o-MEGA: Optimized Methods for Explanation Generation and Analysis

Ľuboš Kriš, Jaroslav Kopčan, Qiwei Peng, Andrej Ridzik, Marcel Veselý, Martin TamajkaEMNLP 2025

TL;DR

Transformer models are powerful but opaque. This paper introduces o-MEGA, an automated tool that optimizes the selection and configuration of explainable AI (XAI) methods for semantic matching tasks, making complex models more transparent and trustworthy, especially for critical applications like fact-checking.

What Problem Does This Paper Address?

The incredible success of modern NLP models, particularly transformers, has come at the cost of increased complexity, making them "black boxes." This lack of transparency hinders their adoption in critical applications where understanding model decisions is paramount, such as automated fact-checking. While numerous explainable AI (XAI) methods exist, choosing the right one and configuring it optimally for a specific task, model, and dataset is a significant challenge. This often involves a time-consuming, resource-intensive, and subjective manual process. The paper addresses this gap by automating the selection and optimization of XAI methods for semantic matching tasks.

Methods

The core contribution is o-MEGA, a framework designed to automate the process of finding the most effective explainable AI (XAI) methods and their configurations. It operates by systematically evaluating a diverse set of XAI algorithms and their hyperparameters against user-defined criteria.

Key aspects of the o-MEGA framework include:

  • Comprehensive XAI Method Space: o-MEGA integrates 11 explainable methods, spanning gradient-based, perturbation-based, and architecture-specific paradigms. These methods are chosen for their ability to capture token-level or word-level importance and cross-sequence interactions, crucial for semantic matching.
  • Model Agnosticism (with caveats): While tested with specific transformer models for semantic matching (listed in Table 1), the framework is designed to work with any transformer-based model for this task.
  • Quantitative Evaluation Metrics: The framework focuses on quantitative metrics for explanation quality, specifically fidelity (how well the explanation reflects the model's actual behavior) and plausibility (how easy the explanation is for humans to understand).
  • Intelligent Optimization: o-MEGA leverages the Optuna library for hyperparameter optimization. Instead of exhaustive grid search, Optuna employs intelligent search strategies like Tree-structured Parzen Estimator (TPE) and Bayesian optimization, which learn from previous evaluations to efficiently explore the hyperparameter space and find near-optimal configurations. This is particularly beneficial given the numerous hyperparameters of XAI methods.
  • Configurability: Users can select models, define weights for fidelity and plausibility metrics, and choose from various search strategies offered by Optuna.

Datasets

The paper uses the MultiClaim dataset for their case study. This dataset consists of post-claim pairs that have been fact-checked by professional fact-checkers. For evaluation, they also utilize human annotations for 512 post-claim pairs, where annotators provide explanations for why specific claims were matched.

Key Insights and Ideas

  • Explainability is Not One-Size-Fits-All: The paper strongly emphasizes that different tasks, models, and datasets require different XAI approaches. Automating this selection is crucial for practical deployment.
  • Trade-off Between Fidelity and Plausibility: While high fidelity and plausibility are desirable, they often conflict. o-MEGA allows users to balance these or prioritize one, reflecting real-world needs.
  • Occlusion as a Top Performer: For the semantic matching task studied, the Occlusion method (both token-level and word-level) emerged as the recommended explainability method, achieving a good balance between technical accuracy (fidelity) and human interpretability (plausibility).
  • Efficiency Gains with Optuna: Using Optuna for hyperparameter optimization significantly reduces the time and computational resources required to find optimal XAI configurations compared to exhaustive search methods. This makes automated XAI selection more feasible for routine deployment.
  • Human Interpretability is Key: While most tested methods achieved high fidelity (>0.94), their plausibility scores varied significantly (0.35-0.68). This highlights that technical accuracy alone is insufficient; user understanding is critical for trust.

Key Results

  • Optimal Method: Occlusion was identified as the best-performing explainability method for the semantic matching task, achieving an average score of 0.819 by balancing faithfulness and plausibility.
  • Efficiency of TPESampler: Among single-objective samplers, TPESampler was found to be the most efficient, delivering the same explanation quality recommendations as exhaustive search methods but reducing optimization time by approximately 66%.
  • Multi-Objective Sampler Performance: BruteForceSampler and TPESampler correctly identified Occlusion as the optimal method, while NSGA-II and NSGA-III variants selected the inferior Saliency method. TPESampler offered a near-identical performance to BruteForceSampler but with significantly less computation time.

Limitations and Future Work

  • Domain Specificity: The current implementation of o-MEGA was developed specifically for claim matching and the MultiClaim dataset, although a text classification pipeline is available. Its immediate applicability to other domains is limited.
  • Metric Restrictions: The evaluation metrics are currently restricted to established ones, excluding experimental metrics like localization and sparseness.
  • Computationally Expensive Methods: Highly computationally expensive explanation methods have not been incorporated, as their high resource requirements might make them less effective for users.
  • Future Work: The authors suggest expanding the tool's applicability to other domains and incorporating a wider range of evaluation metrics. They also plan to explore the integration of computationally expensive explanation methods.

Summary updated

BibTeX
@inproceedings{kris2025o,
  author = {Ľuboš Kriš and Jaroslav Kopčan and Qiwei Peng and Andrej Ridzik and Marcel Veselý and Martin Tamajka},
  title = {{o-MEGA: Optimized Methods for Explanation Generation and Analysis}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure