Entity Linking Based On Deep Neural Networks

Text understanding plays an important role in Natural Language Processing(NLP)and Artificial Intelligence(AI).Entity Linking(EL)is the key for computer to understand texts.EL is to extract mentions from texts and link the mentions to a unique entity in a specified Knowledge Base(KB).If the mentions are ambiguous,Named Entity Disambiguation(NED)is necessary to help computer understand different semantics in different contexts.In this paper,we propose two different EL methods for Chinese Encyclopedias Websites(CEW)infobox and plain text respectively.Our EL methods include Chinese KB construction from CEW,word segmentation,word embedding,document embedding,synonyms mapping and other NLP approaches which follows the habits of text understanding by human that is from word understanding to sentence understanding.We first propose a NED approach for cells in infobox to enhance the relationships among entities in KB.Then we propose a EL approach based on Deep Neural Networks(DNN)for plain text.Mentions are extracted from the given text at first,and entity candidates for these mentions are collected.Then,we apply a bidirectional LSTM networks to mention and context embedding,a Deep Convolutional Neural Network(DCNN)to entity candidates embedding.So,the similarity between the mention and the entity candidates is measured to rank the candidates.Finally,collective graph model and Doc2Vec are used in joint disambiguation.We test our method on both Chinese datasets and English datasets.We achieve 81.07%micro precision on random cases from CEW hyperlinks,state-of-the-art performance on TAC-KBP datasets and 61.47%macro F1 which is also ranked first among 6 other methods provided by GERBIL.At last we present the application of our research on websites.