Using Interpretation Methods for Model Enhancement

TL;DR

This work introduces UIMER, a unified framework that enhances NLP models by leveraging interpretation methods and gold rationales during training. By adding a rationale-alignment loss L_int to the standard task objective, UIMER can accommodate gradient-, erasure/replace-, and extractor-based interpretations; warm-up and, for extractor-based methods, multi-round training further improve stability and performance. The authors instantiate three approaches: UIMER-Gb (gradient-based), UIMER-Im (erasure/replace-based via Input Marginalization), and UIMER-Dm (extractor-based via DiffMask), demonstrating substantial gains especially in low-resource settings across Intent Classification, Slot Filling, and Natural Language Inference. The two novel instantiations often outperform gradient-based baselines, highlighting the value of integrating external rationales into model optimization. The results suggest practical benefits for data-scarce NLP tasks and provide guidance for selecting interpretation methods in real-world scenarios, with code released for replication.

Abstract

In the age of neural natural language processing, there are plenty of works trying to derive interpretations of neural models. Intuitively, when gold rationales exist during training, one can additionally train the model to match its interpretation with the rationales. However, this intuitive idea has not been fully explored. In this paper, we propose a framework of utilizing interpretation methods and gold rationales to enhance models. Our framework is very general in the sense that it can incorporate various interpretation methods. Previously proposed gradient-based methods can be shown as an instance of our framework. We also propose two novel instances utilizing two other types of interpretation methods, erasure/replace-based and extractor-based methods, for model enhancement. We conduct comprehensive experiments on a variety of tasks. Experimental results show that our framework is effective especially in low-resource settings in enhancing models with various interpretation methods, and our two newly-proposed methods outperform gradient-based methods in most settings. Code is available at https://github.com/Chord-Chen-30/UIMER.

Figures (5)

• Figure 1: Our framework illustration utilizing interpretation methods to enhance models. The dotted green line indicates how the parameters of the model are optimized.
• Figure 2: An example of attribution scores and gold rationales.
• Figure 3: The curves of $L_{int}$ with and without warm-up training on 1-shot Intent Classification over 4 random seeds.
• Figure 4: Ablation study of multi-round training on 10-shot Slot Filling.
• Figure 5: The underline marks a rationale. Replace the rationale by BERT v.s. randomly.