United States

The extensive memory footprint of language model (LM) fine-tuning poses a challenge for both researchers and practitioners. LMs use an embedding matrix to represent extensive vocabularies, forming a substantial proportion of the model parameters. While previous work towards memory-efficient fine-tuning has focused on minimizing the number of trainable parameters, reducing the memory footprint of the embedding matrix has yet to be explored. We first demonstrate that a significant proportion of the vocabulary remains unused during fine-tuning. We then propose a simple yet effective approach that leverages this finding to minimize memory usage. We show that our approach provides substantial reductions in memory usage across a wide range of models and tasks. Notably, our approach does not impact downstream task performance, while allowing more efficient use of computational resources.

NAACL 2025

Vocabulary-level Memory Efficiency for Language Model Fine-tuning

hierarchy

encoder-decoder

discourse

pre-training

### 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)

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

Large language model development relies on the pre-train-then-align paradigm, in which the model is typically pre-trained on a large text corpus and undergoes a tuning stage to align the model with human preference or downstream tasks. We investigate the relationship between pre-training and supervised fine-tuning by considering multiple tasks as well as different pre-trained model checkpoints. Our results on 18 datasets and two models suggest that i) although the model benefits significantly through supervised fine-tuning, it may forget previously known domain knowledge and tasks that are not seen during fine-tuning; ii) the model exhibits high sensitivity to evaluation prompts after supervised fine-tuning, but this sensitivity can be alleviated through further pre-training; iii) continual pre-training improves the model in a latent way that manifests after fine-tuning; iv) The model can already solve some tasks after pre-training while fine-tuning most benefits datasets where the model does not show capability during pre-training.

Amuro & Char: Analyzing the Relationship between Pre-Training and Fine-Tuning of Large Language Models

In this work in progress, we explore the use of large language models (LLMs) to develop linguistic tools that benefit Guarani-Jopara, a variant of the indigenous Guarani language influenced by Spanish and widely spoken in Paraguay as well as in Guarani-speaking communities in other regions. Despite its cultural and social importance, Guarani-Jopara has been underrepresented in modern natural language processing (NLP) applications due to the scarcity of digital resources available in this language. To mitigate this gap, we employed the {Google Translate} API to translate the Alpaca dataset into Guarani-Jopara, thereby generating a synthetic corpus. Training was carried out using tools such as {HuggingFace's AutoTrain} and {UnsloothAI's unsloth}, allowing for fine-tuning on several pre-trained models. The preliminary results indicate that the trained model can capture some of the characteristics of Guarani-Jopara, suggesting a considerable margin for improvement. These initial experiments represent another step toward the inclusion of Guarani-Jopara in NLP applications, with future efforts directed to expanding the corpus and optimizing the model for specific tasks such as machine translation or sentiment analysis.

Building a Large Language Model for Guarani-Jopara? Methodology, Challenges, and Preliminary Results

Machine translation is a tool that can help teachers, learners, and users of low-resourced languages. However, there are significant challenges in developing these tools, such as the lack of large-scale parallel corpora and complex morphology. We propose a novel hybrid system that combines LLM and rule-based methods in two distinct stages to translate inflected Ojibwe verbs into English. We use an LLM to automatically annotate dictionary data to build translation templates. Then, our rule-based module performs translation using inflection and slot-filling processes built on top of an FST-based analyzer. We test the system with a set of automated tests. Thanks to the ahead-of-time nature of the template-building process and the light-weight rule-based translation module, the end-to-end translation process has an average translation speed of 70 milliseconds per word. The approach has the potential to be extended to other low-resource Indigenous languages with dictionary data.

A hybrid Approach to low-resource machine translation for Ojibwe verbs

Quechua is a low-resource language spoken by more than 7 million people in South America. It is a spoken language without writing. Quechua's writing does not follow any grammatical or lexical rules; its current writing is based on how it is uttered and how the sound is written. Quechua is a family of languages with similarities among the seven variants. The lack of a parallel dataset has reduced the opportunities for developing machine translation. We investigated whether increasing the current Quechua Parallel dataset with synthetic sentences and using a pre-trained large language model improves the performance of a Quechua machine translation. A Large language model has been used to generate synthetic sentences to extend the current parallel dataset. We use the mt5 model to fine-tune it to develop a machine translation for Quechua to Spanish and vice versa. Our survey identified the gaps in the state of the art of Quechua machine translation, and our BLEU/Chrf++ results show an improvement over the state of the art.

