A Clause Boundary Is Not a Composition: What Clause-Factorized Decoding Shows About Compositional Multi-Intent Detection, Why a Segmenter and a Cardinality Head Carry Most of the Reported Gap, and the Splits That Would Separate Them
(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Multi-intent detection asks a dialogue system to recover every request in an utterance such as "cancel my meeting and play some music". A 2026 study argues that the field has been measuring the wrong thing: on standard benchmarks the same intent combinations appear in training and test, so a model can succeed by memorising familiar sets. It builds a controlled benchmark with held-out intent pairs, shifted connectors, noisier wrappers, held-out clause templates and unseen triples, and reports that a decoder trained only on single-intent utterances, which splits each test utterance into clauses and labels them one at a time, beats whole-utterance models by up to 91 points of exact match. In this analysis the author reviews and synthesises published results to show that the premise is right and the attribution is too broad. The comparison moves four things at once: a hand-written segmenter that supplies the number of intents, an output head that cannot emit three intents because it never saw three, a training signal that differs between the systems, and an encoder that differs in kind. The anchor's own gold-boundary ablation shows that every error ClauseCompose still makes on the connector and wrapper splits is a segmentation error. A 2024 benchmark, on which a count-predicting model loses 35 to 43 points when the joins between requests become more varied while a threshold model trained on the same data loses 2.5 to 15, makes the count head the likelier source of the whole-utterance models' losses. Neither result isolates composition, and the paper argues that segmentation and cardinality are the untested alternatives. Published corpora built from or modelled on real requests show where clause factorisation's assumption fails: several intents inside one clause, arguments shared across intents, and omitted material in later requests. The paper consolidates the measured values in Table 1 and Figure 1, states an attribution procedure as Algorithm 1, and specifies six controls that would separate composition from segmentation, cardinality, supervision and pretraining.
Authors
- Pranay M. Mahendrakar (ORCID: https://orcid.org/0009-0003-7224-029X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24221453
- Primary Topic
- Speech and dialogue systems
- Type
- preprint