Can LLMs Faithfully Explain Themselves in Low-Resource Languages? A Case Study on Emotion Detection in Persian

2025-11-24

AI Summary

This study investigates the faithfulness of self-explanations generated by large language models (LLMs) for sentiment classification in low-resource languages—specifically Persian—where ground-truth reasoning processes are scarce and difficult to model. Method: We propose two prompting strategies—“predict-then-explain” and “explain-then-predict”—and introduce a token-level log-probability-based confidence score to quantify explanation reliability. Faithfulness is rigorously evaluated via human annotation assessing alignment between model-generated explanations and actual human reasoning. Contribution/Results: Despite high classification accuracy, LLMs exhibit significantly low agreement with human annotations; inter-model explanation similarity substantially exceeds model–human alignment, revealing systematic faithfulness deficits in self-explanation for low-resource settings. To our knowledge, this is the first work to empirically quantify and validate the unreliability of LLM self-explanations in Persian, establishing a methodological benchmark and issuing an evidence-based caution for explainable AI in low-resource languages.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large language models (LLMs) are increasingly used to generate self-explanations alongside their predictions, a practice that raises concerns about the faithfulness of these explanations, especially in low-resource languages. This study evaluates the faithfulness of LLM-generated explanations in the context of emotion classification in Persian, a low-resource language, by comparing the influential words identified by the model against those identified by human annotators. We assess faithfulness using confidence scores derived from token-level log-probabilities. Two prompting strategies, differing in the order of explanation and prediction (Predict-then-Explain and Explain-then-Predict), are tested for their impact on explanation faithfulness. Our results reveal that while LLMs achieve strong classification performance, their generated explanations often diverge from faithful reasoning, showing greater agreement with each other than with human judgments. These results highlight the limitations of current explanation methods and metrics, emphasizing the need for more robust approaches to ensure LLM reliability in multilingual and low-resource contexts.

Problem

• Evaluating faithfulness of LLM-generated explanations in Persian emotion detection
• Comparing model-identified influential words against human annotations
• Assessing impact of explanation-prediction order on faithfulness metrics

Innovation

• Evaluates faithfulness using token-level confidence scores
• Compares model explanations with human annotations in Persian
• Tests two prompting strategies for explanation generation