Identificação de tipos de relações temporais event-time em português: uma abordagem baseada em regras com classificação associativa.

Abstract

Temporal understanding in natural language plays a fundamental role in communication effectiveness, enabling the comprehension of sequence of event and their order. This study aims to develop a computational method for identifying types of temporal relations between event and temporal expression in Portuguese texts. The adopted approach is rule-based and incorporates linguistic elements, including lexical, morphosyntactic, and contextual information, Reichenbach’s tenses, temporal signals, and prior world knowledge, in addition to TimeML annotations from the TimeBankPT corpus. The method consisted in creating a comprehensive set of relevant features used to construct rule sets. We explore rule learning algorithms such as CBA, CN2, IDS, RIPPER, and manual rules. The rule sets were applied individually, as well as in combination, to pairs composed of an event and a temporal expression, using two application strategies: the first triggered rule and a voting system. It is worth noting that this is the first work we are aware of to employ rule-learning techniques to solve this specific task. The statistical results showed the effectiveness of the rule-based approach, with the rule set generated by the RIPPER algorithm standing out and achieving the best performance. This rule set outperformed the baseline method, achieving an accuracy of 69.2% and an F1-score of 66.1%. There was a significant increase of 2.3 percentage points in accuracy and 3.6 percentage points in F1-score on the test data. The application of rules by the voting system was more effective on unseen data. The significant difference between the rule sets and the baseline used highlights the importance of the additional features employed by the rules in identifying temporal relations. These features provided complementary information and allowed for a more precise analysis of the data. Furthermore, the rule sets demonstrated generalization ability, capturing patterns and regularities in the data that can be applied to new instances, enabling accurate predictions. This underscores the utility and effectiveness of rule sets as a robust approach to dealing with the complexity of temporal relations in texts. This research contributes to the advancement of natural language processing, providing an enhanced and explainable understanding of temporal relations. It also has practical applications in areas such as scene description, story comprehension, document summarization, representation of temporal structure in medical records, and news analysis. The continuation of this work can unveil new possibilities for temporal understanding in texts.