Description Boosting for Zero-Shot Entity and Relation Classification

Zero-shot entity and relation classification models leverage availableexternal information of unseen classes – e.g., textual descriptions – toannotate input text data. Thanks to the minimum data requirement, Zero-ShotLearning (ZSL) methods have high value in practice, especially in applicationswhere labeled data is scarce. Even though recent research in ZSL hasdemonstrated significant results, our analysis reveals that those methods aresensitive to provided textual descriptions of entities (or relations). Even aminor modification of descriptions can lead to a change in the decisionboundary between entity (or relation) classes. In this paper, we formallydefine the problem of identifying effective descriptions for zero shotinference. We propose a strategy for generating variations of an initialdescription, a heuristic for ranking them and an ensemble method capable ofboosting the predictions of zero-shot models through description enhancement.Empirical results on four different entity and relation classification datasetsshow that our proposed method outperform existing approaches and achieve newSOTA results on these datasets under the ZSL settings. The source code of theproposed solutions and the evaluation framework are open-sourced.