📄 A computational account of comparative implicatures for a spoken dialogue agent

Summary

Benotti and Traum study how a dialogue agent can recover the unstated meaning of comparative sentences. A sentence like "downtown is safer than the market" literally orders two items on a scale. When the scale has few possible values, it can also suggest where each item sits on that scale. Classical degree semantics assumes the ordering relation is already known. In dialogue, however, knowledge of the domain is partial and gets negotiated, so the authors argue that relevant scales come from the participants' goals and the task model. They draw on Gricean and scalar implicature and on clarification requests, and they extend an existing information-state dialogue framework with rules that turn a comparative over a binary attribute into implicated values for both compared items. They test this in the SASO-EN negotiation scenario, where virtual agents debate where to put a clinic. With the rules, the agents grasp arguments they previously missed and can challenge an implicated claim directly. This matters because agents need to know why a comparative was said, not just its literal content.

Key points

• The test domain is SASO-EN. A human Captain negotiates a clinic location with two virtual agents, the Doctor and the Elder.
• The domain has over 60 states relevant to planning and negotiation. It uses 11 attributes, of which 7 are binary scales and 4 are non-scalar.
• A comparative arrives as a semantic frame (attribute "safer", object downtown, value market). The rules add two further assertions: safety of downtown = yes and safety of the market = no.
• The paper cites a corpus study of clarification requests. Problems with referring expressions cause 51.74% of them, and unstated task requirements cause 22.17%.
• Implicatures interact with context in three ways:
  • If the implicated content is already present, it is bound and causes no update.
  • If another speaker's comparative implicates a claim that is already in context, that counts as accepting the claim.
  • If the claim conflicts with what the participants already share, the agent drops the implicature.
• Known flaw: with chained comparatives, which implicature survives depends on the order of the utterances.

Interesting ideas

• I1. Scales and their granularity come from the task model and goals, not from the lexicon, so only distinctions tied to the agent's goals produce implicatures.
• I2. When a scale is binary, "X is more P than Y" can be expanded into two explicit state assertions (X is P, Y is not P), each kept separate from the literal content.
• I3. Clarification questions serve as a diagnostic: content a listener could sensibly ask about ("why is the market not safe?") is exactly the implicated material.
• I4. Implicated content is either bound to earlier context, treated as acceptance of another speaker's claim, or discarded when it clashes with the common ground.
• I5. The order-dependence bug with chained comparatives is a good negative control showing that greedy inference needs global consistency checking.

For Mova

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

• I1 (frames): prag could decide whether a comparative or scalar term triggers an implicature by checking which attributes matter for the current story's goals.
• I2 (world): the world builder could add these derived states marked as cancellable inferences, so later contradicting text can drop them without breaking the literal comparison.
• I3 (learn): evaluation could generate "why/are you sure" probes against inferred facts, checking that the system can point to the implicature and its trigger as an explanation.
• I4 (frames): coreference and speech-act handling could use the same three-way check to label a sentence as agreeing with, repeating, or ignoring an earlier claim.
• I5 (answer): QA tests on texts with chained comparisons (A > B, B > C) would show whether the world model reasons over the whole ordering rather than the last sentence.