Evidence Mining for Explainable AI: Methods and Applications

The MLNLP Academic Talk series promotes exchange among machine‑learning and natural‑language‑processing researchers. This session, held on April 18, 2026, features Associate Professor Yue Linan from Southeast University.

Evidence mining extracts key information from raw input text to support model predictions, serving as a crucial technique for explainability in high‑risk decision‑making. Existing methods often fall into the "shortcut learning" trap, over‑relying on superficial features, which reduces the reliability of the extracted evidence, and they are largely confined to small‑model paradigms.

The presentation introduces a shortcut‑avoidance evidence‑mining method that improves the faithfulness and reliability of explanations by addressing the shortcut learning issue. It also describes recent work that combines evidence mining with chain‑of‑thought compression for large models, aiming to enhance the interpretability of efficient reasoning processes in large‑model scenarios.

Speaker bio: Yue Linan is an associate professor in the School of Computer Science and Engineering at Southeast University, PhD graduate of the University of Science and Technology of China, and a member of the PALM lab. His research focuses on knowledge representation and reasoning, efficient large‑model inference, and trustworthy AI. He has authored over 30 papers in CCF‑A/B journals and conferences, received awards such as the Chinese Academy of Sciences President’s Special Award, and serves on editorial boards and AI committees.