Attention Based Short Text Sentiment Classification Approach

With the rapid development of Internet technology,a large number of efficient and convenient social products have applied by people and most industries are closely connected with the Internet,which gradually makes the Internet diversify.Different from traditional tools such as forums and blogs,emerging social tools have the characteristic of immediacy,wide spreading at an alarming rate,which affects and changes sentiment judgments of people.Therefore,sentiment analysis is necessary for its ability to understand the trend of public opinion timely,obtain useful decision-making information,and provide important support for commodity sales,public opinion monitoring and etc.The information broadcast by emerging social tools is brief,which is a typical short text structure,having high degree of colloquialism,lacking of standardized expression and missing semantic features.Thus,how to judge the sentiment tendency of short texts accurately is challenging.This paper adopt Attention mechanism and deep learning model to mine semantic features of short texts,classify the sentiment polarity and recognize sentiment orientation on aspect terms.Moreover,efficient and accurate aspect term extraction is achieved.The main research work are including the following parts:(1)Short text sentiment classification.This part combines Convolutional Neural Network(CNN)and Long Short-Term Memory Neural Network(LSTM)for short text sentiment polarity judgment.At the same time,Self-Attention mechanism and Capsule network are introduced to mine deep semantic features.Finally,based on the diversified network structure,the sentiment classification method is studied for further model optimizing.(2)Aspect term sentiment orientation judgment.This part analyzes the sentiment of different aspects in the text from two angles: the text itself and the interactive information.First,the position weight coding feature is designed to highlight the emotional features which are closer to the aspects.Combined with multiple Attention mechanism,it can realize the aspect-level sentiment classification.Meanwhile,for the interactive learning features,Scaled Dot-Product Attention is used to construct a hierarchical Attention model to integrate the interaction information between the aspects and context for a better performance of sentiment classification.(3)Bert-based model for aspect term extraction.This paper introduces the Bert model,which effectively improves the quality of data preprocessing and pre-training through deep learning.Combined with the Bidirectional Long Short-Term Memory Neural Network(BLSTM)and Conditional Random Field(CRF),the quality ofaspect term extraction is further improved.Based on the Attention mechanism,this paper studies and optimizes the short text sentiment analysis algorithm.Short text sentiment polarity classification and the aspect term sentiment recognition are realized,which can alleviate the problems of text messages being too brief or feature missing effectively and improve the performance of sentiment analysis.Combined with the Bert model,the aspect term extraction becomes more efficient.