Faithful Two-Perspective Emotion Explanations via Crossed Interventions

When someone writes "our team is being restructured to improve efficiency," the writer may feel calm while readers feel dread. Datasets that collect both labels confirm that the two diverge. In WRIME, Japanese social-media authors rated their own emotions while outside readers rated the same posts, and writer–reader agreement was lower than reader–reader agreement. Language models asked why a text carries an emotion tend to point at emotion words ("it says sad") rather than at the event that caused the feeling. We propose to train a small language model to predict both the writer's and the reader's emotional intensity and to name, for each, the span of text driving that perspective. We then check whether those named spans actually drive the predictions.

The check is the main contribution. Existing tests of whether a model's explanation is faithful, meaning that it reflects what actually drove the answer, ask whether corrupting a cited reasoning step flips an objective answer. Emotional intensity has no objective answer to flip. Two perspectives on one text give a substitute. If the model says the reader's fear comes from span A and the writer's calm comes from span B, then rewriting A should move the reader prediction more than the writer prediction, and rewriting B should do the reverse. A model that rationalizes after the fact, citing the same emotion words for both views, fails this crossed test. We use the test both as an audit and as a training reward for a compact specialized model.

Core Claim

On posts where writer and reader disagree, the difference between the two perspectives supplies a counterfactual signal that needs neither gold cause annotations nor a deterministic answer. Optimizing a compact model (1.5B–8B parameters) against this signal should produce perspective-specific explanations that are measurably more faithful than those from three alternatives: prompting, rationale distillation, or the single-perspective intervention training already used for math.

Research Program

Each direction plays a distinct role. The writer–reader gap is the data regime and the source of the signal. WRIME provides about 17,000 posts with parallel subjective (writer) and objective (reader) intensities for eight emotions on a 0–3 scale. It also reports that readers are "emotionally muted," underestimating anger and trust in particular. Uncovering the Limits of Text-based Emotion Detection shows the same asymmetry at scale, with writer intent much harder to predict than reader perception.

Causal sentiment analysis defines the output: a cited trigger per perspective rather than one consensual cause. This departs from benchmarks such as RECCON as used by POP-CEE and EMOTRIGGER, which assume a single third-party cause. The post-hoc rationalization literature motivates the problem and supplies the method. Language Models (Mostly) Do Not Consider Emotion Triggers When Predicting Emotion diagnoses hollow, keyword-citing emotion explanations but proposes no fix. FRIT corrupts chain-of-thought steps, labels steps whose corruption flips the answer as causally important, and trains on faithful-versus-unfaithful pairs with direct preference optimization, a method that learns from paired good and bad outputs. Its evaluation covers only GSM8K, SVAMP and StrategyQA, all tasks with objective answers. Model specialization is the constraint. The target is a compact local model, because the activation-patching study found smaller Gemma-2 models less faithful than the 27B one, and because distilled students may copy a teacher's explanation style (Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation) without copying its computation.

The model outputs, for a post $x$, writer and reader intensity distributions and one extracted span per perspective, $s_w$ and $s_r$. The writer side may also output "not in text," since writer intent often rests on unstated personal context. For each cited span we build a style-matched rewrite that keeps fluency and topic but removes the span's affective content. We use rewrites rather than deletions because the activation-patching study showed that out-of-distribution perturbations mislead input-attribution methods like SHAP nearly 90% of the time in counterfactual cases.

Let $\Delta_p(s)$ be the shift in expected intensity for perspective $p$ when span $s$ is rewritten. On posts where the gold writer and reader intensities differ by at least one point, the crossed faithfulness score is

$$D(x) = \big[\Delta_w(s_w) - \Delta_r(s_w)\big] + \big[\Delta_r(s_r) - \Delta_w(s_r)\big].$$

$D$ is high only when each perspective's cited span selectively moves that perspective's prediction. On agreement posts we report only per-perspective comprehensiveness, $\Delta_p(s_p)$ compared against an equal-length random span, so the model is never pushed to invent divergence. Preference pairs come from sampling several outputs per post. The preferred output has high $D$ and low intensity error, and the dispreferred output has accurate predictions but low $D$. This pairing keeps the reward from trading accuracy for faithfulness.

FRIT reports that faithfulness drifts once weights change. To address this, pairs are regenerated from the current model every round. On a subset we confirm text-level $D$ against internal activation patching in the style of the Causal Faithfulness score, which tests whether the internal components driving the answer also drive the explanation. This guards against a model that games surface rewrites.

