Sensory Semantic Representation And Sentiment Analysis

In recent years,with the development of cognitive science and artificial intelligence,cross-domain sentiment analysis research has received more and more attention.Emotional expression has always been an important part of human communication.Therefore,it is also important to help computers better understand human emotions in the development of artificial intelligence.At present,the mainstream NLP(Natural Language Processing)sentiment analysis method is to analyze,summarize,and infer text by using those words with subjective emotions.While this paper proposes an innovative analysis method,we use the sensory semantic representations of text for sentiment analysis.Relevant research in cognitive science and cutting-edge psychology has pointed out that the processing of sensory information in the human cerebral cortex can cause emotional changes,that is,sensory feelings can affect people's emotions.Therefore,this paper conjectures that the use of sensory information from the text may help improve the effect of sentiment analysis.Regarding how to obtain sensory information from the text,this work uses sound symbolic words(SSW)from linguistics.These sound symbolic words can provide the sensory information from their phonemes.By verifying the role of text containing sensory features in sentiment analysis,we conclude that in addition to vision and hearing,artificial intelligence can also obtain information in different dimensions from other sensory ways,and thus help its own cognitive development.The main contributions of this work can be summarized in the following three points:First,establish a sensory information data set of Japanese vocabulary: In 4 months,in the laboratory of professor Sakamoto Maki,we completed a data collection experiment of 320 Japanese onomatopoeia in 43 sensory dimensions.Plus,the laboratory belongs to the information department of UEC(University of Electronic Communication)in Japan,Tokyo.After the data was cleaned and organized,we drew the association table of vocabulary phonemes and sensory information,and found the way of obtaining sensory information from text vocabulary.Second,the extraction of sensory information and the textual representation of sensory features: On the basis of the above experimental data,we constructed a sensory information extractor which is suitable for common vocabulary,and we established an innovative word embedding training model based on this information extractor.In this way,we can obtain sensory semantic representation of text and verify the effectiveness of the sensory features for classification.Third,sentiment analysis model based on sensory features: In order to enlarge the contribution of sensory information in text,this work proposes three innovative sentiment analysis models in turn.The first model uses Attention mechanism,the second one uses Multi-Head mechanism and the last one chooses CNN architecture respectively.To prove the effectiveness of sensory semantic representation of text,we comparing the Sense NN model proposed in this paper with the current mainstream text classification model BERT.The result proves the superiority of the sentiment analysis model based on sensory feature.