Multi-Perspective Sentiment Analysis on Life Events with Sentiment Cause Identification

As social media plays an important role in people's interaction, extracting personal life events shared on social media becomes a hot research topic. However, identifying the sentiment of a life event is rarely discussed. In this work, we present a task of identifying the sentiment polarity of personal life events from social media posts. We construct the first human-annotated dataset called SentiLiveKB. Notably, the events described by a user may be her/his own experiences or the experiences of others. Analyzing sentiment polarity from the perspectives of the subject and the author is a widely-studied research topic. On the other hand, identifying the causes of sentiment polarity about the life event is essential for applications, such as personalized recommendation and memory recall assistance. To this end, we propose a sentiment cause-aware dual-channel graph convolutional network (CauseDCGCN) with a graph augmentation mechanism to perform sentiment analysis on personal life events from multi-perspectives and identify the corresponding sentiment causes of the events. With two channels, the sentiment from the subject's and the author's perspectives can be identified simultaneously. In addition, the augmentation mechanism applied to both channels is beneficial for detecting the relations of sentiment words and the corresponding causes from different perspectives.