Learning theories from text

Learning theories from text Maria LIAKATA and Stephen PULMAN Centre for Linguistics and Philology Walton Street University of Oxford U.K., maria.liakata@clg.ox.ac.uk, stephen.pulman@clg.ox.ac.uk Abstract In this paper we describe a method of automati- cally learning domain theories from parsed cor- pora of sentences from the relevant domain and use FSA techniques for the graphical represen- tation of such a theory. By a ‘domain theory’ we mean a collection of facts and generalisations or rules which capture what commonly happens (or does not happen) in some domain of interest. As language users, we implicitly draw on such theories in various disambiguation tasks, such as anaphora resolution and prepositional phrase attachment, and formal encodings of domain theories can be used for this purpose in natural language processing. They may also be objects of interest in their own right, that is, as the out- put of a knowledge discovery process. The ap- proach is generizable to different domains pro- vided it is possible to get logical forms for the text in the domain. 1 Introduction It is an old observation that in order to choose the correct reading of an ambiguous sentence, we need a great deal of knowledge about the world. How- ever, the observation that disambiguation decisions depend on knowledge of the world can be made to cut both ways: just as we need a lot of knowledge of the world to make disambiguation decisions, so a given disambiguation decision can be interpreted as telling us a lot about the way we view the structure of the world. Our method for inducing domain theo- ries relies on this inversion, since in the general case it is a much easier job to disambiguate sentences than to directly encode the theory that we are draw- ing on in so doing. Our strategy for trying to build a domain theory is to try to capitalise on the informa- tion that is tacitly contained in those disambiguation decisions. 2 Some background (Pulman, 2000) showed that it was possible to learn a simple domain theory from a disambiguated cor- pus: a subset of the ATIS (air travel information ser- vice) corpus (Doddington and Godfrey, 1990). Am- biguous sentences were annotated as shown to indi- cate the preferred reading: [i,would,like, [the,cheapest,flight, from,washington,to,atlanta] ] [do,they,[serve,a,meal],on, [the,flight,from,san_francisco,to,atlanta]] [i,would,like, [a,flight,from,boston, to,san_francisco, ][that,leaves,before,’8:00’] ] The ‘good’ and the ‘bad’ parses were used to pro- duce simplified first order logical forms represent- ing the semantic content of the various readings of the sentences. The ‘good’ readings were used as positive evidence, and the ‘bad’ readings (or more accurately, the bad parts of some of the readings) were used as negative evidence. Next a particu- lar Inductive Logic Programming algorithm, Pro- gol (Muggleton, 1995), was used to learn a theory of prepositional relations in this domain: i.e. what kinds of entities can be in these relations, and which cannot: on(+any,+any) from(+any,+any) to(+any,+any) at(+any,+any) The +any declaration says that there are no prior assumptions about sortal restrictions on these predicates. Among others generalisations like the …

Content truncated to 500 words