An extractive rationale model explains a text classifier by picking out the words it used. It has two parts: a selector that highlights a subset of the input, and a predictor that sees only the highlighted text. Training such a model means trading off three goals. The first is accuracy. The second is brevity: highlight little. The third is contiguity: highlight whole phrases rather than scattered words. In current practice each of these gets a hand-set penalty weight, and a bad choice of weights sends the selector to an empty highlight, the whole document, or unreadable salt-and-pepper fragments. The weights also have no units. Nobody can say in advance what "contiguity weight 0.3" means for the explanation that comes out.
We propose to flip the usual formulation. Accuracy becomes a hard constraint, set relative to a model that reads the full text. Contiguity becomes a second constraint, stated in tokens per span. Length is the single quantity the selector minimizes. A constrained-optimization method, the Modified Differential Multiplier Method (MDMM), adjusts the penalty weights automatically during one training run. Every knob the user sets now has a meaning: "stay within 0.02 nats of the full-text model's loss" and "average highlighted span at least six tokens." The rationale length is no longer chosen by hand. It becomes an output of training that reports how much text a dataset actually needs.
Core Claim
The research file finds no prior rationalization work that uses automated loss weighting. Every method it covers balances its terms with static weights tuned by hand: the REINFORCE model of Rationalizing Neural Predictions, the fused-lasso length control of LimitedInk in Are Shortest Rationales the Best Explanations for Human Understanding?, the self-trained single Transformer of RTP, and the multi-loss editor of SHARE.
The naive fix would be to bolt an automatic weighting scheme onto the existing three-term loss. That is not new enough to be worth doing. It also runs into a known partial precedent. Lagrangian methods that drive sparsity and transition counts toward a fixed target already exist in rationalization, notably HardKuma, which the file mentions only as a SPECTRA baseline. Those methods still require a human to pick the target length.
Our claim is narrower. If the constraint is placed on predictive fidelity, measured against an external full-text reference model, then minimizing length under a contiguity floor yields three things. First, the accuracy loss of the rationale model is bounded by a tolerance the user sets, where hard pipelines otherwise lose 5–10% (ERASER). Second, the resulting length differs by dataset and is found automatically. Third, sweeping the tolerance traces a better accuracy–brevity curve than static weights given the same number of training runs.
Research Program
Let $\ell_{\text{rat}}$ be the predictor's validation-tracked cross-entropy when it reads only the rationale. Let $\ell_{\text{full}}$ be the loss of a separately trained classifier on the full text. Let $\delta$ be the allowed gap. Let $r$ be the fraction of tokens selected, and let $\bar s$ be the mean length of a selected span. The training problem is
$$\min_{\theta}\; r(\theta) \quad \text{s.t.} \quad \ell_{\text{rat}}(\theta) - \ell_{\text{full}} \le \delta, \qquad \bar s(\theta) \ge s_{\min},$$
where $\theta$ covers the selector and the predictor and $s_{\min}$ is the contiguity floor in tokens. All three quantities are made differentiable with the usual relaxed masks, so that $r$ is the expected mask sum and span count is the expected number of 0-to-1 transitions.
MDMM keeps one multiplier $\mu_j \ge 0$ per constraint $g_j(\theta)\le 0$. It takes gradient descent steps on $\theta$ and gradient ascent steps on the multipliers, $\mu_j \leftarrow \max(0, \mu_j + \eta\, g_j(\theta))$, with $\eta$ the multiplier step size. It adds a damping term that suppresses oscillation.
This design borrows two things from prior work and departs from each in one respect.
- From Park et al.: we take MDMM and the idea of an external standalone model as the setpoint. They constrain an auxiliary reconstruction loss in speech synthesis. We instead constrain the primary task loss and minimize the structural cost.
- From M-HOF-Opt: we take the view of multipliers as a feedback controller, which they use with shrinking Pareto setpoints on image domain generalization. M-HOF-Opt reports that its multipliers oscillate once setpoints become tight. That result shapes our design: it is why we use fixed, interpretable setpoints plus damping rather than setpoints that keep shrinking.
- From CoV-Weighting: we take a warning. Multi-task weighting schemes over-weight auxiliary regularizers in single-task problems. We use it as a baseline, not as the mechanism.
We will build the constrained trainer on two selector types. The first is a generator–predictor pipeline with a BERT backbone and relaxed binary masks. The second is RTP-style continuous masking, to test whether the controller transfers across relaxations. For both, we will analyze three things. We will track the multiplier trajectories, to check whether fidelity gets enforced early and length pressure rises only later. We will measure how well the gap on the validation set predicts the gap on the test set. We will record how the discovered length $r^*$ varies across datasets and tolerances $\delta$.
Baselines
Unconstrained full-text classifier. This is the naive reference point and doubles as the setpoint model. It gives the accuracy ceiling against which the 5–10% hard-pipeline penalty is measured. It is trained with the same backbone and budget as everything else.
