基于双向门控机制和层次注意力的方面级情感分析

针对传统情感分析模型将单词或词语作为单一嵌入，而忽略句子之间依存信息和位置信息的问题，提出基于双向门控机制和层次注意力的方面级情感分析模型（Based on Bi-GRU and Hierarchical Attention，BGHA）。首先，将文本数据转成词向量再加入位置编码信息，得到包含位置和语义信息的词向量后通过双向门控机制提取上下文特征；接着，分别在单词注意力层和句子注意力层用注意力机制对特征分配权重，突出重点词和重点句信息；最后，结合给定的方面信息选择性提取与其较匹配的情感特征。在SemEval 2014、SemEval 2016和Twitter短文本评论数据集上的实验结果表示，BGHA模型的准确率对比其他模型都有不同程度的提高，证明了模型的有效性。

ASPECT-BASED-SENTIMENT ANALYSIS BASED ON BI-GATED RECURRENT UNIT AND HIERARCHICAL ATTENTION

For traditional sentiment analysis models, words or expressions are embedded as a single, and the problems of the dependent information and location information among the sentences are ignored, a BGHA sentiment analysis model based on Bi-GRU and hierarchical attention is proposed. Firstly, the text data is transformed into word vector and then the location encoding information is added to get the word vector containing the location and semantic information, and then the context features are extracted by bidirectional gating mechanism. Then, the attention mechanism is used to assign weight to features in the word attention layer and sentence attention layer respectively, highlighting the key words and sentences information. Finally, the emotion features matching the given aspect information are selectively extracted. The experimental results in the SemEval 2014, SemEval 2016 and Twitter short text comment datasets show that the accuracy of the BGHA model has been variously improved compared with other models, which proves the effectiveness of the model.