A Rule-based Methodology and Feature-based Methodology for Effect Relation Extraction in Chinese Unstructured Text

Abstract

The Chinese language differs significantly from English, both in lexical representation and grammatical structure. These differences lead to problems in the Chinese NLP, such as word segmentation and flexible syntactic structure. Many conventional methods and approaches in Natural Language Processing (NLP) based on English text are shown to be ineffective when attending to these language specific problems in late-started Chinese NLP. Relation Extraction is an area under NLP, looking to identify semantic relationships between entities in the text. The term “Effect Relation” is introduced in this research to refer to a specific content type of relationship between two entities, where one entity has a certain “effect” on the other entity. In this research project, a case study on Chinese text from Traditional Chinese Medicine (TCM) journal publications is built, to closely examine the forms of Effect Relation in this text domain. This case study targets the effect of a prescription or herb, in treatment of a disease, symptom or body part. A rule-based methodology is introduced in this thesis. It utilises predetermined rules and templates, derived from the characteristics and pattern observed in the dataset. This methodology achieves the F-score of 0.85 in its Named Entity Recognition (NER) module; 0.79 in its Semantic Relationship Extraction (SRE) module; and the overall performance of 0.46. A second methodology taking a feature-based approach is also introduced in this thesis. It views the RE task as a classification problem and utilises mathematical classification model and features consisting of contextual information and rules. It achieves the F-scores of: 0.73 (NER), 0.88 (SRE) and overall performance of 0.41. The role of functional words in the contemporary Chinese language and in relation to the ERs in this research is explored. Functional words have been found to be effective in detecting the complex structure ER entities as rules in the rule-based methodology.