基于事件语境的文本情感原因对特征提取

Feature extraction of emotion-cause pairs in text based on event context

中文关键词  情感原因对  情感分析  注意力机制  事件语境

英文关键词  emotion-cause pair  sentiment analysis  attention mechanism  event context

中文摘要

      现有的情感原因对提取任务(ECPE)大多采用将情感从句逐一与原因从句匹配的方法, 或专注于候选对的 排序方法, 忽略了影响情感因果关系成立的从句的事件语境, 导致模型在理解情感因果关系时产生偏差, 并且无法 捕捉长距离的因果关系. 为此, 本文提出了基于注意力机制和情感从句卷积核的分层模型, 将原始文档的事件语境 特征嵌入到情感原因对特征提取器中, 以创建一个集成和增强的特征. 首先, 将情感分析得到的情感从句类别特征 作为卷积核. 然后, 利用文档的事件语境特征提取情感原因对. 本文方法在中文数据集的F1分数上有1.38%6.08% 的提升, 在英文数据集的F1分数上有2.35%~7.27%的提升, 说明情感分析和因果事件语境对于情感原因对提取的 有效性.

英文摘要

      For emotion-cause pair extraction (ECPE) task, most of the existing works only match the emotion clause and the corresponding cause clause one by one, or focus on sorting candidate pairs, ignoring the event context among clauses that would impact significantly on the establishment of emotion causality. This leads to the deviation of the model in learning emotion causality. They also fail to capture long-span causality hidden among plenty of clauses. To address this issue, we propose a hierarchical model based on attention mechanism and the convolution kernel of emotion clause. In our method, we derive event context features from the original documents and embed them into emotion-cause pair feature extractors to create an integrated and enhanced feature. Firstly, the category features of emotion clause obtained from sentiment analysis are used as convolution kernel. Then, the event context features of documents are presented to extract emotion-cause pairs. The experimental results show that the F1 score on the Chinese benchmark emotion cause corpus are improved from 1.38% to 6.08% compared with the state-of-the-art approaches. Meanwhile, the F1 score on the English benchmark emotion cause corpus are improved from 2.35% to 7.27% compared with the state-of-the-art approaches. They verify the effectiveness of sentiment analysis and causal event context in ECPE.