Event detection for biomedical text

• Main Author: Pham, Nguyen Minh Thu
• Other Authors: Hui Siu Cheung
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Document and text processing
• Online Access: https://hdl.handle.net/10356/156520
• Institution: Nanyang Technological University

In the last decade, text mining in biomedical domain has received significant attention in research and many studies have been devoted to advancing the state-of-the-art natural language processing (NLP) techniques to biomedical text. Event detection is the primary step in the event extraction task, whose objective is to detect events via trigger mentions that signify the occurrence of events with a particular type. Essentially, event detection requires the construction of a semantic relationship between input text representations and the set of predefined event type labels. However, existing methods tend to pay most attention to learn the input text representations and simply use one-hot vectors for event type labels, which overlooks the importance of understanding the type label meaning. In this research, we propose a novel Label-Pivoting Biomedical Event Detection model (LPBED) which is pretrained with PubMedBERT language model and exploits the semantic meaning of the type label set. More specifically, our proposed model makes use of the underlying semantic meaning of type labels to pivot event types as clues for detecting trigger candidates. Our model gains significant benefits from the pretrained PubMedBERT model for the domain-specific knowledge of the widely-used biomedical data sources. We conduct experiments based on the benchmark GENIA Event 2011 (GE11) dataset. Without using any external knowledge bases and syntactic tools, the experimental results show that our model is robust in performance under the scenarios of limited data availability. In addition, our proposed LPBED model also outperforms the baseline BERT-CRF model used for the MAVEN dataset in general domain. It demonstrates that our proposed model achieves competitive performance for event detection in biomedical text, which provides the potential for further investigation on the event extraction task.