Using Sentence-Level Classifiers for Cross-Domain Sentiment Analysis

Abstract : DRDC has been developing a suite of capabilities built around models of semantics andvisual analytic tools for Applied Research Project (ARP) 15ah. Recently, we implemented asentiment analyser in a document visualization tool called Handles to allow users to examinethe positive and negative opinions associated with concepts. The results were unimpressive.Specifically, the system does poorly classifying document from domains that are differentfrom the training domain. In the work reported here, we consider and explore the twosolutions. First we explore whether a more fine-grained analysis of sentiment where thesentences of a document are used as the functional unit of analysis rather than the wholedocument improves performance. Second, we increased the granularity of the classificationduring training from binary (positive or negative) to trinary (positive, negative, or neutral) tosee if performance improved. Neither solution worked well. However, when we mixeddocuments from different domains together during training, we did find that the performanceimproved. We take the results to suggest that the best way to build a sentiment classifier thatis agnostic with respect to domain is to train the classifier on examples from as many domainsas possible.