📄 Negotiating causal implicatures

Summary

The paper tackles causal implicatures: things a speaker leaves unsaid but expects the hearer to infer from cause-and-effect relations in a task domain. If a crown sits in a locked chest and someone asks for the crown, they are implicitly asking for the chest to be unlocked. The authors argue that such inferences are interactive and get negotiated. A hearer either fills the gap silently (internal bridging, or accommodation) or opens a clarification sub-dialogue (external bridging).

An earlier annotation of clarification requests in the SCARE corpus suggested that most implicatures made explicit by clarifications have a causal explanation. Building on this, they present Frolog, a text-adventure-style dialogue system. Frolog computes the missing steps with the off-the-shelf classical planner BLACKBOX:

• the goal is the conjunction of the preconditions of the player's command;
• plans are inferred over a common-ground knowledge base;
• plans are then checked against the true world state.

When no plan exists, Frolog asks a causally informed clarification question. Once the gap is resolved, the same kind of command can be bridged internally. The work is a proof of concept, not a quantitative evaluation. It matters as a first computational handle on relation implicatures, which are under-formalized.

Key points

• Frolog keeps two knowledge bases. The world KB holds complete, accurate game facts. The interaction KB holds the common ground. Plans are inferred on the interaction KB and verified on the world KB.
• BLACKBOX (Kautz and Selman, 1999) receives three inputs:
  • the initial state;
  • a goal made of all preconditions of the uttered command;
  • the full game action database, assumed mutually known.
• The empirical motivation is Benotti's (2009) analysis of annotated clarification requests drawn from the SCARE corpus of Stoia et al. (2008). A request for a required but unknown thematic-role filler is the second most frequent clarification type there.
• The design uses planning instead of plan recognition. It fills the local "micro-structure" of discourse, not the global plan, and is computationally cheaper than the Perrault/Allen tradition.
• Under "strict accommodation", executing an action adds both its effects and its preconditions to the common ground. Preconditions that are newly added get spoken aloud, e.g., asking why the player thinks a key will fit.

Interesting ideas

• I1. Treat an instruction's implicit content as a planning problem: the goal is the command's unmet preconditions, and the missing plan steps are the implicature. This yields a precise, inspectable account of what was left unsaid.
• I2. Keep separate stores for what is actually true and what is mutually established, inferring over the second and checking over the first. This cleanly separates speaker assumptions from reality.
• I3. When an utterance succeeds, record the preconditions it relied on as presupposed facts, not just its effects. This captures assumptions the speaker silently committed to.
• I4. A failed bridge should produce a clarification aimed at the specific missing precondition or role filler, not a generic "I don't understand". The failure point itself tells you what to ask.
• I5. Clarification requests in a corpus serve as evidence of where hearers' inferences broke down, giving an empirical signal for otherwise invisible implicatures.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): crate prag could derive causal implicatures by running a small planner over verb preconditions and effects, and output the bridging steps as explicit, explainable frames.
• I2 (world): the world of the text could hold a "narrator/ground truth" layer and per-character belief layers, so answers about what a character knew stay distinct from what happened.
• I3 (world): when a story event like "unlocked the chest with the key" is added, Mova could also assert "key fits chest", enriching the state for later questions.
• I4 (answer): when a reader cannot connect a question to the text world, it could report which precondition or role is missing, making the failure explainable and useful for debugging.
• I5 (learn): Mova could mine dialogues or QA errors for clarification-like failures and use them as negative controls when evaluating its implicature module.