From Scoring to Explanations: Evaluating SHAP and LLM Rationales for Rubric-based Teaching Quality Assessment

In three linesInterpretability framework for automated rubric-based assessment of classroom transcripts. Combines SHAP attributions with LLM-generated rationales. On 6k annotated segments: fine-tuned models outperform LLMs in accuracy but compress scores; SHAP identifies driving sentences with robust cross-architecture transfer, unlike LLM rationales.

Summary generated by Claude — human-verified