📄 Indirect Responses to Loaded Questions

Summary

Kaplan addresses a practical gap in natural language (NL) database interfaces. Casual users often hold mistaken beliefs about what a database contains, and these beliefs surface implicitly in how they phrase questions. A system that only returns literal answers ("None", "Zero") quietly confirms such errors. The paper calls this "stonewalling."

Drawing on conversational cooperation, the author separates two notions. A presupposition is entailed by every direct answer to a question. A broader presumption is entailed by all direct answers but one. Asking a question signals that the asker finds its presumptions plausible.

The central argument is that many such presumptions can be recovered from the linguistic form of a query. This kind of language-driven inference needs no large hand-built stores of domain knowledge, unlike domain-driven inference. The approach is implemented in CO-OP, a cooperative query system whose only knowledge sources are a standard CODASYL database and a lexicon. When a presumption fails, CO-OP replies by correcting the false belief instead of giving an empty answer. The work shows that cooperative, domain-independent behaviour is computationally feasible.

Key points

• Two types of loaded question. A presupposition holds under all direct answers. A presumption holds under all but one, as in "Did Sandy pass the prelims?" presuming Sandy sat them. Asking a question built on a belief the asker knows to be false would break the cooperative convention.
• Corrective indirect responses. When a user asks for the knife-selling departments that also stock blade sharpeners, replying "No departments sell knives" is more cooperative than "None". A CIS500 dialogue illustrates the failure mode: a literal system never reveals that the course was not offered in Spring '77.
• CO-OP system. It uses only a CODASYL database and a lexicon as world knowledge. Queries are represented in MQL, whose subgraphs are partially ordered. When several corrections apply, CO-OP chooses the minimal failing subgraph.
• Further cooperative behaviours. These were shown in a demo on a project/sponsor database (NCAR, Nasa Headquarters). CO-OP adds supportive extra information to answers. It offers suggestive alternatives when an answer is negative and no presumption has failed. It also tolerates sloppy reference, for example assuming a single contact where several exist.
• Main claim. Domain-driven inference must be complemented by language-driven inference. The latter yields efficient, domain-transparent solutions to problems previously thought to require deep world knowledge.