基于内联关系的方面级情感分析方法

方面级情感分析 (aspect-based sentiment analysis, ABSA) 旨在预测给定文本中特定目标的情感极性. 研究表明, 利用注意力机制对目标及其上下文进行建模, 可以获得更有效的情感分类特征表达. 然而, 目前常用的方法是通过对特定目标使用平均向量来计算该目标上下文的注意权值, 这类方法无法突出文本中个别单词对于整个句子的重要性. 因此, 提出了一种基于内联关系的方面级情感分析方法, 该方法可以对目标和上下文进行建模, 将关注点放在目标的关键词上, 以学习更有效的上下文表示. 首先使用门控循环单元 (gated recurrent unit, GRU) 对方面信息和句中单词进行融合分布式表达; 然后将分布式表达输入到结合注意力机制的长短时记忆网络 (long short-term memory network, LSTM), 通过查询机制来增加内联关系的权重, 最终得到方面级情感分类. 该模型在公开数据集上进行的实验结果表明, 该方法是有效的, 精确度均超过基线模型.

Inner-relation modelling with memory networks in aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) predicts the sentiment polarity of particular entities in given sentences. Studies show that the most effective methods use features obtained by modelling entities and their contexts with attention to sentiment prediction. However, these methods calculate the attention weights of contexts by using mean vectors for target entities. In addition, these methods cannot highlight the importance of individual words in the text of whole sentences. Therefore, this study proposes an aspect-based sentiment analysis method that can model the inner relations with memory networks, where networks can learn more effective contextual representations. First, gated recurrent units (GRUs) are used to embed the distributed representations of aspect words and words in sentences. Then, a long short-term memory network (LSTM) takes the distributed representation as input to increase the weights of entities according to the attention-based contextual relationship. Finally, through a query mechanism, aspect-based sentiment polarity is obtained. The proposed model is tested on the open-source datasets Semeval-2014 and Semeval-2016, with results showing that the proposed method is effective and the accuracy is higher than that of the baselines.