Boosting Explainability through Selective Rationalization in Pre-trained Language Models

TL;DR

This work tackles explainability challenges in pre-trained language models by showing that existing selective rationalization methods degenerate when applied to PLMs due to token homogeneity. It introduces PLMR, a generator-predictor framework that splits a PLM into an early-layer generator and a later-layer predictor, with Dim-Reduction layers that prune irrelevant tokens and a regularized prediction objective that uses full-text information. Empirical results on BeerAdvocate and HotelReview across multiple PLMs demonstrate substantial improvements in rationale quality (F1) and prediction fidelity over state-of-the-art baselines, validating the approach's effectiveness in mitigating rationalization degeneration and failure. The findings suggest that reducing token homogeneity and enforcing full-context regularization can significantly enhance explainability in PLMs, with broad implications for interpretable NLP and robust deployment of large models.

Abstract

The widespread application of pre-trained language models (PLMs) in natural language processing (NLP) has led to increasing concerns about their explainability. Selective rationalization is a self-explanatory framework that selects human-intelligible input subsets as rationales for predictions. Recent studies have shown that applying existing rationalization frameworks to PLMs will result in severe degeneration and failure problems, producing sub-optimal or meaningless rationales. Such failures severely damage trust in rationalization methods and constrain the application of rationalization techniques on PLMs. In this paper, we find that the homogeneity of tokens in the sentences produced by PLMs is the primary contributor to these problems. To address these challenges, we propose a method named Pre-trained Language Model's Rationalization (PLMR), which splits PLMs into a generator and a predictor to deal with NLP tasks while providing interpretable rationales. The generator in PLMR also alleviates homogeneity by pruning irrelevant tokens, while the predictor uses full-text information to standardize predictions. Experiments conducted on two widely used datasets across multiple PLMs demonstrate the effectiveness of the proposed method PLMR in addressing the challenge of applying selective rationalization to PLMs. Codes: https://github.com/ylb777/PLMR.

Figures (10)

• Figure 1: The selective rationalization framework RNP.
• Figure 2: Two examples illustrate the rationalization degeneration and rationalization failure. Human-annotated rationales are underlined. Rationales from the rationalization framework are highlighted in red.
• Figure 3: Comparison of spurious correlation strength in BERT and GRU representations.
• Figure 4: The average trace $\operatorname{tr}(\Sigma)$ of all sentences in different layers. Testing on the HotelReview dataset. The results are verified on three models: Bert, Electra, and Roberta.
• Figure 5: The proposed rationalization architecture PLMR. The dashed lines are used only during the training phase.