UCSP Submission to the AmericasNLP 2025 Shared Task

Language is a cornerstone of cultural identity, yet globalization and the dominance of major languages have placed nearly 3,000 languages at risk of extinction. Existing AI-driven translation models prioritize efficiency but often fail to capture cultural nuances, idiomatic expressions, and historical significance, leading to translations that marginalize linguistic diversity. To address these challenges, we propose a multi-agent AI framework designed for culturally adaptive translation in underserved language communities. Our approach leverages specialized agents for translation, interpretation, content synthesis, and bias evaluation, ensuring that linguistic accuracy and cultural relevance are preserved. Using CrewAI and LangChain, our system enhances contextual fidelity while mitigating biases through external validation. Comparative analysis shows that our framework outperforms GPT-4o, producing contextually rich and culturally embedded translations—a critical advancement for Indigenous, regional, and low-resource languages. This research underscores the potential of multi-agent AI in fostering equitable, sustainable, and culturally sensitive NLP technologies, aligning with the AI Governance, Cultural NLP, and Sustainable NLP pillars of Language Models for Underserved Communities. Our full experimental codebase is publicly avail able at: github.com/ciol-researchlab/Context-Aware_Translation_MAS.

Preserving Cultural Identity with Context-Aware Translation Through Multi-Agent AI Systems

This study investigates zero-shot and few-shot cross-lingual transfer effects in Part-of-Speech (POS) tagging and Named Entity Recognition (NER) for Hamshentsnag, an endangered Western Armenian dialect. We examine how different source languages, Western Armenian (contact cognate), Eastern Armenian (ancestral cognate), Turkish (substrate or contact-induced), and English (non-cognate), affect the task performance using multilingual BERT and BERTurk. Results show that cognate varieties improved POS tagging by 8% F1, while the substrate source enhanced NER by 15% F1. BERTurk outperformed mBERT on NER but not on POS. We attribute this to task-specific advantages of different source languages. We also used script conversion and phonetic alignment with the target for non-Latin scripts, which alleviated transfer.

Cognate and Contact-Induced Transfer Learning for Hamshentsnag: A Low-Resource and Endangered Language

Language models have integrated themselves into many aspects of digital life, shaping everything from social media to translation. This paper investigates how large language models (LLMs) respond to LGBTQ+ slang and heteronormative language. Through two experiments, the study assesses the emotional content and the impact of queer slang on responses from models including GPT-3.5, GPT-4o, Llama2, Llama3, Gemma and Mistral. The findings reveal that heteronormative prompts can trigger safety mechanisms, leading to neutral or corrective responses, while LGBTQ+ slang elicits more negative emotions. These insights punctuate the need to provide equitable outcomes for minority slangs and argots, in addition to eliminating explicit bigotry from language models.

Guardrails, not Guidance: Understanding Responses to LGBTQ+ Language in Large Language Models

Given the widespread use of LLM-powered conversational agents such as ChatGPT, analyzing the ways people interact with them could provide valuable insights into human behavior. Prior work has shown that sometimes the use cases of these agents are sexual. For example, they are used to obtain sexual information and advice, as romantic or sexual companions, and to generate erotica and pornography. In this paper, we hone in on this and perform an analysis of dehumanizing tendencies toward LGBTQ+ individuals by human users in their sexual interactions with ChatGPT. Through a series of experiments that model various concept vectors associated with distinct shades of dehumanization, we find that humans prompt ChatGPT to generate LGBTQ+-related content in nuanced ways. Some of our findings signal the reproduction of harmful stereotypes. However, others indicate that not all prompting practices are inherently negative.

Dehumanization of LGBTQ+ Groups in Sexual Interactions with ChatGPT

We study language ideologies in text produced by LLMs through a case study on English gendered language reform (related to role nouns like congressperson/-woman/-man, and singular they). First, we find political bias: when asked to use language that is "correct" or "natural", LLMs use language most similarly to when asked to align with conservative (vs. progressive) values. This shows how LLMs' metalinguistic preferences can implicitly communicate the language ideologies of a particular political group, even in seemingly non-political contexts. Second, we find LLMs exhibit internal inconsistency: LLMs use gender-neutral variants more often when more explicit metalinguistic context is provided. This shows how the language ideologies expressed in text produced by LLMs can vary, which may be unexpected to users. We discuss the broader implications of these findings for value alignment.

Do language models practice what they preach? Examining language ideologies about gendered language reform encoded in LLMs

This study presents BiasAdapt, a novel data augmentation strategy designed to enhance the robustness of automatic media bias detection models. Leveraging the BABE dataset, BiasAdapt uses a generative language model to identify bias-indicative keywords and replace them with alternatives from opposing categories, thus creating adversarial examples that preserve the original bias labels. The contributions of this work are twofold: it proposes a scalable method for augmenting bias datasets with adversarial examples while preserving labels, and it publicly releases an augmented adversarial media bias dataset.
Training on BiasAdapt reduces the reliance on spurious cues in four of the six evaluated media bias categories.

Downloads

Next from NAACL 2025

Amuro & Char: Analyzing the Relationship between Pre-Training and Fine-Tuning of Large Language Models

Similar lecture

Forward Knows Efficient Backward Path: Saliency-Guided Memory-Efficient Fine-tuning of Large Language Models

.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 NAACL 2025

Amuro & Char: Analyzing the Relationship between Pre-Training and Fine-Tuning of Large Language Models

Similar lecture

Forward Knows Efficient Backward Path: Saliency-Guided Memory-Efficient Fine-tuning of Large Language Models

Downloads