Research On Aspect Based Sentiment Classification Algorithm Based On Attention Mechanism

Aspect based sentiment classification aims at identifying sentiment polarity expressed in a sentence for a given aspect.Deep learning methods based on the attention mechanism have currently become one of the primary approaches to fulfilling such a task.In this paper,in-depth analysis was made for design thoughts and modeling mechanism of the existing aspect based sentiment classification algorithms that depends on the attention mechanism.Then,it has been noted that these methods have the following defects.First,according to the attention mechanism adopted in the aspect based sentiment classification field,only a memory matrix is constructed,which fails to utilize potential features at levels of words and phrases in a sentence.Second,the existing attention mechanism only takes partial sentence information,but not the overall sentence context,into consideration while calculating an attention weight.Finally,basic assumptions of the position attention mechanism now commonly used are also unreasonable to some extent.Specific to problems mentioned above,two algorithms are proposed in this paper accordingly.Two memory matrices were introduced for one algorithm so that features at word and phrase levels can be utilized.Regarding the other algorithm,two novel attention mechanisms were put forward to solve the following two issues respectively.Main contributions of this study are as follows.1.An aspect based sentiment classification algorithm based on bi-memory attention mechanism was proposed.As constructed depending on RNN EncoderDecoder,its memory module consists of memory matrixes at word and phrase levels.A decoder has multiple computational layers.Each layer,composed of an attention layer and a gated recurrent unit,is applied to capture and integrating essential information relevant to the sentiment polarity of a given aspect.Additionally,a two-stage decoding mechanism was selected for the decoder.Such two stages are respectively responsible for feature extractions from the words-level and the phrase-level memory matrixes,so as to make effective use of features at word and phrase levels.2.An aspect based sentiment classification algorithm based on content attention mechanism was proposed.This algorithm includes two novel attention modeling mechanisms of sentence-level attention mechanism and context attention mechanism.The former is capable of capturing information critical to sentiment polarity of a given aspect from a global perspective of the sentence.Moreover,the overall sentence information is also embedded in output,which improves complex sentence processing capability of the model.The latter takes charge of modeling for word order information and the correlation between the words in a sentence,and then embedding such data to an aspect-specific memory matrix.In this way,these data can be employed by a classifier,which enhances classification performance of the algorithm.