Chinese Hedge Scope Detection Based on Structure and Semantic Information

Abstract

Hedge detection aims to distinguish factual and uncertain information, which is important in information extraction. The task of hedge detection contains two subtasks: identifying hedge cues and detecting their linguistic scopes. Hedge scope detection is dependent on syntactic and semantic information. Previous researches usually use lexical and syntactic information and ignore deep semantic information. This paper proposes a novel syntactic and semantic information exploitation method for scope detection. Composite kernel model is employed to capture lexical and syntactic information. Long short-term memory (LSTM) model is adopted to explore semantic information. Furthermore, we exploit a hybrid system to integrate composite kernel and LSTM model into a unified framework. Experiments on the Chinese Biomedical Hedge Information (CBHI) corpus show that composite kernel model could effectively capture lexical and syntactic information, LSTM model could capture deep semantic information and their combination could further improve the performance of hedge scope detection.