Domain and task-informed sample selection for cross-domain target-based sentiment analysis

A challenge for target-based sentiment analysis is that most datasets are domain-specific and thus building supervised models for a new target domain requires substantial annotation effort. Domain adaptation for this task has two dimensions: the nature of the targets (e.g., entity types, properties associated with entities, or arbitrary spans) and the opinion words used to describe the sentiment towards the target. We present a data sampling strategy informed by the difference between the target and source domains across these two dimensions (i.e., targets and opinion words) with the goal of selecting a small number of examples that would be hard to learn in the new target domain compared to the source domain, and thus good candidates for annotation. This obtains performance in the 86-100% range compared to the full supervised model using only ⇠4-15% of the full training data.