Revisiting Few-shot Relation Classification: Evaluation Data and Classification Schemes

We explore few-shot learning (FSL) for relation classification (RC). Focusing on the realistic scenario of FSL, in which a test instance might not belong to any target categories (none-of-the-above, [NOTA]), we first revisit recent popular dataset structure pointing out its unrealistic data distribution. To remedy this, propose novel methodology deriving more from available datasets supervised RC, and apply it TACRED dataset. This yields new challenging benchmark FSL-RC, state art models show poor performance. Next, analyze schemes within embedding-based nearest-neighbor approach with respect constraints they impose embedding space. Triggered by this analysis, scheme NOTA category is represented as learned vectors, shown empirically be an appealing option FSL.