Voice assistants and chat agents receive requests that carry facts with them: "book a flight to the city where JFK was assassinated and find a hotel near the museum about it." Two research areas each see half of this sentence. Language-understanding systems split it into goals, and they now handle several goals per utterance. Fact-checkers verify claims, but they assume the input is a standalone declarative statement. Neither area handles a factual premise that is embedded inside a command and decides what the command means. If the premise is false ("the city where the CDC was founded in 1946", when the founding facts are wrong), the assistant either acts on a fabricated slot value or silently "corrects" it without saying so. In both cases the user gets no explanation.
This note proposes to treat embedded premises as first-class objects. The system would jointly decide which words belong to which goal, which words assert or presuppose a checkable fact, whether that fact holds, and which slot value the fact determines. It would then report this back as an explanation tied to specific spans: "I read 'the city where…' as the destination; that premise is supported by this evidence, so the destination is Dallas." The bet is that, for utterances whose task arguments are defined by facts, verification and slot filling must inform each other. A segment-then-verify pipeline will either fail at that point or, if it succeeds, show that the joint machinery is unnecessary. The work is designed so that either outcome is a clean result.
Core Claim
Two assumptions are linked here. The single-intent assumption in spoken language understanding maps each utterance to exactly one goal label. Fact-checking makes a parallel assumption: one input, one claim, one verdict. Multi-intent understanding has already relaxed the first assumption. Joint models predict several goals and the task arguments ("slots", such as destination city or date) of each, and they let slot and goal predictions refine one another (Co-guiding Net; Stack-Propagation). Explainable fact-checking has partly relaxed the second through decomposition. Claims are broken into phrases with their own latent verdicts in LOREN, or into iteratively extracted atomic facts in AFEV.
The specific claim is this. For premise-defined slots, where a slot's value is given by a referring description rather than a name, the correct slot value is only determinable after verification. The correct verification scope is only determinable after goal segmentation. A model that passes information in both directions should therefore beat any one-directional pipeline on exactly these cases. Elsewhere it should be no better.
A second, cheaper hypothesis comes along with the first. Goal boundaries are a better unit for decomposing mixed utterances than maximal atomic facts. Work on decomposition shows that splitting too finely strips context and hurts verification (Decomposition Dilemmas), and that verifiers have a preferred granularity (Optimizing Decomposition). Goal segments are a linguistically motivated, coarser granularity that nobody has tested.
How This Differs From Existing Work
The closest work falls into three groups.
Dialogue fact-checking in DialFact detects verifiable turns and verifies them. It treats each turn as either a claim or chatter, so a turn that is part command and part premise is outside its scope. Its finding that dialogue context raises retrieval recall from 60.8% to 75.0% motivates using goal context for retrieval here.
Adaptive decomposition is represented by AFEV and DYDECOMP (Optimizing Decomposition). AFEV extracts atomic facts conditioned on earlier verdicts. DYDECOMP learns with reinforcement learning when to split a claim, using verifier confidence as its reward. Both operate on declarative text (Wikipedia, news, biographies). Neither has a notion of a task goal, and neither lets a verdict change a downstream action argument.
Multi-intent understanding (Co-guiding Net; GL-GIN; LABAN) disentangles goals and slots but never asks whether a slot's descriptive content is true.
What the proposal reuses is specific. It takes per-token goal labeling from Stack-Propagation. It takes two-stage mutual refinement between slots and goals from Co-guiding Net. It takes LOREN's three-valued latent verdict per span (Supported, Refuted, Not Enough Info), with soft-logic aggregation. The new element is an edge that none of these papers contain: from a premise span's verdict and evidence back into slot resolution and goal scoping.
Two known failure modes make this edge plausible: words from one goal distracting the prediction of another (Cheng et al.), and entities being bound to the wrong goal (Cai et al.). A fact-carrying relative clause is precisely the kind of long, entity-dense span that causes both.
Research Program
The first deliverable is a benchmark, because none exists. Call it premise-carrying multi-intent requests. It starts from MixATIS and MixSNIPS, the multi-intent benchmarks built by concatenating single-goal utterances. A subset of slot values (cities, airlines, artists, restaurants) is replaced with referring descriptions built from FEVER and HOVER claims about that entity. True and false versions are paired, so that "flights to Dallas" becomes "flights to the city where JFK was assassinated in 1963" or a minimally edited false variant.
