Multiple Instance Learning Networks for Fine-Grained Sentiment Analysis

摘要

We consider the task of fine-grained sentiment analysis from the perspectiveof multiple instance learning (MIL). Our neural model is trained on documentsentiment labels, and learns to predict the sentiment of text segments, i.e.sentences or elementary discourse units (EDUs), without segment-levelsupervision. We introduce an attention-based polarity scoring method foridentifying positive and negative text snippets and a new dataset which we callSPOT (as shorthand for Segment-level POlariTy annotations) for evaluatingMIL-style sentiment models like ours. Experimental results demonstrate superiorperformance against multiple baselines, whereas a judgement elicitation studyshows that EDU-level opinion extraction produces more informative summariesthan sentence-based alternatives.