Research And Implementation Of Aspect-level Sentiment Classification Network Based On Part-of-speech Awareness

The exponentially growing Internet online experience data brings huge challenges to traditional data analysis methods.The sentiment analysis algorithm based on deep learning effectively saves resources while also helping users quickly and accurately obtain other user opinions from big data.The aspect-level sentiment analysis breaks through the limitations of coarse-grained text-level and sentence-level analysis that are too ideal,and makes the analysis process more human-like.The analysis result includes the sentiment tendency of all aspect terms in an opinion text,which is also missing in the other levels.Therefore,the aspect-level sentiment analysis can maximize the effective sentiment information in the text,which is an important branch of the current sentiment analysis research.This thesis focuses on aspect-level sentiment analysis task,and first lists four technical challenges in this field based on deep learning,namely:How to make the model distinguish according to the degree of contribution of words to the emotional expression of the text;How to model the semantic relationship between input text and aspect words;How to deal with the problem of "aspect term sensitivity" of some specific words;How to alleviate the limitations of annotated data sets.The limitations of annotated data sets is a technical barrier commonly faced by the current deep learning field.This problem is mainly reflected in the shortage of sample number and the deviation of sample quality.Under such a background,I propose some methods based on part-of-speech knowledge,and some algorithms based on the interaction of aspect term and other contexts to solve these challengese.This thesis finds that some part-of-speech often implies the expresser's emotional tendency.If it is utilized as prior knowledge into the modeling process,it will enhance the model's ability to judge how much emotional information each word contains,then it is helpful to solve the first and fourth challenges.The asepct-level sentiment analysis methods based on the interaction of aspect term and other contexts aim to filter out irrelevant emotional features,and divide the"emotional radiation area " of each aspect term more clearly.As a result,the second and third technical challenges will also be greatly eased.Finally,they are fused to obtain the aspect-level sentiment analysis networks IPAN based on part-of-speech knowledge.The experimental results show that the final aspect-level sentiment classification networks based on part-of-speech awareness have strong competitiveness compared with more than ten baseline models.Not only that,the rich exploratory experiments in this thesis also proved the effectiveness of the proposed algorithms involved.