Learning Word Meanings From Examples

Abstract

This paper d esc ribes work in progress on a computer program that uses sy nt actic con straints to derive the m ea nings of verb s from an analys is of simple En gli sh example stories .* The central idea is an ex tension of Win ston 's (Winston 1975) p rog ram th at le arned the stru ct ura l d esc riptions of blocks world sc en es. In th e new re search, En gl is h ve rb s take th e pl ace of block s wo rld obj ec ts like ARCH and TOW ER, with fr amc·ba se d de sc riptions of ca u sa l rel ation ships se rving as th e structural desc riptions. Synt ac ti c constraints derived from the pars in g of story pl ots a rc u se d to drive an analogical matc hin g p roc edure. An alo g ical matching g iv es a wa y to compare desc ri p tion s of kn o wn word s to unknown word s. Th e "meaning" of a new ve rb is learned by ma tchin g pan of the ca usal network des cription of a story precis cont a inin g the unknown word to a set of such de sc riptions deri ved from simil ar sto rie s that contain only kn o wn words. The be st match for ges an ass ignment between o bje cts and relations such that the unknown verb is matched to a kn ow n ve rb. with the as~ i g nmcnt being guided by syntactic co nstraints. The cau sa l netwo: ·k surrounding the unknown item is th en use d as a sca ffolding to construct a ne tw o rk re pr es enting the usc of th e novel word in a pa rti c ul ar context. Wo rd s (a nd their assoc iated sto ri es) th at arc "b es t m atc hes" arc grouped together into a similarity network, ac conling to the match score.