Each item is labeled with three things: the goals and slots, the premise spans and the goal/slot each one scopes, and the premise verdict with evidence sentences inherited from the source claim. It also gets a resolution label. This label records whether the premise uniquely determines the slot (premise-defined), merely decorates an already-named slot ("Dallas, where JFK…"), or is a free-standing claim with no task role.
Template artifacts are the obvious threat to validity. The benchmark will therefore include a human-written test portion in which annotators paraphrase items without explicit conjunctions. This addresses the brittleness to non-canonical syntax that LABAN documents. Results on templates alone will not be reported as headline numbers. DialFact turns serve as an out-of-distribution test of premise detection on real conversation.
The model is a shared encoder with per-token heads for goal, slot tag, and premise tag, in the Stack-Propagation style. A first stage produces draft goals, slots, and premise spans. Each premise span is turned into an evidence query that includes its scoped goal segment rather than the bare clause, so the query keeps the context that decomposition otherwise loses. Retrieval follows AFEV's bi-encoder-plus-reranker recipe, and a verifier assigns the span a latent verdict $z$.
A second stage, adapted from Co-guiding Net's graph refinement, adds the verdict and the top evidence entity as nodes connected to the slot and goal nodes. Slot values can then be resolved from evidence ("the city where…" → Dallas), and refuted premises can trigger a flag rather than a guess. LOREN-style soft logic ties span verdicts to an utterance-level judgment of whether the request can be executed as stated.
The explanation is the trace itself: spans, scopes, verdicts, evidence, and the resolved slot. Faithfulness is measured as agreement between the aggregated span verdicts and the final decision, which is the metric LOREN reports above 96% on FEVER.
Baselines
The naive reference applies the single-intent assumption on both sides. The whole utterance is sent once to a standard claim verifier: the same retriever and verifier used inside our model, with no segmentation. A standard multi-intent model separately fills slots from the surface string. This shows how much is lost by not separating goals from premises. Any proposed system must clearly beat it on premise verdict accuracy and on slot accuracy for premise-defined items.
The strongest existing structured competitor is a two-stage composition of the best published components. Co-guiding Net parses goals and slots. It reports 51.3% overall accuracy on MixATIS (57.5% with RoBERTa) and 77.5% on MixSNIPS. Its slot spans are then handed to LOREN or AFEV for verification. LOREN reports a 72.93 FEVER score and 74.3 F1 on culprit-phrase identification. AFEV reports 78.87% label accuracy on HOVER. This pipeline has strong parts in both directions but no feedback from verdict to slot, so it isolates the claim directly. There is no reported result found on Lacuna for any system on premise-carrying multi-intent requests, so all numbers on the new benchmark will be produced here.
The strongest unstructured competitor is a frontier instruction-tuned language model. It is given few-shot prompts to segment the request, verify premises with the same retrieval corpus through tool calls, and emit the same structured trace. It also gets a self-critique pass, because Decomposition Dilemmas shows reflection raising verification F1 on BINGCHAT from 24.91 to 67.70. Leaving that step out would make the baseline a strawman.
Fairness rules apply to all baselines. Every system uses the same evidence corpus, the same retriever, and the same encoder size where trainable. Every system gets an equal hyperparameter search budget, measured in training runs, and an equal number of prompt-development iterations on a shared development split.
Two ablations isolate the key ingredient. The first removes the verdict-to-slot edges, leaving our model a jointly trained pipeline. The second keeps the model but decomposes at atomic-fact granularity, using the neo-Davidsonian prompting of Wanner et al., which scored 42.3 DecompScore versus about 32 for FActScore-style decomposition. A third comparison replaces fixed goal-boundary units with DYDECOMP's learned granularity.
Beating the baselines means two things. First, a statistically significant gain in end-to-end exact accuracy and slot accuracy on the premise-defined, human-written split. Second, no loss on decorative and free-standing items.
Why not just…?
Why not just prompt a strong language model to do everything? This is the most serious threat, and the proposal does not assume it fails. Existing evidence is suggestive but indirect. GPT-3.5 reaches only 64.9% as a planner on multi-hop EX-FEVER. Unreflected decomposition on BINGCHAT scored 24.91 F1. Prompted baselines in FACT5 left 78 of 150 real claims unverified. None of these involve task requests. The direct comparison on the human-written split settles the question.
If the prompted model matches the joint model on premise-defined items, the report will say so. The contribution then becomes the benchmark, the contamination metric, and the finding that structure is unnecessary. One gap would remain: a token-level model offers inspectable span scoping and lower latency, which matter for on-device assistants.
