When someone says "I just found out I failed my board exam for the second time," the emotional weight sits less in "board exam" than in "for the second time." Empathetic dialogue systems now try to find the emotion cause, the event that provoked a feeling, and build their reply around it. But the tools they use to find that event often cut off its trailing qualifiers. Work on multi-word entity boundaries shows the same failure in other domains: models identify the core noun correctly but truncate or overextend the modifiers around it. My claim is that in emotional conversation these modifiers are disproportionately where the emotional meaning lives. A system that cuts them off ends up empathizing with the wrong thing.
The proposal therefore treats boundary ambiguity as information, not only as noise to be engineered away. I will build a cause extractor that outputs a calibrated probability distribution over candidate cause spans, so it can report which specific segment of the cause it is unsure about. A response policy then uses that report. When the boundary is confident, the system reflects the full cause back to the speaker. When it is contested, the system asks a short, targeted question about the uncertain segment. This is the "exploration" move that counseling theory prescribes, and it is used here exactly where the model's own boundary uncertainty says the situation is unclear. Put briefly: find out where the system is unsure of the cause's edges, and either ask about that part or show that it heard it.
What already exists and where this differs
The obvious combination, a better span extractor feeding an empathetic generator, is largely covered already. Recognizing Emotion Cause in Conversations introduced RECCON and framed conversational cause extraction as start/end span prediction, with SpanBERT reaching about 60% F1 on the DailyDialog portion. MuTEC pushed this to about 66.9% positive F1 by learning span extraction jointly with utterance-level entailment. Gao et al. fed token-level cause probabilities into a generator through soft gated attention. Its finding that soft gating beats hard gating is important here: it suggests generators already tolerate imprecise boundaries by keeping the whole context visible. ECC used a masked-word objective over heuristic cause tokens. Kim et al. inferred single cause words without supervision and reached 17.3% Top-1 recall on EMOCAUSE against 41.3% for humans.
None of these papers asks whether boundary errors matter for empathy specifically, or uses boundary uncertainty to drive behavior. All of them commit to a single cause estimate. The Gao et al. result constrains the design: if soft gating absorbs boundary errors, then a sharper extractor alone may change nothing downstream. That pushes the idea away from "more accurate spans" and toward "know which part of the span you don't know, and act on it."
The method draws on two sources. BENSC scores each candidate span as the product of a start probability, an end probability and a span-classification probability. That product naturally yields a distribution over overlapping spans, and its pruning keeps the number of candidates small, which matters for multi-turn histories. Towards Emotional Support Dialog Systems supplies the ESConv benchmark and its strategy labels, including Question, grounded in an exploration–comforting–action progression. In ESConv, the decision to ask a question is predicted from the dialogue as a strategy token. It is not tied to any representation of what the system does not yet understand.
Core Claim
The proposal makes two testable claims.
The first is diagnostic. In conversational emotion causes, the tokens that span extractors most often drop or wrongly include are post-modifiers, quantifiers, temporal and repetition markers, and coordinated second clauses. These are also the tokens that human raters judge most decisive for an appropriate empathetic reply. This mirrors the Direction 1 finding that end boundaries are hardest under complex post-modifiers. If the overlap between "boundary-error tokens" and "emotionally decisive tokens" is no larger than chance, the rest of the program is not worth doing, and I will report that.
The second is behavioral. Let $p(s \mid x)$ be the model's normalized score over candidate cause spans $s$ for a dialogue context $x$. For each token $t$, define its contested mass $c_t$ as the probability that $t$ lies inside the cause minus the probability that the most likely span contains it:
$$c_t = \Big|\sum_{s \ni t} p(s \mid x) - \mathbb{1}[t \in \hat{s}]\Big|,$$
where $\hat{s}$ is the highest-scoring span. Tokens with high $c_t$ form the contested segment. A response policy that reflects $\hat{s}$ when the contested mass is low, and asks about the contested segment when it is high, should cut cause misattribution in replies more than two alternatives. The first is a policy driven by the dialogue model's overall uncertainty. The second is an extractor that is simply more accurate. The policy should achieve this without making the conversation feel interrogative.
