Event causality extraction from natural language text

Abstract

Extraction of events has been an important task in the field of Natural Language Processing. However, in the recent years establishing and studying of relationships between events has garnered wide importance. One such relationship between events is the cause-and-effect relationship which is present in major Socio- Political, Biomedical, Legal and Financial events and holds great importance. Hence detailed research on the causality between events has become so significant. Utilizing the Causal News Corpus (CNC) dataset, this thesis delves into the analysis of Socio-Political events worldwide, aiming to extract cause-effect spans to enhance our understanding of event causality. Leveraging the capabilities of Gen-AI (Llama-2), new data was generated and integrated into the CNC dataset, leading to a notable improvement in performance metrics. Specifically, the F1 Score for test data increased from 46% to 56% when employing the BERT model. Furthermore, to broaden the scope of the research, a comprehensive Causal News Corpus Bengali (CNCB) dataset was curated, encompassing news articles covering events from diverse geographical regions. This dataset, spanning over 100 news articles and consisting of 245 cause-effect relationships across 5 primary categories, serves as a valuable resource for studying event causality in Bengali. Additionally, by extracting the causes and effects from the CNCB dataset, a question answering dataset was also created called the CNCB-QA dataset, containing 355 questions in total—187 questions for the cause extraction subtask and 168 questions for the effect extraction subtask, respectively. Remarkable results were achieved through token classification and question answering tasks utilizing BERT as the underlying model, with an F1 score of 82.3% for token classification and 100% similarity score for the question answering task, highlighting the efficacy of the approach in extracting cause-effect spans. Overall, this research contributes to the growing field of event causality extraction, offering insights into the intricate relationships between events. The findings underscore the importance of causality studies in advancing our understanding of global phenomena and inform decision-making processes in various domains.