Event Detection and Analysis in Thai News Using Bi-LSTM

Abstract

Event extraction is one of the most challenging tasks in information extraction. Extracted knowledge not only supports semantic search and natural language understanding but also serves as foundational representations in many applications. However, there are few event extraction researches on Thai datasets in Thailand and these prior works are based on classical machine learning techniques. Therefore, this paper presents a research of event detection and event analysis in Thai news using a bidirectional long short-term memory (Bi-LSTM) model. Within the model, event types are classified and event components, i.e., trigger types and argument roles, are identified. The results show that the model could achieve the accuracy of 73.41% in event type classification and 81.71% in event component identification, respectively. It outperforms many classical machine learning techniques on Thai datasets and comparable to other works in other languages.