Research Program
The first component is data. RECCON spans have a single annotation, which cannot show where boundaries are genuinely ambiguous. I will re-annotate a subset of RECCON-DD causal utterances (on the order of a thousand) with three annotators each. Each annotator marks every span they would accept as the cause, and separately marks the tokens without which an empathetic reply would miss the point. This gives three things: a set of acceptable boundaries, a per-token agreement signal to calibrate $c_t$ against, and a "decisive token" layer for the diagnostic claim. A smaller parallel annotation on ESConv seeker turns lets me test the same effects in multi-turn support dialogue, where causes are spread across turns.
The second component is the extractor. A RoBERTa or SpanBERT encoder runs over the dialogue history with BENSC-style start, end and span heads, plus a multi-span output so that RECCON's Hybrid and multi-utterance causes are not forced into one span. The single-span limitation of pointer networks noted by Li et al. is exactly what this is meant to avoid. The extractor is trained with soft targets taken from the multi-annotator acceptable-span sets, so that $p(s \mid x)$ learns real disagreement instead of a single gold span. Calibration is measured directly: does $c_t$ predict annotator disagreement on token $t$?
The third component is the response policy. A frozen instruction-tuned language model generates the reply. It is conditioned on the cause span, the contested segment and a binary commit-or-ask decision, which a threshold on contested mass makes from validation data. In ask mode the prompt requires one question that names the uncertain segment, for example "Is it mainly that this is the second time?" In commit mode the reply must reflect the whole $\hat{s}$, modifiers included. Keeping the generator frozen and identical across conditions means any difference comes from what it is told about the cause, not from generator capacity.
Baselines
The naive reference point is the same frozen language model prompted with the dialogue alone and no cause information. It shows how much any explicit cause signal helps when a modern generator already reads the whole context.
The strongest span-extraction baseline is MuTEC's end-to-end model from Multi-Task Learning Framework for Extracting Emotion Cause Span and Entailment in Conversations, which reports about 66.9% positive F1 for causal span extraction on RECCON-DD with conversational context. The SpanBERT baseline from Recognizing Emotion Cause in Conversations, at about 60% F1, is a second reference. I will retrain both on the same RECCON-DD splits with the same encoder size and the same hyperparameter search budget as my extractor. Beating them on exact-match F1 is not the goal. The bar is to match them on F1 while clearly improving retention of annotator-marked decisive tokens and calibration of contested mass.
The strongest generation baseline that uses causes is the soft-gated generator of Gao et al., which reports a human empathy rating of 3.244 and emotion accuracy of 42.4% on EMPATHETICDIALOGUES. Because those numbers come from an older, smaller generator, a fair comparison reimplements soft gating as a cause-weighting signal for the same frozen language model used in my system. The comparison then reflects the cause signal, not the decoder. I will cite ECC, which reports 48.4% emotion accuracy and 16.23 perplexity, as the strongest cause-aware generator on that benchmark. It will not be rerun, because its training pipeline is not comparable to a frozen-generator setup.
For questioning behavior, the baseline is an ESConv-style strategy predictor, following Towards Emotional Support Dialog Systems, which decides when to use the Question strategy. I found no reported result on Lacuna for clarification-question targeting accuracy on ESConv, so this baseline's numbers will come from my own runs.
Two ablations isolate the key ingredient. In the first, global uncertainty triggers asking: the commit-or-ask decision comes from the language model's own entropy or self-reported confidence, and the question is generic. In the second, span uncertainty decides when to ask but the question is untargeted. If either matches the full system, the localized contested segment is not doing the work.
Why not just…?
