📄 Speaker's Intentions and Beliefs in Negative Imperatives

Summary

This short workshop paper asks which factors shape a speaker's communicative intentions. It argues that one key factor in instructions is the set of choices the speaker believes the hearer faces while acting. The test case is negative imperatives. Di Eugenio separates two classes:

• DONT: don't or negative polarity items.
• neg-TC: verbs like take care, be sure or be careful followed by a negated infinitive.

The data is a small corpus taken from two how-to books and some product containers. Semantically, the neg-TC form entails the DONT form, yet the two do different jobs:

• DONT fits when the hearer knows about a decision point but might deliberately pick a wrong option.
• neg-TC fits when the hearer might not notice the decision point at all. Such a point often shows up only through an unwanted side effect, and it is always tied to another action in the discourse.

A real but odd example, don't damage adjoining tiles, supports the claim. The analysis matters for instruction generation: it shows how deep task intentions could become surface forms, which earlier work (DSH93) did not explain.

Key points

• The corpus has 36 DONT and 27 neg-TC imperatives, drawn from two how-to books plus detergent and cosmetics containers.
• Take care not to do α entails don't do α, and terse instructions use only DONT forms.
• DONT can stand alone, as in fire warnings. It also states general rules or rules out an alternative to a recommended action, with both serving a shared third goal (scrub vs. vacuum to clean parquet).
• neg-TC appears in two cases:
  • α is a bad specialization of an underspecified action β (drilling "near a corner" but through the pattern line).
  • α is an unwanted effect of β (dripping paste onto the wall).
• The work builds on the author's earlier analysis of purpose clauses and on the deep/shallow intention levels of Delin, Scott and Hartley.

Interesting ideas

• I1. Two forms can be near-equivalent in truth conditions yet still encode different beliefs about the hearer, here deliberate choice versus oversight, which makes surface form a cue to inferred mental state.
• I2. A neg-TC directive is anchored to another action. The negated act is either a bad way of carrying out that action or a bad side effect of it, and this relation can be made explicit.
• I3. Separate two ways a prohibited act relates to a recommended one: a sibling alternative under a shared goal, or a narrower version of an underspecified action.
• I4. Felicity judgments make a useful test: DONT is odd when the negated outcome, like "damage", is not something the hearer would deliberately choose.
• I5. Infelicitous but genuinely attested examples give the analysis a natural negative control.

For Mova

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

• I1 (frames): prag could tag negative imperatives as DONT vs. neg-TC and record whether the speaker presumes the hearer might choose the act or merely overlook it.
• I2 (world): link the prohibited event to its parent action as "manner-specialization" or "side-effect" in the event graph, so answers can explain why it is forbidden.
• I3 (world): store goal–alternative and action–specialization edges in the text world to support "why not X?" questions on instructional texts.
• I4 (back): when generating instructions, choose take care not to for side-effect outcomes and plain don't for deliberate options, and check that choice against such cases.
• I5 (learn): add real-world mismatched negative imperatives as negative controls when evaluating the prag speech-act classifier.