Scaling and Adapting Large Language Models for Portuguese Open Information Extraction: A Comparative Study of Fine-Tuning and LoRA

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F26%3A4232JKTX" target="_blank" >RIV/00216208:11320/26:4232JKTX - isvavai.cz</a>

Výsledek na webu

<a href="http://dx.doi.org/10.1007/978-3-031-79035-5_30" target="_blank" >http://dx.doi.org/10.1007/978-3-031-79035-5_30</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-3-031-79035-5_30" target="_blank" >10.1007/978-3-031-79035-5_30</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Scaling and Adapting Large Language Models for Portuguese Open Information Extraction: A Comparative Study of Fine-Tuning and LoRA

Popis výsledku v původním jazyce

This paper comprehensively investigates the efficacy of different adaptation techniques for Large Language Models (LLMs) in the context of Open Information Extraction (OpenIE) for Portuguese. We compare Full Fine-Tuning (FFT) and Low-Rank Adaptation (LoRA) across a model with 0.5B parameters. Our study evaluates the impact of model size and adaptation method on OpenIE performance, considering precision, recall, and F1 scores, as well as computational efficiency during training and inference phases. We contribute to a high-performing LLM and novel insights into the trade-offs between model scale, adaptation technique, and cross-lingual transferability in the OpenIE task. Our findings reveal significant performance variations across different configurations, with LoRA demonstrating competitive results. We also analyze the linguistic nuances in the Portuguese OpenIE that pose challenges for models primarily trained on English data. This research advances our understanding of LLM adaptation for specialized NLP tasks and provides practical guidelines for deploying these models in resource-constrained and multilingual scenarios. Our work has implications for the broader cross-lingual open information extraction field and contributes to the ongoing discourse on efficient fine-tuning strategies for large pre-trained models. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

Název v anglickém jazyce

Scaling and Adapting Large Language Models for Portuguese Open Information Extraction: A Comparative Study of Fine-Tuning and LoRA

Popis výsledku anglicky

This paper comprehensively investigates the efficacy of different adaptation techniques for Large Language Models (LLMs) in the context of Open Information Extraction (OpenIE) for Portuguese. We compare Full Fine-Tuning (FFT) and Low-Rank Adaptation (LoRA) across a model with 0.5B parameters. Our study evaluates the impact of model size and adaptation method on OpenIE performance, considering precision, recall, and F1 scores, as well as computational efficiency during training and inference phases. We contribute to a high-performing LLM and novel insights into the trade-offs between model scale, adaptation technique, and cross-lingual transferability in the OpenIE task. Our findings reveal significant performance variations across different configurations, with LoRA demonstrating competitive results. We also analyze the linguistic nuances in the Portuguese OpenIE that pose challenges for models primarily trained on English data. This research advances our understanding of LLM adaptation for specialized NLP tasks and provides practical guidelines for deploying these models in resource-constrained and multilingual scenarios. Our work has implications for the broader cross-lingual open information extraction field and contributes to the ongoing discourse on efficient fine-tuning strategies for large pre-trained models. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

—

Rok uplatnění

2025

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Lect. Notes Comput. Sci.

ISBN

978-3-031-79034-8

ISSN

—

e-ISSN

—

Počet stran výsledku

15

Strana od-do

427-441

Název nakladatele

Springer Science and Business Media Deutschland GmbH

Místo vydání

—

Místo konání akce

Belém do Pará

Datum konání akce

1. 1. 2026

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—