Research On Chinese Hypernym Relation Classification

Research on Chinese hypernym relation classification is a key point in knowledge graph and has significance in knowledge base.Our research focuses on hypernym relation classifi-cation.Firstly,we build a concise guideline and construct a Chinese hypo-hypernym dataset containing 17,813 word pairs.Then,we propose a word-pattern based relation classification method to further utilize cocntext information.Finally,based on an up-to-data classification research,we propose a method based on multi-feature integration which use dependency information.Main researches in the thesis are as follows:(1)Guideline of Chinese Hypernym Relation and Data AnnotationBased on previous work,we propose a concise guideline of Chinese hypernym relation.Considering the meanings of words in dictionary and corpus,this guideline aims at illustrat-ing the semantic relationship between hyponyms and hypernyms.Under the instruction of the guideline,we extract positive hypernym word pairs using Cilin and HowNet while neg-ative pairs contain a mixture of meronymy,co-hyponym and others.Finally,we construct a large scale of Chinese hypo-hypernym word pairs.(2)Hypernym Relation Classification Based on Word PatternIn order to capture the context information of a word pair in sentences,we apply the pattern-based method to convert the information into sparse vectors.We propose the con-cept of word pattern,which improves the matching degree of patterns and make the vector more dense.Embedding contains sementic information while word pattern expresses con-text information.According to the complementarity between embedding and word pattern,we integrate pattern-based method with distributional method.The positive effect of word pattern is verified by being applied in the maximum entropy model.The key point is to build a hypernym relation classifier,using both the semantic meaning of word embedding and the context information in word pattern.(3)Hypernym Relation Classification with Multi-feature FusionIn order to make better use of context information,especially the dependency informa-tion,we attempt to encode the dependency path with LSTM.While constructing the hypo-hypernym database,we discover some surface features on hypo-hypernym word pairs.We build a relation classifier with multi-feature fusion including embedding,dependency path and surface features.Experiments show that the multi-featured classifier outperforms other models.In conclusion,the thesis designs a guideline of Chinese hypernym relation and con-structs a Chinese hypo-hypernym database.Word pattern is proposed and combined with embedding to improve the accuarcy of the relational classification task.Also,the use of dependency path and surface features further improve the performance.So far,we have accomplished some primitive progress and we hope this thesis could further motivate the progress of natural language processing and knowledge graph.