The Research And Implementation Of Attention-Based Aspect-Level Sentiment Analysis

With the development of e-commerce,review sites,social networking,online users have created a huge number of texts on the web.These texts are commercially valuable,by analyzing the aspect-level sentiment of these texts,users' opinions on products could be determined in a fine-grained manner,which provide decision-making evidences for merchants.Complete aspect-level sentiment analysis consists of two independent subtasks,namely aspect extraction and document-level multi-aspect sentiment analysis.The former aims to extract the objects on which users express their sentiments,and the latter aims to determine the sentiments toward these aspects.This paper proposes two attention-based deep neural networks to achieve unsupervised aspect extraction and weakly supervised document-level multi-aspect sentiment analysis.For aspect extraction,this paper proposes a Linguistic Attention-Based Model(LABM)to achieve both explicit and implicit aspect extraction.LABM first finds the aspect indicators in the text,then utilizes the linguistic attention mechanism to determine the weights of the aspect indicators and obtain the vector representation of the aspect by weighted summing the vectors of the indicators.This paper also proposes an unsupervised training method,Distributed Aspect Learning(DAL).The main idea of DAL is that the word vector of the noun with the most relevance to the aspect indicator should be similar to the vector representation of the aspect.Experimental results on six data sets show that LABM not only exceeds baseline models,but also has strong interpretability.For document-level multi-aspect sentiment analysis,this paper proposes a Diversified Multiple Instance Learning Network(D-MILN),which merely relies on document-level sentiment annotations to learn the aspect-level sentiment classifier.In D-MILN,the document-level sentiment distribution is obtained by combining the aspect-level sentiment distributions through attention.Then,the aspect-level sentiment classifier is learned through back-propagation of document-level supervised signals.However,due to the indirect supervision,the aspect-level sentiment classifier tends to over-fit the document-level signals.To alleviate this problem,this paper further introduces two kinds of diversified regularizations.The diversified textual regularization encourages the classifier to select aspect-relevant snippets,and the diversified sentimental regularization allows the aspect-level sentiments to differ from document-level sentiment.Experimental results on TripAdvisor and BeerAdvocate show that D-MILN outperforms current weakly supervised methods significantly.Based on the research results of aspect extraction and document-level multi-aspect sentiment analysis,this paper designs and implements an aspect-level sentiment analysis prototype system.Given a piece of text input,the system extracts aspects from the text and predicts the sentiments on these aspects.At the same time,the system visualizes the attention mechanism and thus enhances the interpretability of the models.