Research On Question Answering Style Sentiment Analysis Task Based On Text Segmentation

In the Internet age,websites are flooded with massive text information.Question-answering(QA)style sentiment analysis aims to mine the attitudes from user interaction QA pairs such as e-commerce reviews,Weibo updates and Zhihu questions.Existing method of QA-style sentiment analysis divides the question and answer text into different sentences,and encodes each sentence separately.This method ignores the relation between sentences,resulting in information loss.To address the issue,a Hierarchical Recurrent Attention(HRA)model is proposed.After text segmentation,RNN(Recurrent Neural Network)is used to encode the internal features of sentences,and self-attention mechanism is used to encode the features between sentences.According to the importance of each sentence,incorporate the inter-sentence features into each word within the sentence.For the information matching problem of QA pairs,bidirectional attention mechanism is adopted to learn the interactive features of question and answer.BERT(Bidirectional Encoder Representation from Transformers)has achieved top results in many NLP(natural language processing)tasks,but has restrictions on the input length of the model,which leads to truncating sentences when processing long text.To solve this problem,a Hierarchical BERT Attention(HBA)model based on text segmentation is proposed.Firstly,the long text is divided into several short sentences that BERT can process,and BERT is used to extract the complete features of each short sentence.Then self-attention mechanism is used to merge the features of all short sentences,so that the overall features after encoding can represent the whole long text,and the information of each part of the long text can be effectively utilized.To verify the effectiveness of HRA and HBA model,experiments were conducted on three e-commerce QA sentiment analysis datasets.It can be seen from the results that HRA can effectively solve the problem of information loss between sentences in existing methods,and it's faster than RNN without text segmentation.HBA can still achieve very good results when limiting the input length of BERT.