民生话题下政务微博评论Emotion-Cause Pair抽取方法研究

Study on Method of Emotion-Cause Pair Extraction of Government Microblog Comments Under the Topic of People's Livelihood

【目的/意义】微博已成为政府部门与公众间互动的一个重要途径,针对政务微博进行细粒度的情感和原因分析有利于提高政府部门舆情治理能力,为此本文提出一套政务微博评论Emotion-Cause Pair抽取架构。【方法/过程】本文在定义Emotion&Cause共现句侦测任务的基础上,基于文本分类模型识别出E&C共现句,构建GATECPE模型抽取Emotion-Cause Pair,并通过模型迁移和微调手段减少数据标注。【结果/结论】经过多个数据集验证,Emotion&Cause共现句侦测阶段识别P值在70%以上,Emotion-Cause Pair抽取阶段识别F1值在60%以上。通过模型微调可以有效缓解模型直接迁移产生的效果下降,本文提出的情感原因抽取流程可以有效抽取出政务微博评论的情感原因。【创新/局限】实验数据来源受限,Emotion&Cause共现句侦测和Emotion-Cause Pair抽取两阶段存在误差传播。

【Purpose/significance】Microblog has become an important way of interaction between government departments and the public.Fine-grained emotion and cause analysis on government microblog is conducive to improving government departments'public opinion governance ability.Therefore,this paper proposes a set of Emotion-Cause Pair extraction architecture for government microblog comments.【Method/process】This paper defines the Emotion&Cause co-occurrence sentence detection task,uses the text classification model to identify the E&C co-occurrence sentence,constructs the GATECPE model to extract the Emotion-Cause Pair,and reduces the data annotation by model migration and fine-tuning.【Result/conclusion】After validation of multiple datasets,the P value maintained above 70%and the F1 value maintained above 60%during Emotion-Cause Pair extraction.Model fine-tuning can effectively alleviate the effect of direct model migration.The emotional reason extraction process proposed in this paper can effectively extract the emotional reason of government microblog comments.【Innovation/limitation】The source of experimental data is limited,and error propagation exists in the two phases of Emotion&Cause co-occurrence sentence detection and Emotion-Cause Pair extraction.