Research On Chinese Semantic Dependency Analysis

Semantic Dependency Parsing (SDP) integrates dependency structure and semantic information in the sentence, based on dependency grammar, which can present the implicit semantic information of a full sentence. Semantic information is extremely valuable for many applications, such as Information Retrieval, Question & Answering and Machine Translation, etc.This thesis researches the building of corpus and auto-tagging the semantic information for a full sentence.The building of corpus has to revolve semantic scheme problem and the manual annotating. The semantic dependency relation tag set used in this thesis was extracted from HowNet.The building of corpus is semi-automatic: First, use rules to annotate some arcs; then manually marked and modified; when a certain scale corpus is built, use maxent method to label, and then manually modified.Automatically determining the semantic relations for a full sentence is strongly desirable. At present, there is no algorithm for semantic dependency analysis, and the most relevant algorithm is dependency parsing algorithm.In this thesis, Graph-based algorithm is used for automatic semantic dependency analysis. As the scale of the current corpus is small, the number of semantic relations is large, data sparseness is serious, which result in the lower accuracy, unlabeled attachment score reached 79.45%, label attachment score reached 63.93%.Knowledge from dependency parsing as additional features is used to guide the establishment of dependency structures, the predicate semantic frames are used to enhance the accuracy of relationship labeling. After adding information, label attachment score is increased by 2.33%.In order to adapt to different applications and prevent the data sparse, the semantic relationships are generaled. After the general, label attachment score is increased by 2.9%.