基于BiLSTM+Attention的体育领域情感分析研究

Sentiment Analysis Based on BiLSTM+Attention in Sports Field

针对体育领域情感分析资源不足、分析性能不高的现状,对体育领域的情感分析开展了研究.首先从"新浪体育"和"直播吧"等平台经过人工筛选、标注,构建了中文情感标注语料库CH-SPORT,共标记评论10 000条,其中积极评论5 000条,消极评论5 000条.然后选用了SVM、TextCNN、BiLSTM、RCNN、fastText、BiLSTM+Attention等模型对CH-SPORT进行了评估.实验结果表明,BiLSTM+Attention模型在CH-SPORT上的分类效果最佳,Acc为87.75%,比基准数据集ChnSentiCorp和NLPCC2014分别高出18.65%、11.75%.本文构建的数据集能有效应用于体育情感分析研究中.

At present,the research on sentiment analysis in sports field and publicly available corpus resources are very rare and low effect of performance.Therefore,this paper construct an emotional annotation corpus for sports.Firstly,the corpus resources come from "Sina Sports Network" and "Live Bar".After data preprocessing,the emotional polarity annotation is manually performed and then different algorithms and external dataset are used for analysis and comparison.A total of 10 000 comments were marked,including 5 000 positive comments and5 000 negative comments.Then,SVM,TextCNN,BiLSTM,RCNN,fast Text,BiLSTM+Attention were selected to evaluate CH-SPORT.The experimental results show that BiLSTM+Attention model on CH-SPORT is the best,the accuracy can reach 87.75%,18.65% and 11.75% higher than the benchmark datasets ChnSentiCorp and NLPCC2014,respectively,the B:LSTM+Attention model can effectively improve the classification effect.The corpus constructed in this paper can be effectively used in sports sentiment analysis.