Static-weight grid search, matched in compute. This is the three-term Lei-style loss of Rationalizing Neural Predictions with fused-lasso contiguity as in LimitedInk. The grid is given exactly as many full runs as our $\delta$ sweep uses, and a second, larger grid gets five times that budget. To beat it, our method must dominate the grid's accuracy–length curve at equal runs, and it must match the five-times grid within seed noise.
SPECTRA, the strongest structured competitor. SPECTRA hard-codes contiguity and a budget in a factor graph. Its reported results are a token F1 of 0.61 against human rationales on BeerAdvocate and a Macro F1 of 0.91 on Hotels. We run it at the budget our method discovers, so both methods produce equal lengths. Beating it means equal or higher Clf-F1 and IoU-F1 at that length. It also means our method's accuracy gap stays within $\delta$ on datasets where SPECTRA's fixed budget is either too tight or too loose.
Plausibility baselines. On USR Movies, the strongest reported plausibility number is from BA2R+NI in Unsupervised Selective Rationalization with Noise Injection: rationale F1 of 38.6 at 85.7% accuracy. We treat this as a plausibility bar to match, not to exceed. Our method targets fidelity and brevity, not agreement with human highlights.
Other automatic weighting schemes on the same three-term loss. These are CoV-Weighting and M-HOF-Opt. Beating them means a smaller accuracy gap and lower seed variance.
Ablations. The first swaps the roles, so length is the constraint (fixed at 40% following Shen et al.) and loss is the objective. The second removes the external setpoint, so the gap is measured against the rationale model's own running full-text loss instead of a separate model. The third removes damping.
Why not just…?
Why not just grid-search static weights? This is the most serious threat. If a small grid already finds points on our curve, the controller adds little. Prior evidence cuts the other way: fixed multipliers reached 0.748 accuracy on PACS Sketch versus 0.806 for M-HOF-Opt (Sun et al.), and static weights gave 36.28% word error rate versus 19.36% with MDMM (Park et al.). But those results come from other domains. The matched-compute grid comparison settles the question for text, and we will report it even if the grid wins.
Why not just fix the length at the 40% sweet spot? Shen et al. found that human accuracy rose from about 61% at 10% length to about 85% at 50%, saturating near 40%. That finding comes from Movie Reviews alone. For short FEVER claims or long Evidence Inference documents, 40% is unlikely to be the right length. The role-swap ablation tests this directly. If it matches our method's accuracy on every dataset, the fidelity constraint is unnecessary and we will say so.
Why not just warm up the penalties with a schedule? Ramping the sparsity and contiguity weights up from zero captures the intuition of "learn to predict first, then prune" for free. We include a linear warm-up baseline that also gets its own grid. Our method must show that it bounds the accuracy gap; a schedule only hopes to.
Why not just use SPECTRA's hard structure and skip loss weighting? SPECTRA removes the need for weights, but only by fixing a budget per dataset, and the authors report that its inference adds computational overhead. Our method supplies the budget SPECTRA lacks. A useful hybrid is SPECTRA run at our discovered $r^*$, which we include.
Proposed Experiments
We use the ERASER tasks with human rationales: Movies, BoolQ, Evidence Inference, FEVER, and MultiRC (ERASER). We add USR Movies and Hotels. We skip BeerAdvocate for training, given the file's note that USR was built as a replacement for it.
For each dataset we sweep $\delta \in \{0.01, 0.05, 0.1\}$ nats and $s_{\min} \in \{3, 6\}$, with five seeds each. We report Clf-F1, token F1, IoU-F1, AUC-PR, sufficiency, comprehensiveness, length fraction, and mean span length. We also report hypervolume under the accuracy-versus-length curve and the standard deviation of the final length across seeds. For RTP we found no reported numbers on Lacuna, so we rerun it.
Three checks decide whether the idea holds. First, the test-set accuracy gap must stay at or below $\delta$ on at least four of the six datasets. If it does not, the core guarantee fails. Second, the discovered $r^*$ must vary meaningfully across datasets, not collapse to one value. If it collapses, a fixed budget suffices. Third, our method must dominate the matched-compute static grid in hypervolume, with lower seed variance.
A small follow-up human study, run on Movies with the Shen et al. protocol, would test whether the automatically chosen length lands near the comprehension plateau.
References
- Rationalizing Neural Predictions
- Are Shortest Rationales the Best Explanations for Human Understanding?
- Rationalizing Transformer Predictions via End-To-End Differentiable Self-Training
- A Shapley-value Guided Rationale Editor for Rationale Learning
- ERASER: A Benchmark to Evaluate Rationalized NLP Models
- Automatic Tuning of Loss Trade-offs without Hyper-parameter Search in End-to-End Zero-Shot Speech Synthesis
- M-HOF-Opt: Multi-Objective Hierarchical Output Feedback Optimization via Multiplier Induced Loss Landscape Scheduling
- Multi-Loss Weighting with Coefficient of Variations
- SPECTRA: Sparse Structured Text Rationalization
- Unsupervised Selective Rationalization with Noise Injection