Intent classification

Intent classification maps a user's message to the goal it expresses, drawn from a fixed set of intents: refund, cancel_subscription, reset_password, set_alarm. It is the first step of most chatbots, voice assistants, support routers and agent planners, and it is usually paired with slot or entity extraction.

Traditional approaches train a classifier on labelled utterances per intent. That works well but needs data per intent and retraining whenever the intent list changes. Zero-shot approaches, including LLM prompting and NLI-based classifiers, accept the intent list at request time.

Intent classification with Laya

In Laya an intent set is a choice question. Labels and optional descriptions are passed per request, so you can add or rename intents without retraining:

{"intent": {"type": "choice",
  "instructions": "What does the customer want in `message`?",
  "criteria": {"refund": "money returned or a duplicate charge reversed",
               "cancellation": "wants to cancel or downgrade",
               "other": "none of the other options fits"}}}

Short descriptions help; an explicit other option gives uncertain inputs somewhere to go.

Reported benchmarks

• MASSIVE (multilingual, the card uses 20 options): 0.783 on English with the English checkpoint; 0.451 across 13 other languages with the multilingual one.
• Banking77 (77 fine-grained banking intents): 0.425 at default settings, against Jev's 0.870 on 72 labels. With 77 options, each label gets only about 3–4 tokens of the head budget.

The practical rule: Laya handles tens of intents well; for 50+ fine-grained intents either raise the head budget when self-hosting or use a two-step coarse-to-fine design (pick a group, then an intent within it). See the intent detection use case and Laya vs embeddings kNN.