Joint Linking Of Entity And Relation For Question Answering Over Knowledge Base

Entity linking and relation linking are two crucial components in many question answering systems over knowledge base,which aims to identify relevant entity or relation mention in question and link them to the target entity or relation in knowledge graph.Prior works typically solve these two tasks independently or as sequential tasks,which usually lead to poor performance,as we are facing the case of short text fragments in most questions,the context information for disambiguation is not sufficient.In this paper,we perform entity linking and relation linking jointly,which involves three tasks: entity and relation mention recognition,candidate generation,joint disambiguation for candidate.The key point in our work is to exploit both independent features and joint features of candidates for disambiguation,which captures different characteristics of entity and relation words while considering the information of knowledge graph and semantics of question.The main research contents of this thesis are as follows:(1)The joint recognition of entity and relation mention was regarded as a sequential labeling problem,and solved by the neural network model.We verify the effects of characterlevel features,decoding of conditional random field and pre-training models by comparing the results of this model with other two baseline models.(2)For construct the mention-entity mapping dictionary,we leverage triples containing entity label attributes,entity redirection information and entity alias information.Then we collect synonyms and derivations of relation words to construct mention-relation mapping dictionary.Finally,we create index of mention-URI pairs to provide candidates for entity and relation mention.(3)The candidate disambiguation task is regarded as a ranking process,which we propose to utilize both independent feature and joint feature to focus on different aspects for disambiguation.We first calculate feature scores based on entity and relation respectively to get independent feature.Then we consider entity and relation jointly by constructing a query graph,and later the joint feature is obtained by calculating the similarity score between this query graph and question.The final score for each candidate is calculated by a linear function of independent feature score and joint feature score.This paper implements an entity and relation joint linking system for question answering over knowledge base.We have evaluated our approach on different standard benchmarks.Our empirical study shows that our approach significantly outperforms existing entity and relation linking systems.Our work provides an effective solution for entity linking and relation linking by using a joint method,which is of great significance for solving the task of question answering over knowledge base.