Multi-Intent Spoken Language Understanding

272 concepts178 papers3 proposals

Traditional Spoken Language Understanding (SLU) systems operate under a simplifying constraint known as the single-intent assumption. In this paradigm, every user utterance is mapped to exactly one actionable goal. While this works for simple commands like "Set an alarm," it fails to capture the complexity of natural human speech. As noted by Gangadharaiah et al., real-world users frequently issue compound requests—such as "book a cab and find a hotel"—where multiple distinct goals coexist within a single sentence.

Overcoming this assumption requires a shift from global classification to granular, relational modeling. The challenge is not just identifying that multiple intents exist, but correctly disentangling which specific entities (slots) belong to which intent.

The Problem of Semantic Ambiguity

When an utterance contains multiple intents, the relationship between words and goals becomes highly complex. Standard models often treat intent as a global property of the entire sequence. However, Cai et al. argue that this is logically flawed: in a multi-intent sentence, a specific entity like a "song title" usually relates to only one of the expressed intents (e.g., "Play Music") rather than all of them.

Without explicit mapping, systems encounter several failure modes:

  • Intent Distraction: Tokens relevant to one goal act as semantic noise for others, as described by Cheng et al..
  • Slot Confusion: If a user mentions two different locations for two different tasks, a global model may struggle to assign the correct constraint to the correct goal.
  • Structural Complexity: Intents are often nested or expressed through implicit linguistic cues rather than clear conjunctions like "and," making simple segmentation difficult (Wu et al.).

From Utterance to Token-Level Detection

To resolve these ambiguities, recent research moves away from classifying the whole sentence at once. Qin et al. proposed a Stack-Propagation framework that introduces token-level intent detection. By predicting an intent label for every individual word, the model can capture fine-grained semantic boundaries. This information is then "piped" directly into the slot-filling process, ensuring that the system understands the local context of each entity.

This transition allows the model to treat the utterance as a sequence of potentially overlapping segments. Morbini et al. demonstrated that models trained on single-intent data can generalize to complex speech by learning to decompose these inputs into simpler, actionable segments.

Mutual Guidance and Graph Interactions

The most effective multi-intent models leverage the inherent dependency between intent detection and slot filling. Rather than treating them as independent tasks, researchers use mutual guidance to let each task inform the other.

Xing et al. developed the Co-guiding Net, which uses heterogeneous semantics-label graphs to model these interactions. By allowing slot predictions to refine intent detection and vice versa, the model can resolve conflicting labels that might appear in isolation. Similarly, Qin et al. introduced a Global-Locally Graph Interaction Network (GL-GIN) to capture both the broad context of the utterance and the specific local relationships between words, significantly increasing processing speed without sacrificing accuracy.

Generalizing to Unseen Goals

Multi-intent detection becomes even more difficult when some of the user's goals are "unknown" or not present in the training data. Wu et al. addressed this through LABAN (Label-Aware BERT Attention Network), which treats intent labels as semantic descriptions rather than arbitrary IDs. This allows the system to detect multiple intents—even new ones—by matching the input text to the meaning of the labels themselves.

For scenarios with very little data, Zhang et al. proposed Coarse-to-Fine Prototype Learning (CFPL). This approach uses a universal "teacher" model to provide broad semantic knowledge, which is then refined by specific examples to help the model distinguish between overlapping intent boundaries in low-resource settings.

By moving beyond the single-intent constraint, these frameworks enable conversational agents to handle the fluid, multi-faceted nature of human interaction, transforming them from simple command-executors into sophisticated dialogue partners.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 186 papers
  2. Version 1 · 2026-10-04 · 280 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_single_intent_assumption_2ead5a83,
  title = {single-intent assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-intent-assumption/rdr_2ead5a838a76497fa6585d916fffaf35},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_single_intent_assumption_2ead5a83_v2,
  title = {single-intent assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-intent-assumption/rdr_2ead5a838a76497fa6585d916fffaf35/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake