📄 Normal State Implicature

Summary

Green asks how a speaker can say something like "I'd like to see a parrot" in a pet shop and be understood as wanting a live parrot. In a china shop, the same words would instead suggest a porcelain one. A literal first-order translation of the request does not entail the richer reading, so the extra content must come from context and from what speaker and hearer believe. The paper names this inference normal state implicature and treats it as a class of Gricean conversational implicature.

It follows Hirschberg's method for scalar implicature. First it shows that the example meets her six conditions for an implicature. Then it states a formal convention that picks out every instance of the class. The convention says the hearer may enrich a request about an entity with the state that entity must be in for the speaker's plan. This holds when both parties mutually believe the speaker has a plan in which:

• the entity plays a role,
• being in that normal state is a precondition, and
• the utterance itself is a step.

The paper reports no corpus or experiments. Its contribution is a formal account that matters both for interpreting users and for generating cooperative language.

Key points

• Hirschberg's six conditions frame the analysis: the speaker intends Q, cooperativity is mutually believed, saying U in context licenses Q, and Q is cancelable, nondetachable and reinforceable. Two follow-ups provide the evidence: "...a dead one" cancels the reading, and "...a live one" reinforces it without sounding redundant.

• The convention (6) quantifies over classifications, states, actions and plans. It is built from these predicates: BMB, HAS-PLAN, PRECOND, NORMAL-STATE, STEP, REALIZE, REQUEST, IN-STATE and NORMAL-STATE-IMP.

• Plans follow STRIPS, with two extensions:

  • a list of entities that take part in the plan, either as instruments or as the goal;
  • preconditions understood as applicability conditions, following Carberry, meaning states the agent should not reasonably try to bring about.

• The paper separates three notions:

  • stereotype, e.g. birds are built to fly;
  • plan-independent normal state, e.g. a bird is alive and uninjured;
  • normal state with respect to a plan, e.g. a carrier pigeon must be able to fly.

  Only the third drives the implicature.

• States are defined through causal event chains in the lexicon. For example, live covers the span from birth until death, and ready-to-use covers an artifact from creation or repair until destruction. A dead parrot still counts as a parrot.

Interesting ideas

• I1. Default enrichment of a bare indefinite should come from the precondition of the inferred plan, not from a global stereotype; this keeps defaults context-dependent, so the same noun can be enriched differently in a pet shop and a china shop.
• I2. An implicature can be validated by two checks: whether an explicit denial cancels it, and whether stating it outright reinforces it without redundancy.
• I3. Defining a state as the interval between two event types gives each normal state a precise lifecycle. For example, live runs from birth to death and ready-to-use from creation or repair to destruction.
• I4. In generation, the speaker should omit qualifications the hearer would infer anyway. It must add them whenever the default reading would be false, in line with Joshi's revised Maxim of Quality.
• I5. A cooperative answer flags an abnormal state that would defeat the asker's inferred plan, as in "yes, there is a truck, but it's out of service."

For Mova

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

• I1 (frames): when prag recognises a request speech act with an inferred goal frame, it can attach the goal's applicability conditions on the role filler as an explicit, cancelable implicature with a traceable justification.
• I2 (learn): the cancel and reinforce tests can serve as negative controls when evaluating implicature rules, using templated "…, a dead one" versus "…, a live one" continuations to check that the rule fires and retracts correctly.
• I3 (world): entity states in the text world can be stored as intervals bounded by events in the timeline, so a mention of dying or breaking automatically ends the normal state and blocks the default.
• I4 (back): before verbalising an indefinite, the generator can check whether the entity in the meaning graph violates the plan's normal state and, if so, insert the blocking qualifier.
• I5 (answer): QA readers can go beyond a literal yes when the world model shows the referenced entity failing a precondition of the asker's evident goal.