Why not just segment first and verify afterwards? Segmentation from single-goal data is known to work (Morbini et al.), and the Co-guiding Net + LOREN/AFEV pipeline is exactly this alternative run at full strength. The prediction is narrow. On decorative premises the pipeline should tie, because the slot is already named. On premise-defined slots it should lose, because the parser must commit to a slot value before anyone has looked up what the description refers to.
The no-feedback ablation tests the same prediction inside one model. If the gap appears only on the premise-defined stratum, the mechanism is confirmed. If it appears nowhere, the bidirectional edge is not worth its cost.
Why not just use a dialogue fact-checker such as Aug-WoW from DialFact? That system decides whether a whole turn is verifiable and verifies it as a unit. It has no output for "this clause scopes the destination slot of the second goal." Its own numbers show verification dropping from about 69% to 51% once evidence must be retrieved rather than given. We will run it as a turn-level verifier on our benchmark to measure how much a single verdict per turn misses. It cannot fill slots, though, so it competes only on the verification metrics.
Why not just parse with constituency trees, as LOREN does? Relative clauses are syntactically easy to find, so this is a fair cheap alternative for premise detection. Syntax does not say which goal a clause belongs to when conjunctions are absent, and it does not resolve the slot. We include a parser-based premise detector as a component swap and expect it to close most of the detection gap on templates but little of the scoping gap on human paraphrases.
Proposed Experiments
The headline metrics are three. MixATIS-style overall accuracy is exact match on all goals, slots, premise spans, and verdicts. Premise verdict accuracy and culprit-span F1 follow LOREN's protocol. Premise-contamination rate is the fraction of false-premise requests that a system executes with a slot value derived from the false premise without raising a flag; this is the failure that matters for deployed assistants. Faithfulness is reported as verdict agreement.
Every metric is broken down by the resolution label: premise-defined, decorative, and free-standing. All results are reported separately for template and human-written test items.
The decisive check is an interaction effect. The joint model's advantage over the pipeline and no-feedback ablation should be concentrated in the premise-defined stratum and near zero in the decorative stratum. A uniform gain would point to extra capacity rather than the proposed mechanism. No gain would falsify the claim.
The granularity hypothesis is tested by holding the verifier fixed and comparing whole-utterance, goal-segment, atomic-fact, and DYDECOMP-learned units. Verification accuracy and retrieval quality (NDCG@10, the metric QuanTemp++ uses) are measured for each. The prediction, following Decomposition Dilemmas, is that goal segments beat atomic facts on mixed requests.
A small user study modeled on Generating Fact Checking Briefs will check whether span-level traces help people catch false premises. That paper found 45% of workers relied on briefs without searching, so the study will also measure over-reliance when the trace is wrong.
References
- Co-guiding Net: Achieving Mutual Guidances between Multiple Intent Detection and Slot Filling via Heterogeneous Semantics-Label Graphs
- A Stack-Propagation Framework with Token-Level Intent Detection for Spoken Language Understanding
- LOREN: Logic-Regularized Reasoning for Interpretable Fact Verification
- Fact in fragments: Deconstructing complex claims via LLM-based atomic fact extraction and verification
- Decomposition Dilemmas: Does Claim Decomposition Boost or Burden Fact-Checking Performance?
- Optimizing Decomposition for Optimal Claim Verification
- DialFact: A Benchmark for Fact-Checking in Dialogue
- GL-GIN: Fast and Accurate Non-Autoregressive Model for Joint Multiple Intent Detection and Slot Filling
- A Label-Aware BERT Attention Network for Zero-Shot Multi-Intent Detection in Spoken Language Understanding
- A Scope Sensitive and Result Attentive Model for Multi-Intent Spoken Language Understanding
- Slim: Explicit Slot-Intent Mapping with Bert for Joint Multi-Intent Detection and Slot Filling
- A Closer Look at Claim Decomposition
- EX-FEVER: A Dataset for Multi-hop Explainable Fact Verification
- FACT5: A Novel Benchmark and Pipeline for Nuanced Fact-Checking of Complex Statements
- Joint Identification and Segmentation of Domain-Specific Dialogue Acts for Conversational Dialogue Systems
- A Benchmark for Open-Domain Numerical Fact-Checking Enhanced by Claim Decomposition
- Generating Fact Checking Briefs