基于情感增强词向量的朝鲜语情感分析方法

Sentiment enhanced word vector based Korean sentiment analysis method

现有的句子级情感分析方法把重点放在针对句子的语义及情感倾向建模上,忽略了词向量的情感倾向性信息。针对这一问题,提出一种情感分析方法 SE-LSTM,并将该方法应用于朝鲜语情感分析任务中。以句子片段的情感评分作为目标,训练多层神经网络,得到情感增强词向量,将之作为LSTM网络的输入,预测句子的情感分类。实验结果表明,与传统的LSTM或CNN模型相比,融合了情感增强词向量的LSTM模型将F1值分别提高了2.55个百分点和1.94个百分点。

Existing sentence-level sentiment analysis methods focus on modeling semantic and sentiment polarity of sentence.However,sentiment polarity of word is ignored.To solve this problem,a sentiment analysis method named SE-LSTM was proposed and applied to Korean sentiment analysis task.Multi-layer neural network for obtaining sentiment score of sentence segment was trained to get sentiment enhanced word vector,and the vector from previous step was inputted to the LSTM which predicted sentiment classification of sentence.Experimental results show that compared with the traditional LSTM or CNN model,the LSTM model combined with the sentiment enhanced word vector increases the F1 score by 2.55 percentage points and 1.94 percentage points respectively.