Research On Relationship Extraction Method Based On Piecewise Convolution Neural Network

Natural language processing is an important sub-field of artificial intelligence,and its impact on human life is important and profound.At present,how to extract users' interesting and meaningful content from massive and redundant information is an urgent problem to be solved.The two core tasks of information extraction are named entity recognition and relationship extraction.In recent years,deep learning technology has been widely used in both tasks,and great progress has been made in effect.The relationship extraction task is based on the entity recognition task,and the purpose is to extract the relationship from the entities that are meaningful.This paper uses a pipeline approach to study a named entity recognition model based on a word vector pre-training model,and then uses the identified entities to complete a relationship extraction task with a neural network.The following introduces the work and innovations of this article:1.Completed named entity recognition based on BERT-BiLSTM-CRF model.The vector representation of words is obtained through the BERT pre-training model.This method effectively shortens the training time and has a significant effect on the treatment of polysemy problems.The vector representation sequence is then input to the BiLSTM network layer to obtain the text context features,and finally the sequence features are annotated through the CRF layer to obtain the final entity recognition result.2.A piecewise convolutional neural network is proposed to complete the relation extraction task.The input of the model is a sentence and its entity.The input is converted into a representation vector and sent to the convolution layer.In the convolutional layer,the vector is divided into two segments and different convolution operations are used respectively.The model that integrates sentence and entity information can better extract text features.After concatenating the two convolutional results,they are input to the pooling layer for dimensionality reduction,and at the same time,they can obtain richer corpus structure features.Finally,the classifier classifies and outputs the result of relation extraction.3.Introduce the self-attention mechanism in the relation extraction model,and give the feature self-attention weight.The key features that express the relationships between entities are given higher weight values,and the weights that represent the wrong relationship features are reduced.It enables the model to capture important semantic information,reduces the impact of corpus labeling problems,and optimizes the relationship extraction effect of the model.The text uses distant supervision methods to obtain training data and extract relationships,reducing the dependence on manual annotation.The realized named entity recognition function has made certain progress in recognition accuracy and time.The performance of relation extraction based on the evaluation pipeline is compared with other related models.The experimental results prove that the method proposed in this paper improves the effect of relation extraction.