Research On Causal Relationship Extraction Based On Deep Learning Method

In natural language processing(NLP),causal identification is an important task of information extraction and knowledge discovery.Causal relationship is widely used in the fields of question and answer,decision making and knowledge discovery.People can extract causal relationship from multiple data sources,such as web pages,online journals and electronic medical records,establish corresponding causal chain and provide references for relevant researches.Due to the ambiguity and diversity of natural language texts,causal relationship extraction is still a difficult NLP problem to solve.Traditional causal relationship extraction methods use natural language processing tools and machine learning related classifiers for causal relationship extraction,but they rely heavily on natural language processing tagging such as part of speech tagging and syntactic parsing to provide classification features,and also rely heavily on knowledge base.However,natural language processing labeling tools often have a large number of errors,which will constantly propagate and amplify in the relational extraction system and eventually affect the effect of relational extraction.In recent years,deep learning has been more and more widely used in natural language processing.Since convolutional neural network and cyclic neural network can extract global and local features from sentences,they have achieved good results in relational extraction,machine translation,sentence classification and other tasks.Therefore,this paper mainly use convolutional neural network and Gate Recurrent Unit,an improved version of cyclic neural network,to extract causal relationship.The main contents of this paper are as follows:1.A new word vector model: ELMO,was used to pre-train word vectors as input to the neural network,and the model of multi-attention mechanism based on convolutional neural network(MUL-PT-CNN)and the model of entity feature perception multi-attention mechanism based on bi-directional GRU(MUL-ET-CNN)were proposed for causal relationship extraction.2.When training the neural network,the loss function used is not the traditional cross entropy loss function,but the space-based sorting loss function,which increases the score of the correct classification in the sample and decreases the score of the classification error with the largest score.3.In this paper,causal relationship extraction was conducted by using causal relationship data and non-causal relationship data of semeval-2010-task8.Due to the small amount of data in this causal relationship data set,Altlex data set was also used in this paper,and its form was transformed into the form of semeval-2010-task8 data set.The experimental results show that the proposed model is effective in causal relationship extraction.