NoNE Found: Explaining the Output of Sequence-to-Sequence Models When No Named Entity Is Recognized

Jane Arleth dela Cruz, Iris Hendrickx, Martha A. Larson

Abstract: We introduce a feature-relevance approach for seq2seq Named Entity Recognition (NER) models. Our approach provides an explanation when no named entity (NoNE) is detected in the input sequence. Existing work on explainability for NER focuses on explaining positive cases. In contrast, we point out that the problem of false negatives is critical in certain applications domains. Specifically, our approach is motivated by the needs of Disaster Risk Management (DRM). When DRM practitioners use NER to extract the location of disasters from social media, in this case, tweets, they must be certain that the NER model did not miss a tweet that contained a location. Our NoNE explanation approach makes it possible to identify a set of tweets in which a location was potentially missed and to carry out an efficient manual review of the set by looking at the explanation words our approach identifies. We focus on sequence-to-sequence (seq2seq) NER, due to the effectiveness and extendability of these models. Our NoNE explanation approach introduces a NoNE-related class-of-input (COIN) into a Translation between Augmented Natural Languages (TANL) seq2seq NER model. We report the results of experiments that demonstrate that our NoNE explanation approach is able to deliver important information about the false negatives produced by the NER model and also to uncover gaps in the formulation and application of the protocol used to annotate the data.