Extraction Of Entity Hyponymy And Synonymy Relations From Open Domain Texts

The entity hyponymy and synonymy relations are two kinds of core semantic relationships between entities,and they are the important cornerstone of building semantic network,knowledge graph and other large-scale knowledge bases,which have important research significance and application value.The open domain texts have a wide range of sources,different structures,and are rich in a large amount of entity semantic information,which are the important sources for the extraction of entity hyponymy and synonymy relations.However,with the rapid development of Internet technology,the texts on the Web present an exponential growth.The domain boundaries of the texts in the Web become increasingly open and their structures are diverse,which makes it difficult for people to obtain the target entity semantic knowledge intuitively and accurately.Currently,the extraction of entity hyponymy and synonymy relations in open domain texts faces the following problems.1)Entity semantics are diverse,complex and obscure,which makes it difficult to model entity hyponymy and synonymy relation extraction.2)The sparse or missing features of hyponymy and synonymy relations result in unsatisfactory extraction effect and poor robustness for existing methods.This paper studies the extraction of entity hyponymy and synonymy relations from open domain texts based on the background knowledge,relation description and association information of entities.To solve the above problem 1),the paper studies entity hyponymy relation extraction using unsupervised hyponymy semantic clustering,and studies entity synonym extraction based on multi-source of encyclopedia knowledge.To solve the above problem 2),the paper studies entity hyponymy relation extraction based on representation learning and relation description,and studies entity synonym extraction using association information and entity constraint.Specifically,the main research contents of the paper are as follows:(1)Entity hyponymy relation extraction using unsupervised hyponymy semantic clustering.Domain independent unstructured texts are the most common open domain texts.However,these texts come from different sources and are complex and obscure,causing problems such as difficulty in modeling complex texts and relation error propagation.To solve these problems,this paper studies an unsupervised hyponymy semantic clustering for extracting entity hyponymy relations from unstructured texts.First,the semantic iterative pattern-based matching and syntactic methods are employed to obtain high-quality seed entity hyponymy relations.Second,an unsupervised hyponymy semantic clustering method is used to infer novel entity hyponymy relations.Third,a two-step detection strategy is proposed for entity hyponymy relation error detection.In addition,a confidence value is assigned to each hyponymy relation to specify the credibility of the learned hyponymy relations.(2)Research on entity hyponymy relation extraction based on representation learning and relation description.Due to the diversity of open domain texts and sparse features of hyponymy relations,it is difficult to extract semantic features from the open domain texts.In view of this,the paper considers the background knowledge and relation description of entities,and studies the extraction method of entity hyponymy relations based on representation learning and relation description.First,the paper studies the construction method of entity relation description triples,which are rich in entity background knowledge and hyponymy semantic information.Second,based on the entity relation description triples,a joint training model of representation learning is proposed.Third,based on the above joint training model,two learning models named offset-based classification model and offset-based similarity model are presented to identify and predict the entity hyponymy relations from the texts.(3)Research on entity synonym extraction based on multi-source of encyclopedia knowledge.In the encyclopedia texts,there are many entity synonyms.Such synonyms can realize the association and expansion of entities in the hyponymy relations,and can also provide synonymous semantic knowledge support for semantic search,knowledge reasoning and question answering.In this paper,we propose a simple yet effective extraction and cleaning framework that automatically extracts entity synonyms from encyclopedia texts.In the extraction phase,three extraction methods named direct extraction,pattern-based extraction and neural mining extraction are employed to acquire a large number of candidate entity synonym sets from multiple sources of online encyclopedia texts.In the cleaning phase,three error detection strategies named lexical and semantic rules,domain filtering and similarity filtering are used to improve the precision of the extracted candidate entity synonym relations.(4)Entity synonym extraction using association information and entity constraint.The above entity synonym extraction methods are mainly for encyclopedia texts,which are difficult to be used in other open domain texts,and many valuable entity synonyms have not been identified.This paper studies a neural entity synonym extraction method using association information and entity constrain.First,an association-aware set-term neural network classifier is proposed to learn whether a new entity should be added into the entity synonym sets.Second,an entity-constraint-based synonym relation extraction algorithm is employed to apply the above trained set-term neural network classifier to extract the entity synonym relations from the entity term vocabulary.To sum up,this paper takes the background knowledge,relation description and association information of entities to study the extraction methods of entity hyponymy and synonymy relations from open domain texts.The research results of this paper can provide semantic knowledge support for information retrieval,knowledge graph and knowledge reasoning.