If you've got it, flaunt it: Making the most of fine-grained sentiment annotations

Fine-grained sentiment analysis attempts to extract sentiment holders,\ntargets and polar expressions and resolve the relationship between them, but\nprogress has been hampered by the difficulty of annotation. Targeted sentiment\nanalysis, on the other hand, is a more narrow task, focusing on extracting\nsentiment targets and classifying their polarity.In this paper, we explore\nwhether incorporating holder and expression information can improve target\nextraction and classification and perform experiments on eight English\ndatasets. We conclude that jointly predicting target and polarity BIO labels\nimproves target extraction, and that augmenting the input text with gold\nexpressions generally improves targeted polarity classification. This\nhighlights the potential importance of annotating expressions for fine-grained\nsentiment datasets. At the same time, our results show that performance of\ncurrent models for predicting polar expressions is poor, hampering the benefit\nof this information in practice.\n