Banca de DEFESA: DÁRCIO SANTOS ROCHA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : DÁRCIO SANTOS ROCHA
DATE: 14/08/2023
TIME: 14:00
LOCAL: https://conferenciaweb.rnp.br/ufba/defesas-do-formas
TITLE:

IDENTIFICATION OF EVENT-TIME TEMPORAL RELATIONS IN PORTUGUESE: A RULES-BASED APPROACH WITH ASSOCIATIVE CLASSIFICATION

KEY WORDS:

Information Extraction, Temporal Relations, Natural Language Processing

PAGES: 127
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SUMMARY:

This work aims to develop a computational method to identify types of temporal relations between events and temporal expressions in texts written in Portuguese. In order to achieve this goal, rule learning techniques will be employed to discover the best combinations of available linguistic information, formulating decision rules that can efficiently identify the types of temporal relations between events and temporal expressions. Most related works adopted a machine learning-based approach, while only one used a hybrid approach, combining manual rules. The methodology proposed in this work consists of a rule-based approach, which incorporates lexical, morphosyntactic and contextual information, Reichenbach tenses, temporal signals and knowledge about the world, in addition to TimeML annotations in the corpus. Unlike a purely machine learning approach, the rule sets generated by our method allow the combination of rules generated by different algorithms, or the combination of complete sets, which can result in better performance. In short, the method takes event/time expression pairs as input and uses a filtering strategy to select the pairs most likely to have been annotated in the corpus. It then applies sets of rules to each pair to identify the type of existing temporal relationship and a data augmentation strategy to calculate the temporal closure of all pairs and their respective identified relationships. In preliminary experiments, we proposed an initial set of manual rules for the Portuguese language. However, the results obtained showed that this set was limited, resulting in low coverage and consequently low accuracy. The maximum value achieved was 45.1% accuracy and 34.1% coverage in the test data. To improve these results, we propose to incorporate rule learning techniques to the method, aiming to increase the set of rules. These techniques are able to handle noisy data well, work well on unseen data and generate more efficient rules, in addition to offering competitive performance and working efficiently. With this incorporation, we hope that the proposed experiments will produce a set of rules capable of identifying types of event-time temporal relations efficiently and achieving superior results. This will contribute to the advancement of the state of the art in the area, in addition to disseminating the research carried out and contributing to the scientific community.

COMMITTEE MEMBERS:
Presidente - 2352062 - MARLO VIEIRA DOS SANTOS E SOUZA
Interno - 3069553 - ROBESPIERRE DANTAS DA ROCHA PITA
Externo ao Programa - 3551858 - RERISSON CAVALCANTE DE ARAUJO - UFBA