Many text classifiers can now point to the words that justify their answer. These highlighted words are called a rationale. The most trustworthy version is a two-step pipeline: one model selects a snippet of the input, and a second model predicts the label from that snippet alone. Because the second model never sees anything else, the explanation is faithful by construction. The catch is accuracy. The ERASER benchmark, which standardized rationale evaluation across seven tasks, found that such pipelines lose substantial accuracy compared with ordinary classifiers.
This note argues that part of that loss comes from how the snippet selector is taught. It is trained to reproduce one human's highlight token by token, even though humans disagree about exactly where evidence starts and stops. That disagreement is well documented in extraction work. On clinical notes, for example, entity F1 rises from 0.7724 under exact matching to 0.8447 under overlap matching (Dutch ADE corpus). We propose to move "relaxed" span matching out of the evaluation script and into the training loss of the snippet selector. The loss gives credit for boundary-shifted and alternative valid highlights. It is also set-based, so it still punishes bloated spans. We then test whether this recovers pipeline accuracy without loosening faithfulness.
Core Claim
The claim is specific. A select-then-predict pipeline whose selector is trained with a differentiable, multi-reference soft intersection-over-union (IoU) loss will match or beat token-level cross-entropy supervision on downstream accuracy at equal rationale length. IoU is the overlap between two sets divided by their union. The pipeline will also keep faithful-by-construction guarantees. The expected mechanism is that exact-token supervision teaches the selector to drop the modifiers, negations and transitional words that annotators include inconsistently. The predictor often needs exactly those words. The unsupervised editor SHARE observed the same problem when selectors lose context such as "not" next to an adjective.
Relaxed evaluation work has one standing weakness: loose rules like "any shared token" reward spans that swallow the whole sentence. IoU is not vulnerable to this, because a larger prediction enlarges the union and lowers the score. Relaxation here therefore forgives boundary jitter but not bloat.
Mechanism
Let $m_i \in [0,1]$ be the selector's soft mask on token $i$. Following ID-ExpO, the mask comes from a sigmoid with temperature $T$, which is annealed toward hard 0/1 masks during training. Let $r^{(k)}_i \in \{0,1\}$ be the $k$-th available human highlight. The rationale loss is
$$\mathcal{L}_{rat} = 1 - \max_k \frac{\sum_i m_i r^{(k)}_i}{\sum_i m_i + \sum_i r^{(k)}_i - \sum_i m_i r^{(k)}_i}.$$
The maximum over references means the selector is rewarded for matching any one valid annotator's highlight, rather than an average of them. Averaging would blur disjoint alternatives into a meaningless middle. This is the rationale analogue of cluster-based matching in extraction evaluation. In practice a smooth log-sum-exp stands in for the hard max.
The full objective adds the predictor's cross-entropy on the masked input, which rewards sufficiency. At inference the mask is binarized, so the predictor sees only the selected tokens.
The idea borrows from two lines of work but differs from both. ID-ExpO relaxed hard top-$s$ feature selection into soft masks, but for image and tabular attribution heatmaps with no human references. We apply the same relaxation to human-supervised text spans. MARE deliberately binarizes masks to stop information leaking across aspects. Our design is compatible with that concern, because binarization still happens at inference. What we relax is only the training target.
Research Program
We will build one selector–predictor pipeline on a shared encoder and vary only the rationale supervision. This isolates the effect of the loss.
Beyond ERASER's own metrics, we add an evaluation that the training loss cannot directly optimize. Adapting the optimal-assignment idea of AESOP, we match predicted rationale segments to reference segments one-to-one and score each pair by Jaccard overlap. This multi-span score, the pipeline's task accuracy, and ERASER's comprehensiveness and sufficiency form the primary evidence. We do not rely on IoU F1 as primary evidence, because the model is trained on a close relative of it and that would be circular.
Where multiple annotator highlights per example exist, we use them as the multiple references. We expect this in e-SNLI evaluation splits, though it must be confirmed from the release. Elsewhere $k=1$, and the method reduces to single-reference soft IoU. That reduction is itself an ablation.
We will also run a failure analysis. It categorizes the tokens gained or lost relative to the baseline selector (negations, modifiers, function words) and checks whether gains in accuracy trace to those tokens. This directly tests the proposed mechanism rather than just the outcome.
