Limit 02 of 11

Non-conventional implicature

Detection is achievable. Resolution is not, and it is open for everyone.

Detection
Achievable
Resolution
Open problem
Scope
Everyone
Sound
Yes, flags

The problem

What actually goes wrong

Conventional implicature is carried by the words: "but" signals contrast whatever the context. Non-conventional implicature is carried by the situation. "It is cold in here" can be a report of temperature or a request to close a window, and nothing in the sentence distinguishes them.

Resolving it requires a model of what the speaker wants, in this room, at this moment. Ackren has a grammar and an authored world, not a theory of the person in front of it.

The impact

Who this affects, and how much

In practice this shows up as literal-mindedness: an answer that is correct about the proposition and wrong about the request. In a regulated intake conversation that is usually recoverable, because the next turn corrects it.

The severity depends entirely on whether acting on the literal reading is harmful. In pharmacovigilance intake it is a nuisance. In a device that acts on the utterance it would not be, which is why we are not selling into that case yet.

What it means

The honest reading

It means the engine is a good literal listener and a poor mind-reader, and it will stay that way. What it does not mean is that this is a deterministic-systems problem. Transformer models do not resolve non-conventional implicature reliably either.

The difference is disclosure. When Ackren cannot resolve an implicature it raises a flag that appears in the derivation record. A statistical model produces its best guess with the same confidence it produces everything else, and the failure is invisible in the log. Our limit is worse-looking and better-behaved.

Stated position

This limit is published before anyone else finds it. It is one of eleven rows in the failure accounting, and it will not be quietly removed from this page if it becomes inconvenient.

Roadmap

What we are doing about it

  • 01

    Now

    Detection of the ambiguous case, with a flag in the record. The engine can tell you that a reading was underdetermined even when it cannot tell you which reading was meant.

  • 02

    Next

    Clarification turns in the conversational core: asking rather than guessing, which is available to us precisely because we know when we do not know.

  • 03

    Later

    Serena, which reads stance and affect, narrows a subset of these cases without resolving them. A request delivered as a complaint has a shape.

  • 04

    Never

    Claiming this as solved. It is open for the whole field, and any vendor claiming resolution is reporting benchmark performance on conventional cases.