XAI meets Natural Language Processing
summary
Traditional XAI often fails on text so the session shows concrete options and what actually worked: using human label explanations as predicted outputs, extracting attention and feature importances from BERT layers, and applying LIME SHAP ELI5 for local and global views. A PoC dashboard combines custom TF‑IDF embeddings plus engineered features and displays SHAP aggregated by paragraph so customers see which paragraph drove a binary decision — proof that the approach is usable in production.
AI-written summary · about
chapters (11)
- 0:00Intro
- 1:51What is Explainable Al?
- 4:42Why bother?
- 6:25Common Approaches & Evaluations
- 9:52Feature Imporance & Model Structures
- 11:33Understandable Models
- 15:31Difference between LIME & SHAP
- 18:30Graphical Representations
- 20:43Difficulties for NLP Models
- 23:00Real World Example - Setup
- 24:38Lessons Learned
related talks
47mML: Explainable Artificial Intelligence explained
31mWhat We Expect from XAI - A scientist’s experience
30mTrustworthy AI - opening up the black box
39mThe Search for Transparency and Accountability in the Age of...
15mIntroduction to Explainable AI- 39mSponsored Session: The Search for Transparency and Accountability in the Age of AI: RAI & XAI as Essential Tools
35mSHAPtivating Insights: unravelling blackbox AI models
20mBridging the gap between AI and UI