Enhancing the Interpretability of NLI Models Using LLMs and Active Learning Algorithms

Qi Wang, Junqiang LiuInformation 2026 · DOI

TL;DR

This paper introduces EGM-PM, a novel framework that uses Large Language Models (LLMs) to automatically generate natural language explanations for Natural Language Inference (NLI) tasks. Coupled with a new active learning algorithm called ETCM, this approach significantly reduces annotation costs and improves model performance and interpretability, even with limited data.

What Problem Does This Paper Address?

Interpreting how Natural Language Inference (NLI) models arrive at their decisions is a major challenge. While manually creating natural language explanations for NLI tasks can improve interpretability, it's incredibly expensive, time-consuming, and prone to inconsistencies. Existing datasets often lack these explanations, and current active learning methods don't effectively leverage the semantic richness of explanations for efficient data selection. This creates a bottleneck for developing truly understandable and reliable NLI systems, especially in low-resource settings.

Methods

The core of this work is the EGM-PM (Explanation Generation Model-Prediction Model) framework.

  • LLMs as Explanation Annotators: Instead of relying on human experts, the paper uses LLMs, guided by carefully engineered prompts, to automatically generate natural language explanations for premise-hypothesis pairs. This significantly cuts down annotation costs and time.
  • ETCM (Explanation-aware Transition from Clustering to Margin) Active Learning Algorithm: This novel algorithm intelligently selects the most informative samples for annotation. It operates in two stages:
  • TypiClust-Explanation (Stage 1): In the initial "cold-start" phase, it uses a combination of textual semantics and early-stage LLM explanations to ensure diversity and representativeness in sample selection.
  • MarginNLI (Stage 2): As more data is labeled, it shifts to an uncertainty-based approach, incorporating both prediction uncertainty and an "explanation semantic discrepancy" to focus on samples that are both uncertain and offer novel explanatory insights.
  • Dual-Model Architecture: The EGM generates explanations, which then serve as an additional input to the Prediction Model (PM), alongside the premise-hypothesis pair, to improve its inference accuracy and interpretability.

Datasets

  • e-SNLI: A publicly available dataset used for evaluating the framework.
  • ExNLI (Explainable Natural Language Inference): A newly constructed dataset created by the authors. It's built upon XNLI and OCNLI, augmenting existing premise-hypothesis pairs with LLM-generated natural language explanations. ExNLI aims to provide a scalable and cost-effective resource for explainable NLI research, offering cross-lingual coverage and diverse Chinese expressions.

Key Insights and Ideas

  • LLMs as Scalable Explanation Generators: The paper demonstrates that LLMs, with proper prompt engineering, can effectively replace human annotators for generating natural language explanations, making explainable AI more accessible.
  • Active Learning Enhanced by Explanations: Integrating explanation semantics into active learning sampling (ETCM) is crucial. It allows the model to learn relational and logical structures more effectively than relying solely on textual data or prediction uncertainty.
  • Dynamic Sampling Strategy is Key: The two-stage ETCM algorithm, which transitions from diversity-focused to uncertainty-focused sampling, is more effective than static approaches. This dynamic adaptation reflects the evolving nature of information gain as more data becomes available.
  • Explanations as Auxiliary Signals: Natural language explanations aren't just for human understanding; they can act as valuable supervisory signals for the prediction model itself, improving its performance even with smaller models.

Key Results

  • Improved Performance with Less Data: ETCM-based active learning approaches full-data performance significantly faster and with fewer labeled samples compared to other active learning algorithms.
  • Accuracy Gains: The EGM-PM framework, particularly with ETCM, achieved an accuracy of 88.89% on the e-SNLI dataset, outperforming previous state-of-the-art methods, including those using much larger models or more complex training regimes.
  • Efficiency: The proposed approach significantly reduces annotation costs and enables efficient model training, even on a single GPU for explanation generation.
  • Cross-Task Transferability: The EGM-PM framework showed promise in question answering tasks, indicating that the learned "evidence searching" capability from NLI explanations can generalize.

Limitations and Future Work

  • Human Evaluation Subjectivity: While rigorous controls were in place, human evaluation of explanations can still be influenced by subjective factors like fluency and conciseness. Future work could involve more comprehensive, large-scale evaluations.
  • Logical Validity of Explanations: LLM-generated explanations, while helpful, aren't guaranteed to be logically sound in all cases. They might sometimes rely on tautological paraphrasing or hallucinated details. The current framework partially mitigates this through active selection but doesn't explicitly filter out such "plausible but ungrounded" explanations.
  • Computational Overhead: The ETCM algorithm introduces additional computational cost due to the need for encoder inference for explanation embeddings.
  • Dependence on Base LLM Capability: The effectiveness of the generated explanations relies on the underlying LLM's reasoning capabilities, which might be a limitation in extremely low-resource languages.
  • Future Work: The authors suggest improving the faithfulness of generated explanations and extending the EGM-PM framework to more complex reasoning tasks beyond NLI, such as multi-hop question answering.

Summary updated

Abstract

In the field of Natural Language Inference (NLI), model interpretability remains an urgent and unresolved challenge. Existing interpretability-oriented annotated datasets are highly limited, and manually constructing natural language explanations is both costly and inconsistent, making it difficult to balance model performance and interpretability. To address this issue, this paper proposes an interpretable NLI framework based on active learning, Explanation Generation Model-Prediction Model (EGM-PM), and designs an active learning sampling algorithm, Explanation-aware Transition from Clustering to Margin (ETCM), that incorporates natural-language explanation information. In this framework, Large Language Models (LLMs) are employed to automate explanation annotation, reducing dependence on human experts in traditional active learning. A small number of high-value samples obtained via ETCM sampling are used to train the EGM, whose generated natural-language explanations are then used to guide the PM in label inference. Experimental results show that data sampled by ETCM substantially enhance the model’s ability to learn relational and logical structures between premise–hypothesis pairs. Compared with other active learning algorithms, ETCM approaches full-data performance more rapidly while using significantly fewer labeled samples. This finding confirms the value of natural language explanation semantics in improving both model performance and interpretability. Furthermore, this paper employs prompt engineering to construct an interpretability-oriented NLI dataset, Explainable Natural Language Inference (ExNLI), which augments traditional premise–hypothesis pairs with natural-language explanations. Human and automated evaluations confirm the consistency and faithfulness of these explanations. The dataset has been publicly released, offering a low-cost and scalable data construction approach for future research on explainable NLI.
BibTeX
@inproceedings{wang2026enhancing,
  author = {Qi Wang and Junqiang Liu},
  title = {{Enhancing the Interpretability of NLI Models Using LLMs and Active Learning Algorithms}},
  booktitle = {Information},
  year = {2026},
  doi = {10.3390/info17020119},
}
Figure