Research And Application Of Multi-label Sentiment Classification Based On UGC

With the rapid development of the Internet,a large amount of user-generated content(UGC)information is accumulated on e-commerce and social networking platforms.Efficient analysis and processing of UGC text data can greatly help enterprises or businesses to conduct product feedback and research,which has great commercial and academic value.At present,research work on fine-grained multi-label sentiment analysis is relatively rare and has more difficulties than traditional sentiment analysis tasks.Combined with the widely used recurrent neural network unit in deep learning,this paper proposes a hierarchical multi-input and output sentiment classification model(HMIO),which both considers the semantic and grammatical information of texts and introduces auxiliary tags to improve the learning ability of network layer weights.In addition,inspired by the Hinge loss function,this paper proposes a relax margin loss function(RMG),which effectively reduces the over-fitting problem of multi-label models.In the feature representation layer,we innovatively propose the word-POS attention mechanism based on part of speech to make full use of the part of speech information in emotional expression.Meanwhile,the word-POS attention mechanism makes it easier to capture the semantic expression of the emotional word combination,which helps the model to converge to a higher level more quickly.The experimental results verify the effectiveness of the HMIO model and the word-POS attention mechanism.This paper improves the C&W model and proposes a sentiment-polarized word embedding(SPWE)model based on the sentiment dictionary.As a weak supervised model,the SPWE can better differentiate the vector representation of sentiment words under the same lexical form,and it is easier to extend to large-scale corpus for word embedding learning.The experimental r-esults show that SPWE can improve the performance of the sentiment classification model.