Representation Mapping And Its Application In Relation Extraction And Knowledge Base Question Answering

With the development of internet technology,the network is full of natural lan-guage text information.How to efficiently and accurately obtain the information that users need from large-scale text has become an urgent problem.Relation extraction and knowledge base question answering,as the core tasks of natural language process,play an important role in solving such problems.Relational extraction can extract structured information from a large number of natural language texts.This information can be used to complete the knowledge base,and the knowledge base can also help to relation extraction.However,the knowledge base itself has a much larger coverage of the entity than that of the training sample,This thesis defines the entities that do not exist in the training samples as unseen enti-ties.Because of the lack of representation of unseen entities,most methods of relation extraction do not handle samples involving these entities well.KBQA generally can be divided into two steps:entity linking and relation de-tection.Compared with entity linking,which only need lexical level matching of the question and the knowledge base entity,the relation detection needs to understand the semantics of the whole sentence,which is more challenging.Due to the wide variety of relations covered by the knowledge base,it is unrealistic to label a sufficient amount of corpus for all relations in the knowledge base to train the relation detection model.Similarly,this thesis defines the relations that are not annotated by training sample as unseen relations.In actual use,the KBQA system does not answer questions that in-volve unseen relations.This phenomenon is particularly serious in the open domain KBQA.This thesis mainly studies the representation of unseen entities and relations.The main works are as follows:1.This thesis proposes a representation mapping method called Adapter.This struc-ture can learn a mapping from a domain-independent general representation to the representation of a specific domain.Through this mapping,the representations of unseen entities and relations can be obtained.2.In order to alleviate the negative impact of the lack of representation of the unseen entity on the relation extraction,the adapter is applied to the entity representation of the relation extraction task.The experiment shows that the adapter brings im-provement to relation extraction.3.To alleviate the problem of unseen relations in KBQA,this thesis use adapter into the state-of-the-art relation detection model.Since SimpleQuestion is not sufficient to reflect the performance of unseen relation detection,this thesis re-organizes the SimpleQuestion dataset and evaluate the performance of seen and unseen relation,separately.Experiments show that the adapter brings a significant improvement to the detection of unseen relations,while still keep comparable to the state-of-the-art method for the seen relation.