Evolving Open Information Extraction for Portuguese employing Language Models

Bruno Souza Cabral

Evolving Open Information Extraction for Portuguese employing Language Models

Open Information Extraction (OpenIE) is an important task in Computer Science focused on extracting structured information from text, typically as (argument 1, relation,
argument 2) triples, without requiring predened target relations. OpenIE aims to extract valuable information for uses such as enhancing language understanding, populating knowledge bases, and text comprehension. The extraction of OpenIE relations from
Portuguese text presents substantial challenges, primarily due to its rich morphology,
frequent use of clitic pronouns, exible word order, inected nature, and other linguistic
peculiarities.
Deep Learning has signicantly advanced OpenIE for the English language, with
sequence labeling being a common approach. Recently, a new approach, Generative Information Extraction, particularly leveraging generative Large Language Models (LLMs),
has emerged as a fruitful alternative. Generative techniques can take a sentence as input
and generate structured semantic representations.
Despite numerous OpenIE studies focusing on English, research on OpenIE for the
Portuguese language, particularly employing Deep Learning methods, remains limited.
Existing work often relies on datasets automatically translated from English. Moreover,
most Deep Learning approaches for OpenIE in Portuguese have adopted a multilingual
perspective, treating it as just one language among many in training datasets, thereby
often neglecting its unique linguistic characteristics.
This thesis investigates neural methods for the automated extraction of OpenIE relations from Portuguese texts. A core contribution is the development and curation of
diverse Portuguese OpenIE datasets to address data scarcity and enable robust evaluation. These include both manually annotated corpora and novel corpora generated using
LLMs. The study involves developing and evaluating a sequence labeling model and
assessing the performance of generative LLMs on these Portuguese datasets.
A comprehensive comparative analysis of these methods is conducted, focusing on
their ecacy in extracting OpenIE relations, including abstractive ones, from Portuguese
text. This research signicantly contributes to the growing body of literature on the
application of Deep Learning techniques for OpenIE in the Portuguese language, addresses
critical resource gaps, and lays the foundation for further advancements in this eld,
particularly in exploring generative and abstractive extraction capabilities.

Daniela Barreiro Claro

Marlo Vieira dos Santos e Souza

Renata Vieira

http://lattes.cnpq.br/6218967777630412

Marcos Garcia

https://citius.gal/team/marcos-garcia-gonzalez/

Aline Marins Paes Carvalho

http://lattes.cnpq.br/0506389215528790

Vládia Célia Monteiro Pinheiro

http://lattes.cnpq.br/2991281565518934

Helena de Medeiros Caseli

http://lattes.cnpq.br/6608582057810385

Pablo Gamallo

https://fegalaz.usc.es/~gamallo/

Tatiane Rios

http://lattes.cnpq.br/0851148137941240

Lais do Nascimento Salvador

http://lattes.cnpq.br/1972531466861737

15 Sep, 2025

8:00 AM

O principal impacto científico deste trabalho é no avanço do estado da arte da área de OpenIE para a língua portuguesa, com o objetivo de impulsionar novos LLMs para extração de informação aberta. Impacto social corresponde a uma melhoria no pipeline de distribuição para diversas tarefas do PT que fazem uso da OpenIE.