基于多跳推理的篇章情感分析模型

A Multi-Hop Reasoning Architecture for Document Level Sentiment Analysis

为进一步提高篇章情感分析准确率,考虑到文档语义的复杂语言学层级结构,结合多跳推理网络将情感分析任务转化为阅读理解任务,提出了一种基于多跳推理的篇章情感分析模型。篇章常常包含情感倾向不同的句子,该结构首先在单词层级使用辅以位置编码的循环网络学习得到句子向量表示,然后在句子层级使用基于注意力机制的循环网络得到当前轮次文档向量,迭代得到最终文档向量,再经过全连接层预测输出。在IMDB和Yelp数据集上的实验结果表示,相较于不考虑文档层级结构的模型和不使用多跳推理的层级模型,所提出的模型具有更好的实验结果。

In order to further improve the accuracy of document level sentiment analysis,considering the complex linguistic hierarchical structure of documents,the document level sentiment analysis tasks can be regarded as machine reading tasks.A sentiment analysis model incorporating both multi-hop reasoning architecture and hierarchical structure is proposed.The documents often contain sentences with different sentiment polarities.So the model first utilizes a recurrent neural network with position encoding at the word level to obtain the sentence vector representation.Then an attention mechanism-based recurrent neural network at the sentence level is employed to leverage different sentences and obtain the document vector representations in the current iteration,followed by a fully-connected network.Empirical results conducted on IMDB and Yelp datasets demonstrate that the proposed model has better performances than the models without document hierarchy or the models without multi-hop reasoning architecture.