C7.14.2Free-form speech and parse ambiguitydesignresearch

Free expression lets people describe intent in natural language, at a large rise in parse ambiguity

Aliases: NLU · intent ambiguity · open-domain voice

What it is

Free expression lets people describe intent in natural language without a fixed pattern. The back end guesses what to do with intent classification, slot filling, or a large-model alignment. Coverage widens and parse ambiguity jumps: the same sentence could address a calendar, a message, or search; negations and conditions are easier to drop. What you buy is not having to memorise phrases.

Why it happens

Open decoding first emits text, then (or at the same time) parses semantics. The hypothesis space is the whole language; one WER substitution can twist the intent onto another skill. Ellipsis, anaphora, and indirect speech acts (“it’s a bit cold in here” as close the window or set the A/C) have no grammar graph to sit on. Language models prefer fluency and may fill slots the user never said. Relative to a command grammar, rejects there are explicit; here the system often produces an execution that sounds reasonable and is not what was meant. Intent error can occur at very low WER because the sentence is legal and the skill is wrong.

Studying it

On the same batch of natural commands, compare grammar-constrained decoding with open NLU on intent accuracy, slot F1, and “reasonable but wrong” executions. Let untrained people improvise. Label ambiguous sentences with more than one legal reading. Do not use only the perfect questions from an advertisement script.

Where it stops holding

In a very narrow domain, priors can crush free-form ambiguity until it looks like a grammar. Chat with no executable intent pays for ambiguity with a wrong answer, not a wrong action. High-cost actions on free expression need a confirm, or ambiguity becomes an incident. Irony and rhetoric scramble parsing further; polite lab requests will not show it.

Applying it

  • For free expression that can have side effects, read back the understanding before executing—especially delete, pay, and send-to-someone.
  • Log intent mix and harvest frequent stable intents into a grammar, shrinking ambiguity on those paths.
  • Put sentences with two legal readings in acceptance tests and confirm the system asks rather than silently picking one.

Related

  • Same group: C7.14.1 Command grammars require fixed phrasings: more reliable recognition, at a memory cost · C7.14.3 When only a command grammar is supported and the phrasings are unstated, people trial-and-error · C7.14.4 Command grammar and free expression can coexist in layers
  • Adjacent: C7.03 Types of Recognition Errors · C7.04 Accent, Dialect, and Code-Switching
  • Search: NLU · intent ambiguity · free-form speech

Cards in the same group

Quick Actions

Share

Share this page

ios_share

https://hci.top/en/handbook/C7.14.2