resources.

Don't stop pretraining! Efficiently building specialised language models in resource-constrained settings.

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

This study investigates the internal representations of verb-particle combinations, called multi-word verbs, within transformer-based large language models (LLMs), specifically examining how these models capture lexical and syntactic properties at different neural network layers. Using the BERT architecture, we analyze the representations of its layers for two different verb-particle constructions: phrasal verbs like "give up" and prepositional verbs like "look at". Our methodology includes training probing classifiers on the model output to classify these categories at both word and sentence levels. The results indicate that the model’s middle layers achieve the highest classification accuracies. To further analyze the nature of these distinctions, we conduct a data separability test using the Generalized Discrimination Value (GDV). While GDV results show weak linear separability between the two verb types, probing classifiers still achieve high accuracy, suggesting that representations of these linguistic categories may be "non-linearly separable". This aligns with previous research indicating that linguistic distinctions in neural networks are not always encoded in a linearly separable manner. These findings computationally support usage-based claims on the representation of verb-particle constructions and highlight the complex interaction between neural network architectures and linguistic structures.

Probing Internal Representations of Multi-Word Verbs in Large Language Models

The growing volume of digitized historical texts requires effective semantic search using text embeddings. However, pre-trained multilingual models face challenges with historical content due to OCR noise and outdated spellings. This study examines multilingual embeddings for cross-lingual semantic search in historical Luxembourgish (LB), a low-resource language. We collect historical Luxembourgish news articles from various periods and use GPT-4o for sentence segmentation and translation, generating 20,000 parallel training sentences per language pair. Additionally, we create a semantic search (Historical LB Bitext Mining) evaluation set and find that existing models perform poorly on cross-lingual search for historical Luxembourgish. Using our historical and additional modern parallel training data, we adapt several multilingual embedding models through contrastive learning or knowledge distillation and increase accuracy significantly for all models. We release our adapted models and historical Luxembourgish-German/French/English bitexts to support further research.

Adapting Multilingual Embedding Models to Historical Luxembourgish

Automatic poetry generation is an immensely complex task, even for the most advanced Large Language Models (LLMs) that requires a profound understanding of intelligence, world and linguistic knowledge, and a touch of creativity.
This paper investigates the use of LLMs in generating Russian syllabo-tonic poetry of various genres and styles. The study explores a character-level tokenization architectures and demonstrates how a language model can be pretrained and finetuned to generate poetry requiring knowledge of a language's phonetics. Additionally, the paper assesses the quality of the generated poetry and the effectiveness of the approach in producing different genres and styles. The study's main contribution is the introduction of two end-to-end architectures for syllabo-tonic Russian poetry: pretrained models, a comparative analysis of the approaches, and poetry evaluation metrics.

Generation of Russian Poetry of Different Genres and Styles Using Neural Networks with Character-Level Tokenization

Algorithmic hate speech detection faces significant challenges due to the diverse definitions and datasets used in research and practice. Social media platforms, legal frameworks, and institutions each apply distinct yet overlapping definitions, complicating classification efforts. This study addresses these challenges by demonstrating that existing datasets and taxonomies can be integrated into a unified model, enhancing prediction performance and reducing reliance on multiple specialized classifiers. The work introduces a universal taxonomy and a hate speech classifier capable of detecting a wide range of definitions within a single framework. Our approach is validated by combining two widely used but differently annotated datasets, showing improved classification performance on an independent test set. This work highlights the potential of dataset and taxonomy integration in advancing hate speech detection, increasing efficiency, and ensuring broader applicability across contexts.

Improving Hate Speech Classification with Cross-Taxonomy Dataset Integration

This paper investigates the application of prompt-answering Large Language Models (LLMs) for the task of Named Entity Recognition (NER) in historical texts. Historical NER presents unique challenges due to language change through time, spelling variation, limited availability of digitized data (and, in particular, labeled data), and errors introduced by Optical Character Recognition (OCR) and Handwritten Text Recognition (HTR) processes. Leveraging the zero-shot capabilities of prompt-answering LLMs, we address these challenges by prompting the model to extract entities such as persons, locations, organizations, and dates from historical documents. We then conduct an extensive error analysis of the model output in order to identify and address potential weaknesses in the entity recognition process. The results show that, while such models display ability for extracting named entities, their overall performance is lackluster. Our analysis reveals that model performance is significantly affected by hallucinations in the model output, as well as by challenges imposed by the evaluation of NER output.

Prompting the Past: Exploring Zero-Shot Learning for Named Entity Recognition in Historical Texts Using Prompt-Answering LLMs

We present a clear distinction between the phenomena of comparisons and similes along with a fine-grained annotation guideline that facilitates the structural annotation and assessment of the two classes, with three major contributions: 1) a publicly available annotated data set of 100 comparative statements; 2) theoretically grounded annotation guidelines for human annotators; and 3) results of machine learning experiments to establish how the--often subtle--distinction between the two phenomena can be automated.

`... like a needle in a haystack'': Annotation and Classification of Comparative Statements

Recent studies suggest that canonical works possess unique textual profiles, often tied to innovation and higher cognitive demands. However, recent work on Danish 19th century literary novels has shown that some non-canonical works shared similar textual qualities with canonical works, underscoring the role of text-extrinsic factors in shaping canonicity. The present study examines the same corpus (more than 800 Danish novels from the Modern Breakthrough era (1870–1900)) to explore socio-economic and institutional factors, as well as demographic features, specifically, book prices, publishers, and the author's nationality – in determining canonical status. We combine expert-based and national definitions of canon to set up a classification experiment to test the predictive power of these external features, and to understand how they relate to that of text-intrinsic features. We show that the canonization process is influenced by external factors – such as publisher and nationality – but that text-intrinsic features nevertheless maintain predictive power in a dynamic interplay of text and context.

Why Novels (Don't) Break Through: Dynamics of Canonicity in the Danish Modern Breakthrough (1870-1900)

This study examines sentiment analysis in Tamil-English code-mixed texts using advanced transformer-based architectures. The unique linguistic challenges, including mixed grammar, orthographic variability, and phonetic inconsistencies, are addressed. Data limitations and annotation gaps are discussed, highlighting the need for larger datasets. The performance of models such as XLM-RoBERTa, mT5, IndicBERT, and RemBERT is evaluated, with insights into their optimization for low-resource, code-mixed environments.

Advancing Sentiment Analysis in Tamil-English Code-Mixed Texts: Challenges and Transformer-Based Solutions

The emergence of artificial intelligence has proven beneficial to numerous organizations, particularly in its various applications for social welfare. One notable application lies in AI-driven image generation tools. These tools produce images based on provided prompts. While this technology holds potential for constructive use, it also carries the risk of being exploited for malicious purposes, such as propagating hate. To address this we propose a novel dataset "HateImgPrompts". We have benchmarked the dataset with the latest models including GPT-3.5, LLAMA 2, etc. The dataset consists of 9467 prompts and the accuracy of the classifier after finetuning of the dataset is around 81%.

Downloads

Next from NAACL 2025

Probing Internal Representations of Multi-Word Verbs in Large Language Models

Similar lecture

Causal Estimation of Tokenisation Bias