Interactive Attention Mechanism For Aspect-Level Sentiment Classification

In recent years,many social media have used some recommendation algorithms,such as content-based recommendation and product-based recommendation,to attract a lot of traffic for them.The recommendation algorithms are based on the analysis of users’ emotional orientation,so sentiment analysis has its unique research value.Aspect-level sentiment classification is a more fine-grained sentiment classification.And it aims to identify the sentiment for a given aspect based on context.At present,the neural network method based on the attention mechanism is one of the mainstream methods to solve aspect-level sentiment classification,and has achieved great success.However,the existing aspect-level sentiment analysis methods still have some shortcomings as follows.First,the attention mechanism cannot obtain the positional relationship between words in a sentence.Some researchers use position coding to obtain the positional relationship between words in the attention mechanism.However,the position coding used in the the attention mechanism is simple.Second,most of neural network models,which is based on the attention mechanism,tried to get the relationship between aspect and context,but lack the analysis of the semantic relationship at the sentence level.Third,the attention mechanism-based aspect-level sentiment analysis methods lack the capacity to analyse the sentence of complex grammar and structure.We design two aspect-level sentiment analysis models based on the attention mechanism in response to the above-mentioned deficiencies.Based on the aspect-level sentiment analysis model,we implement a system for aspect-level sentiment analysis.The main work of our method is as follows:1.We integrate the improved sinusoidal position coding into the aspect-level sentiment analysis model,which is based on the interactive attention mechanism.In order to strengthen the attention to key context words,this model interactively processes aspect words and context in the attention mechanism,so as to obtain aspect-based key semantic information,obtain context features and aspect word features that are more conducive to classification.At the same time,the attention mechanism has some shortcomings in the acquisition of order information between words,so this model introduces improved sinusoidal position coding to get the positional relationship between the aspect and the context,and integrate the positional relationship between the context words.On this way the attention mechanism with sinusoidal position coding could obtain the order relationship between words.2.We implement an aspect-level sentiment analysis model based on multiple interactive sentence-level content attention mechanisms.In order to analyze the key semantic information within the sentence and aspect words,this model introduces the self-attention mechanism.In order to make up for the difficulty of analyzing sentences with complex structure,we try to introduce external knowledge to improve attention distribution.On the other hand,the context and aspect words should be interactive,so we introduces two methods of interaction between aspect words and context: sentence-level interaction and Attention-over-Attention.With this two methods,the key semantic information of the sentence based on the aspect is obtained through the interaction between the sentence and the aspect word,and the relationship between aspect and context is obtained through the Attention-over-Attention model.3.Based on the algorithm model described in this article,we design and implement a system for aspect-level sentiment analysis of user comments.The system includes functions such as data processing,sentiment analysis,and result display.It can analyze a single aspect-based review text,or analyze batches of aspect-based review texts,and visually display the analyzed results.