Research On Relation Extraction In Chinese News Text

With the rapid development of Internet technology,there is a massive growth in Internet text.It becomes a hotspot to extract needed knowledge and information from mass text quickly and accurately,in which entity relation extraction from text is an important issue.Currently,the most existing research works on entity relation exaction focus on English text,and carried out with traditional machine learning methods.Furthermore,the influence of abundant no-relation instances is rarely considered.To this end,this study focuses on entity relation extraction from Chinese news text based on deep learning.To reduce the influence of no-relation instances,the entity extraction task is divided into two subtasks,namely entity relation discrimination and relation classification.In the entity relationship determination subtask,this study first realizes an entity discrimination method based on bag of words model and Logical Regression algorithm.To solve the problem of too many feather space dimensions and long running time existing in this method,the entity discrimination method based on Convolutional Neural Network(CNN)model is studied.The input of CNN are the word vectors obtained from pre-training on Sogou CS news text corpus and the vector reflection of the words from ACE2005 Chinese text entity relation extraction dataset.The experimental results on the ACE 2005 dataset show that the CNN-based method performs well.The achieved F-value is 81.78%.In the entity relation classification subtask,an entity relation classification method based on Bi-directional Long-Short Term Memory model(BLSTM)and feature fusion is proposed.This method firstly learned pre-training word vectors from corpus.The entity related features including the entity type,the entity length,the entity relative position,and so on,are identified.Through analyze the features related to entity types and corresponding contextual features in the text corpus,a set of rules are constructed.By incorporating the word vectors,entity related features and rules as the input of the BLSTM model,an entity relation type classifier is constructed.The experimental results on ACE2005 dataset show that this method achieves 91.74% F-value,which shows the effeteness of the proposed method on entity relation extraction from Chinese news text.