NoNE Found: Explaining the Output of Sequence-to-Sequence Models when No Named Entity is Recognized

Anonymous

Abstract: Sequence-to-sequence (seq2seq) models are known to be effective for named-entity recognition (NER). Here we focus on explainability of seq2seq NER models. Contrary to most efforts that focus on explaining why a certain named entity has been recognized, we concentrate on negative cases i.e., sequence-level true negative or false negative, in which no named entity (NoNE) is recognized. We introduce an approach to feature-relevance explainability for seq2seq models that leverages, a special class-of-input (COIN) token to capture whether or not a named entity was present in the input sequence. We run experiments on a location extraction task using a modified translation model (TANL) and generate NoNE explanation for the sequence-level negatives. We carry out a systematic use case-based validation procedure for our NoNE explanation approach. The experiments demonstrate that our NoNE approach is able to deliver important information about shortcomings of the seq2seq model and to uncover gaps in the formulation and application of the protocol used to annotate the data.

Paper Type: long

Research Area: Interpretability and Analysis of Models for NLP

Contribution Types: Model analysis & interpretability, Publicly available software and/or pre-trained models

Languages Studied: English