Fine-Grained Reviews Sentiment Analysis Based On Aspect Hierarchical Structure

Fine-grained sentiment analysis can help people improve the efficiency of information acquisition and learn about the public attitudes toward various aspects of one thing.Fine-grained product reviews sentiment analysis can assist consumers make consumption decision and help merchants improve products.We study three involved tasks:aspect extraction,aspect-based sentiment analysis and aspect hierarchical structure construction.Sentiment analysis based on aspect hierarchical structure can help consumers locate to aspect nodes they are interested in.Meanwhile,sentiment polarity about the aspect are aggregated under the node,which can help to learn the overall evaluation of the aspect quickly.For aspect extraction,considering the tagging result of a word is mainly affected by its near context,we propose Word-level CNN to capture n-grams information by convolution window.Besides,observing morphology of word can reflect part of speech and affect the tagging result,we propose to join Char-level CNN to learn character constitution of word(morphological information),which can enhance feature representation,namely Two-level CNN(TCNN).For Aspect-based sentiment analysis,existing methods mainly utilize context to construct features for Aspect.However,it can cause classification error when there exist multiple aspects in a sentence,for different aspects may have similar features.Therefore,we propose ATtention-based CNN(AT-CNN),which can automatically learn the weight of words in context for an aspect sentiment polarity.Then,considering the importance of aspect information,we propose ATtention-based CNN with Aspect Embedding(ATAE-CNN)based on AT-CNN,which couple aspect information as part of input.We adopt semi-supervised learning method and obtain an initial aspect hierarchical structure from CNet.com,for existing unsupervised methods cause low accuracy.As existing aspect vectorization methods can not distinguish various aspects which have similar context,it can cause error when cluster aspects.We propose attention-based aspect vectorization method which can weight context.then use aspect clustering result to improve the initial hierarchical structure and aggregate sentiment polarity in reviews toward the related aspect.To prove the effectiveness of our methods,we conduct comparative experiments on two datasets provided by SemEval to evaluate our methods.Experimental results show that our methods obtain higher accuracy and F score compared to existing methods,which demonstrate the effectiveness of our proposed methods.