Research map / Label-efficient and robust learning / Recognition and long-tail
Is the MMI Criterion Necessary for Interpretability? Degenerating Non-causal Features to Plain Noise for Self-Rationalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Maximizing the Remaining Discrepancy (MRD), a rationale-extraction criterion that treats spurious features as plain noise by maximizing the KL divergence between the predictor's outputs on the full input and on the input with the rationale removed.
Problem
Rationale extraction via the maximum mutual information (MMI) criterion is confused by spurious features that are correlated with the label, which complicates the extractor's loss landscape and yields non-causal rationales.
Approach
Instead of scoring the selected rationale Z with MMI, score the remaining input X_{-Z}: maximize the KL divergence between P(Y|X) and P(Y|X_{-Z}). Removing noise or spurious features leaves the conditional label distribution unchanged, so both are treated equivalently, and only removing causal features changes it. No penalty regularizers are needed.
Key results
- MRD improves rationale F1 over the strongest baseline MCD by up to 10.4 points (63.5 vs 53.1 on Beer-Palate at ~10% sparsity).
- MRD reaches F1 65.7 on Beer-Appearance at ~10% sparsity versus 63.9 for MCD.
- With BERT at ~10% sparsity, MRD improves F1 by 15.6 points on Beer-Appearance over CR (54.6 vs 39.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace MMI with MRD | RNP rationale extraction with a GRU encoder · Beer-Appearance (S≈10%) | F1: 63.9 -> 65.7 | Helped |
| Replace MMI with MRD | RNP rationale extraction with a GRU encoder · Beer-Aroma (S≈10%) | F1: 65.8 -> 68.1 | Helped |
| Replace MMI with MRD | RNP rationale extraction with a GRU encoder · Beer-Palate (S≈10%) | F1: 53.1 -> 63.5 | Helped |
| Replace MMI with MRD | RNP rationale extraction with a GRU encoder · Hotel-Location (S≈10%) | F1: 52.7 -> 54.4 | Helped |
| Replace MMI with MRD | RNP rationale extraction with a GRU encoder · Hotel-Service (S≈10%) | F1: 44.7 -> 46.3 | Helped |
| Replace MMI with MRD | RNP rationale extraction with a GRU encoder · Hotel-Cleanliness (S≈10%) | F1: 35.9 -> 36.6 | Helped |
| Replace MMI with MRD | RNP rationale extraction with a BERT encoder · Beer-Appearance (S≈10%) | F1: 39.0 -> 54.6 | Helped |
| Replace MMI with MRD | RNP rationale extraction with a BERT encoder · Beer-Aroma (S≈10%) | F1: 49.1 -> 55.1 | Helped |
Surprising findings
- Penalty-based Inter_RAT is outperformed by more than 10 F1 points in 14 of 18 settings, which the authors take as evidence of limitations of penalty-based methods.
- Some baselines become highly hyperparameter-sensitive when switched to BERT, so the BERT comparison is limited to methods already implemented with BERT.
Limitations
- Does not evaluate on LLMs or other domains such as computer vision and graph learning.
- Results with BERT are a supplementary comparison with fewer baselines due to hyperparameter sensitivity and limited compute.
- Baseline results are mostly copied from prior papers rather than re-run, given hyperparameter sensitivity.
Experimental setup
Datasets: BeerAdvocate HotelReview
Benchmarks: BeerAdvocate HotelReview
Metrics: precision recall F1 sparsity
Baselines: RNP INVRAT Inter_RAT NIR MCD CR VIB A2R
Builds on: Rationalizing Neural Predictions (RNP) minimum conditional dependence (MCD) invariant risk minimization (INVRAT) Inter_RAT CR
Closest papers by meaning
- Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization (ICLR 2025 · similarity 0.89)
- The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability? (NeurIPS 2025 · similarity 0.81)
- Identifiability Guarantees for Causal Disentanglement from Purely Observational Data (NeurIPS 2024 · similarity 0.81)
- Rethinking Dimensional Rationale in Graph Contrastive Learning from Causal Perspective (AAAI 2024 · similarity 0.81)
- Mitigating Spurious Correlations via Disagreement Probability (NeurIPS 2024 · similarity 0.80)
- Learning Robust Rationales for Model Explainability: A Guidance-Based Approach (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.