Why not just use a better span extractor and let the generator handle the rest? This is the most serious threat, because soft gating's advantage over hard gating in Gao et al. shows that generators route around boundary errors when the full context is visible. The committed-span condition, with no ask mode, run with my extractor and with MuTEC, tests this directly. If more accurate spans alone remove most cause misattributions, the questioning policy adds little, and the contribution reduces to the diagnostic finding plus a calibrated extractor. I will report it that way.
Why not just let the language model decide when to ask a clarifying question? A large model can notice that a situation is vague. But global uncertainty says whether to ask, not what to ask. The evidence from Chun et al., where Llama-3.1-8B reached only 46.9 F1 on biomedical term boundaries with semantic prompting and 49.4 with syntactic retrieval, suggests these models are poor judges of exactly which modifier belongs to a span. The global-uncertainty ablation is the test. The proposal survives only if targeted questions reduce misattribution more than generic ones at a matched ask rate.
Why not just infer the situation with commonsense knowledge instead of locating exact text? CEM infers intents, needs and reactions with COMET. However, the ECC ablation reported in the file shows that removing the cause module hurts badly even with the commonsense encoder active. Commonsense inference also cannot tell "failed the exam" from "failed the exam for the second time," because the difference is in the text, not in general knowledge. I will still include a COMET-augmented prompt condition so this is measured rather than assumed.
Why not just annotate causes at the utterance level, as graph models do? DAM and ECQED pick whole cause utterances. Knowing which utterance holds the cause cannot tell the system which qualifier inside that utterance is ambiguous, so it cannot support targeted questions. The utterance-level condition passes the whole cause utterance to the generator, testing whether full-utterance conditioning already captures decisive modifiers.
Proposed Experiments
The first decisive check is the diagnostic overlap. On the re-annotated RECCON-DD subset, compare the tokens that MuTEC and SpanBERT get wrong at the boundary with annotator-marked decisive tokens, against a chance baseline matched for position and length. If the observed overlap does not clearly exceed chance, the program stops at a published negative result.
The second check is calibration. Does contested mass $c_t$ predict per-token annotator disagreement better than the language model's token-level confidence? This is measured as rank correlation and expected calibration error on held-out dialogues, with the same test on EMOCAUSE to check transfer beyond RECCON.
The third check is the end-to-end human study on held-out RECCON-DD dialogues and ESConv seeker turns, with raters blind to condition. Raters judge empathy, specificity and, most importantly, whether the reply misattributes the cause, meaning it responds to the core event while ignoring the qualifier that matters. Because over-asking is a real cost and the commit-or-ask threshold could be tuned to inflate scores, all questioning conditions are compared at matched ask rates. Raters also judge whether the conversation felt like an interrogation.
The success criterion is this: compared with the strongest committed-span system and with the global-uncertainty ablation, at the same ask rate, the full system produces significantly fewer cause misattributions, with no drop in empathy ratings. Failure on the third check after success on the first two would still be informative. It would mean boundary ambiguity is real and emotionally loaded, but that current generators cannot use it, and that would point future work toward training rather than prompting.
References
- Recognizing Emotion Cause in Conversations
- Multi-Task Learning Framework for Extracting Emotion Cause Span and Entailment in Conversations
- Improving Empathetic Response Generation by Recognizing Emotion Cause in Conversations
- ECC: Synergizing Emotion, Cause and Commonsense for Empathetic Dialogue Generation
- Perspective-taking and Pragmatics for Generating Empathetic Responses Focused on Emotion Causes
- Boundary Enhanced Neural Span Classification for Nested Named Entity Recognition
- Towards Emotional Support Dialog Systems
- Boundary Detection with BERT for Span-level Emotion Cause Analysis
- Enhancing Automatic Term Extraction with Large Language Models via Syntactic Retrieval
- CEM: Commonsense-Aware Empathetic Response Generation
- Discourse-Aware Emotion Cause Extraction in Conversations
- ECQED: Emotion-Cause Quadruple Extraction in Dialogs