Exposing Shallow Heuristics of Relation Extraction Models with Challenge Data

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3702–3710, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3702 Exposing Shallow Heuristics of Relation Extraction Models with Challenge Data Shachar Rosenman1 Alon Jacovi1 Yoav Goldberg1,2 1Computer Science Department, Bar Ilan University 2Allen Institute for Artificial Intelligence {shacharosn, alonjacovi, yoav.goldberg}@gmail.com Abstract The process of collecting and annotating train- ing data may introduce distribution artifacts which may limit the ability of models to learn correct generalization behavior. We identify failure modes of SOTA relation extraction (RE) models trained on TACRED, which we attribute to limitations in the data annotation process. We collect and annotate a challenge- set we call Challenging RE (CRE), based on naturally occurring corpus examples, to bench- mark this behavior. Our experiments with four state-of-the-art RE models show that they have indeed adopted shallow heuristics that do not generalize to the challenge-set data. Further, we find that alternative question an- swering modeling performs significantly bet- ter than the SOTA models on the challenge-set, despite worse overall TACRED performance. By adding some of the challenge data as train- ing examples, the performance of the model improves. Finally, we provide concrete sug- gestion on how to improve RE data collection to alleviate this behavior. 1 Introduction In the relation extraction (RE) task, our goal is, given a set of sentences s ∈ S to extract tuples (s, e1, e2, r) where a relation r ∈ R holds between e1 and e2 (entities that appear in the sentence s, each represented as a span over s). RE is often represented as relation classification (RC): given a triplet of (s, e1, e2), determine which relation r ∈ R holds between e1 and e2 in s, or indicate no- relation (∅). This can be presented as a set of |R| binary decision problems, (s, e1, e2, r) �→ {0, 1}: return 1 for tuples for which the relation holds, and 0 otherwise. The reduction between RE and RC is clear: given a sentence, extract all entity pair candidates (given a NER system), and run the RC problem on each of them. Indeed, contemporary methods are all RC meth- ods, and the popular TACRED large-scale relation extraction dataset is annotated for RC: each in- stance in the dataset is a triplet of (s, e1, e2) and is associated with a label r ∈ R ∪ {∅}. Import- antly, the annotation is non exhaustive: not all e1, e2 pairs in the dataset are annotated (only 17.2% of the entity pairs whose type match a TACRED relation are). While this saves a lot of annotation effort, as we show this also leads to sub-optimal behavior of the trained models, and hinders our ability to properly assess their real-world utility. We show that state of the art models trained on TACRED are often “right for the wrong reasons” (McCoy et al., 2019): instead of learning to per- form the intended task, they rely on shallow heur- istics which are effective for solving many data- …

Content truncated to 500 words