The Research On Text Relation Extraction Based On Convolution Neural Network

With the rapid development and extensive application of information technology,a large amount of text information has been generated in the Internet.How to extract useful information from these textual information is one of the most important topics for domestic and foreign researchers.Relation extraction is an important subtask in information extraction and is mainly used to predict the semantic relations between pairs of nominals contained in a given sentence.At present,many relational extraction systems use existing natural language processing tools to generate features.However,there are many mistakes when generating features,which may lead to errors in relation detection and relationship classification.To solve these problems,researchers began to apply deep learning techniques to relational extraction tasks in recent years.The convolutional neural networks in deep learning technology can be a good solution to those problems that caused by the use of existing natural language processing tools to generate features.However,the current convolutional neural network model still has a problem of low accuracy.Therefore,this paper makes further research and improvement on the current convolutional neural network based on words space characteristics,so as to obtain better relationship extraction performance.Main results of this paper are as follows:(1)Preprocessing.First,remove the non-text parts and stopwords in the data and restoring words to their original forms.Then,by truncating and repeating the long sentences,the sentences are unified to a fixed length.Then,using the word2vec to train the word embedding model.Finally,in order to highlight the lexical features of labeled noun pairs,this paper proposes a method that uses '1' as the placeholder to replace the word embedding vector of labeled nouns.(2)Convolutional neural network model design.The feature vector generated by the word embedding model only considers the semantic information of words,and ignores the influence of the sentence structure on the relation extraction.In order to extract sentence-level features,we need to consider the relationship between each word in the sentence.Based on the spatial relationship between vocabulary feature vectors of words,this paper proposes a method of extracting sentence-level features using the vocabulary space feature vector as an input channel of the convolutional neural network model.(3)Convolutional neural network model optimization.Model optimization is divided into two parts.On the one hand,weights are added for cross-entropy to balance the problems caused by uneven distribution and quantity of various types of data.On the other hand,it adjusts the learning rate according to the training situation during the training process to solve the problem of over-fitting in the process of training.Experimental result shows that the convolutional neural network model proposed in this paper reaches 83.2%of FI value,it proves that the model proposed is effective.