Backbones are a 1.5B and a roughly 7B open multilingual model, since WRIME is Japanese. WRIME also supplies writers' Big Five personality scores. These are given as an optional writer-side input, because that paper showed they reduce error on writer labels.

Baselines

The naive reference is the WRIME paper's fine-tuned BERT, which reports a mean absolute error (MAE) of 0.519 on reader labels and 0.553 on writer labels. It gives no explanations, so it bounds prediction quality only. Our model must stay within about 0.02 MAE of it to show that faithfulness is not bought with accuracy.

The strongest prediction-only competitor is the same compact backbone fine-tuned with two regression heads on identical WRIME splits. Its explanations are extracted by integrated gradients, an attribution method that sums input gradients along a path from a blank input to the real input.

The strongest explanation competitor is frontier-model prompting. EMOTRIGGER reports GPT-4 trigger-identification F1 of about 0.68 and Llama2/Alpaca around 0.25, against emotion-classification Macro-F1 of 0.907–0.920. On WRIME the frontier model receives the same two-perspective output format and few-shot examples. Its $D$ is measured with the same rewrites; activation patching is impossible for a closed model.

The distillation baseline follows Specializing Smaller Language Models towards Multi-Step Reasoning. The student is trained on the frontier model's two-perspective rationales and labels, with the same compute and token budget as our method.

The key ablation is single-perspective FRIT. It keeps the identical preference-optimization pipeline but rewards only diagonal effects, $\Delta_p(s_p)$, and drops the crossed off-diagonal terms. It isolates whether the second perspective is what adds the signal.

All trained systems share data, backbone, number of preference rounds and a tuning budget of equal size on a shared validation split. No faithfulness number on WRIME is reported on Lacuna, since no faithfulness result on WRIME exists. Beating the competitors means a higher mean $D$ than both the distillation baseline and single-perspective FRIT, with bootstrap confidence intervals that do not overlap, at matched MAE.

Why not just…?

Why not just prompt GPT-4 to explain both perspectives? Its trigger F1 of about 0.68 on EMOTRIGGER is far below its labeling accuracy. That gap is the trigger-label confusion this proposal targets, and fluent explanations from it can still be decorative. Prompting remains a real competitor on plausibility. If its crossed score $D$ matches our compact model's, the specialization claim fails, and the experiment reports exactly that.

Why not just fine-tune a two-head regressor and attribute afterwards? It may well match our MAE, and we expect it to. But EMOTRIGGER's analysis found SHAP saliency on EmoBERTa aligned 27–40% more with topic keyphrases than with true triggers. Post-hoc attribution therefore inherits the classifier's shortcuts. The comparison of $D$ settles whether this matters for the two-perspective case.

Why not just run FRIT as published? This is the most serious threat. Rewarding only diagonal effects could already yield selective spans. The single-perspective ablation tests this directly. If it reaches our $D$, the crossed term contributes nothing, and the contribution narrows to porting FRIT to affect.

Why not just distill teacher rationales into the small model? Distillation transfers reasoning skill, as the Flan-T5 results show (27.1% on GSM8K for the 11B student). But trace-distillation studies report students whose explanations come apart from their predictions. The activation-patching check measures whether that happens here.

Proposed Experiments

The first experiment is the main comparison on WRIME gap posts. It reports $D$, per-perspective comprehensiveness and MAE for all systems at both model sizes, with emotions analyzed separately because muting is concentrated in anger and trust.

The second is external validity. We evaluate zero-shot reader-side span F1 on EMOTRIGGER and ask whether crossed training lifts the compact model above the roughly 0.25 reported for open models of similar scale. That would be direct evidence that the method reduces trigger-label confusion rather than overfitting WRIME. On a few hundred WRIME gap posts, fresh reader annotators will also mark the span that drives their reaction. Writer spans cannot be validated this way, since the authors are unavailable, and we state that limitation openly.

The third is mechanistic agreement. Activation patching on a subset tests whether text-level $D$ agrees in rank with internal causal faithfulness.

The fourth measures the specialization cost. We test on SemEval-2025 Task 11 Track B intensity prediction for available languages, where the best systems reach Pearson correlations up to about 0.65 according to the task overview, and on a small general reasoning suite.

The decisive result is crossed training beating single-perspective FRIT on $D$ at matched MAE, while also improving EMOTRIGGER span F1. If either fails, the crossed signal is not doing the work claimed.