Analogical inference over a common sense database

Abstract

This paper shows that by applying analogical inference techniques to a large natural language common sense database, we can generate new, plausible common sense facts. Two systems that do this are described. Being able to generate new facts in this manner allows quick augmentation of the common sense database. The role of analogical inference in common sense reasoning has been discussed before (Carbonell 1983), but only recently have large common sense databases become publicly available. We used the Open Mind Common Sense (OMCS) database, which contains several hundred thousand English common sense statements (Singh 2002). The OMCS database is built by internet users who are prompted to enter common sense facts at a website. The reason we use the OMCS database instead of the CYC common sense database (Lenat and Guha 1994) is because this analogical inference work is concurrently being integrated into OMCS's data collection mechanism. The idea is that an internet user can enter a fact at the OMCS website, then the system uses techniques described by this paper to respond with 30 plausible deductions based off the fact. Then, the user decides whether each deduction is true or false. True deductions become additional database facts, while false ones become negative expertise. Doing this allows for much faster data input by users. To carry out the analogical inference, we first represent the OMCS database as a set of concepts and a set of relations. A concept is a noun phrase or an adjective. A relation looks like “A ? is for playing ?”. Each original sentence becomes a relation after its concepts are replaced by question marks.