Influence Functions for Sequence Tagging Models

TL;DR

This paper extends influence functions to sequence tagging models, enabling interpretation of predictions by tracing them back to training data segments, and demonstrates practical utility in error identification.

Contribution

It introduces a novel method to compute influence of training segments on sequence tagging predictions, enhancing interpretability of such models.

Findings

Segment influence correlates well with true influence empirically.

Method effectively identifies annotation errors in NER datasets.

Provides an efficient approximation for influence computation.

Abstract

Many language tasks (e.g., Named Entity Recognition, Part-of-Speech tagging, and Semantic Role Labeling) are naturally framed as sequence tagging problems. However, there has been comparatively little work on interpretability methods for sequence tagging models. In this paper, we extend influence functions - which aim to trace predictions back to the training points that informed them - to sequence tagging tasks. We define the influence of a training instance segment as the effect that perturbing the labels within this segment has on a test segment level prediction. We provide an efficient approximation to compute this, and show that it tracks with the true segment influence, measured empirically. We show the practical utility of segment influence by using the method to identify systematic annotation errors in two named entity recognition corpora. Code to reproduce our results is available at https://github.com/successar/Segment_Influence_Functions.