Baselines
The naive reference point is the ERASER-style hard pipeline. Its selector is trained with token-level binary cross-entropy against a single gold highlight, and its predictor sees only the selection. This is exactly the configuration that shows the accuracy penalty. We run it with the same encoder, data, epochs and hyperparameter search budget as our method, and report it at matched rationale length.
The upper reference is the same encoder trained as an ordinary full-input classifier. It shows how much of the pipeline gap remains. On e-SNLI, the older e-InferSent system reports about 84% accuracy (e-SNLI), which calibrates the expected range.
The strongest existing rationale-only competitor is SHARE. It reports 0.894 accuracy on Movie Reviews using only its extracted rationale, against 0.769 for the VCU baseline (SHARE). We will quote this number and also rerun SHARE on our encoder. Beating it means exceeding its rationale-only accuracy at no greater rationale length, while also scoring higher on the multi-span plausibility metric. SHARE uses no human highlights, so it has no plausibility anchor.
For ERASER's supervised pipeline, no reported result was found on Lacuna for specific ERASER baseline scores. We will therefore produce these numbers ourselves under the shared protocol.
Two ablations isolate the key ingredient. The first is single-reference soft IoU on e-SNLI, which tests whether multiple references matter. The second is soft IoU without the predictor's sufficiency term, which tests whether relaxation helps plausibility alone or specifically helps the predictor.
Why not just…?
Why not just smooth the gold mask near its boundaries and keep cross-entropy? This is the real threat. Spreading label mass onto neighbouring tokens gives much of the boundary tolerance at almost no cost. It also has two predicted weaknesses. First, it cannot represent two disjoint valid highlights. Second, it penalizes tokens independently, so it has no set-level pressure against extending a span. We include boundary-smoothed cross-entropy as a full baseline under the same tuning budget. If it matches soft IoU on single-reference datasets, we will say so. The contribution then narrows to the multi-reference case on e-SNLI, which smoothing cannot express.
Why not just evaluate leniently, as ERASER already does with IoU F1? Lenient evaluation changes the score, not the model. The pipeline's accuracy penalty in ERASER persists regardless of how plausibility is scored. Our decisive test is downstream accuracy at fixed length, which evaluation choices cannot move.
Why not skip human highlights and use unsupervised extraction like SHARE? Unsupervised selectors can be accurate, but nothing ties them to what humans consider evidence. RORA shows that label-predictive text can be vacuous or leaky, which is why predictive usefulness alone is a weak quality signal. Our comparison against SHARE reports both accuracy and multi-span plausibility. That shows whether human supervision buys plausibility without costing accuracy.
Why not use a full classifier plus a post-hoc attribution method? ERASER found that attention and post-hoc attributions give weaker faithfulness guarantees than a pipeline in which the predictor literally cannot see unselected text. We keep the guarantee and attack its cost.
Proposed Experiments
We run on the ERASER suite: Movie Reviews, e-SNLI, BoolQ, FEVER, MultiRC and Evidence Inference. The last is the clinical-trial task where boundary disagreement should be largest.
The first decisive check is pipeline accuracy at matched rationale length, relative to cross-entropy and boundary-smoothed supervision, with seeds and confidence intervals. The second is a bloat control. We train a variant under the permissive "any shared token" rule. We expect its rationales to grow and its comprehensiveness to become uninformative, while soft IoU keeps length near the references. The third is the e-SNLI multi-reference ablation. The fourth is the token-category analysis of which gained tokens drive accuracy.
The hypothesis is supported if soft IoU improves accuracy over cross-entropy on at least the long-document tasks at equal length, and the smoothing baseline does not close the gap on multi-reference data. If boundary-smoothed cross-entropy matches soft IoU everywhere, the conclusion is still useful: the accuracy penalty of faithful pipelines is largely a boundary-supervision artifact that a trivial fix removes.
References
- ERASER: A Benchmark to Evaluate Rationalized NLP Models
- Creation of a gold standard Dutch corpus of clinical notes for adverse drug event detection: the Dutch ADE corpus
- A Shapley-value Guided Rationale Editor for Rationale Learning
- Explanation-based Training with Differentiable Insertion/Deletion Metric-aware Regularizers
- MARE: Multi-Aspect Rationale Extractor on Unsupervised Rationale Extraction
- Learning to Extract Structured Entities Using Language Models
- e-SNLI: Natural Language Inference with Natural Language Explanations
- RORA: Robust Free-Text Rationale Evaluation