📄 Semantic Interpretation of Pragmatic Clues: Connectives, Modal Verbs, and Indirect Speech Acts

Summary

The paper tackles a first-pass pragmatic analysis of single user utterances in a German financial consultation system, WISBER. The goal is to turn each utterance into user-model entries: what the user wants, what the user believes, and how these relate. The authors argue that modal verbs and connectives carry much of the speaker's attitude but had received little attention in AI and computational linguistics.

Utterances arrive in the semantic language IRS, which still holds uninterpreted features such as modals, hedges, connectives and tense. Rewrite rules then process them in stages:

• split compound sentences at connectives;
• turn idiomatic indirect questions into direct ones;
• disambiguate each modal into one or more propositional attitudes;
• link these attitudes according to the connective's class;
• add time information from tense and attitudes from sentence type.

The results are stored in A-Box contexts for wants and beliefs. The module is implemented in Interlisp and covers all German modal verbs in assertions and questions, a few connectives and the most frequent indirect questions. No quantitative evaluation is reported. The work matters because it shows that small grammatical cues can reveal a user's goals and even their misconceptions.

Key points

• Part of WISBER, a German consultation dialog system built by a consortium: Nixdorf, SCS, Siemens, and the Universities of Hamburg and Saarbrücken.
• The input representation is IRS. It allows only one- and two-place predicates, which map to concepts and roles of the terminological knowledge base QUIRK. An operator PROP lets a variable name a whole proposition.
• The user model sits inside the central A-Box. It has separate contexts for user wants, user beliefs, facts and mutual knowledge.
• Modal disambiguation draws on three cues:
  • the semantic class of the subject (e.g., ANIMATE, GENERIC, DEFINITE);
  • how the modal's tense relates to the time of the proposition;
  • whether the proposition is a state or an event. For example, sollen can mean "is supposed to" (a belief) or "should" (a want).
• Implemented in Interlisp. Coverage: every German modal in assertions and questions, connectives such as and, so that and because, and the commonest indirect questions.

Interesting ideas

• I1. Polite indirect questions ("I don't know whether X", "Could you tell me…") are rewritten at the semantic level into a canonical direct question, so later stages see only one form; this is safe because the system checks whether it can answer anyway.
• I2. When tense cannot settle a modal's reading, the system asks whether the proposition's object already sits in the speaker's want context from an earlier sentence, and if so files the new proposition as a want too.
• I3. The domain changes the reading: in a consultation, "Should I…?" asks for a recommendation from the advisor rather than about someone's wishes.
• I4. A purpose connective ("X so that Y") marks Y as the real goal and X as an assumed necessary condition; storing that condition lets the system spot a user's mistaken belief, such as relying on investment term instead of liquidity.
• I5. Rules are pairs of parameterized semantic patterns; when the right-hand side is instantiated, any unbound variable becomes a fresh, properly quantified variable.

For Mova

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

• I1 (frames): crate prag could normalize idiomatic indirect requests into a single question frame and keep a record of the original surface form, so the rewrite stays explainable.
• I2 (world): the text world could hold separate want and belief contexts per agent, which the prag stage would consult when choosing between epistemic and deontic readings of "should" or "must".
• I3 (frames): speech-act rules could take the genre or setting as a parameter, e.g. a fable versus a word problem, so the same modal question maps to different act types.
• I4 (answer): readers could store "means → goal" links from purpose clauses and check them against world knowledge, flagging false premises in questions.
• I5 (learn): self-induced repair rules could be written as pattern pairs over frame graphs that create fresh variables explicitly, which keeps every rewrite inspectable and reversible.