Syntactic disambiguation using presupposition resolution

Syntactic disambiguation using presupposition resolution Alistair Knott and Peter Vlugter Department of Computer Science University of Otago alik/pvlugter@cs.otago.ac.nz Abstract If a sentence is ambiguous, it often hap- pens that the correct reading is the one which can most easily be incorporated into the discourse context. In this paper we present a simple method for imple- menting this intuition using the mecha- nism of presupposition resolution. The basic idea is that we can choose be- tween the alternative readings of an am- biguous sentence by picking the read- ing which has the greatest number of satistified presuppositions. We present two uses of the disambiguation algo- rithm in our bilingual human-machine dialogue system. 1 Introduction Syntactic ambiguity is a well-known problem for natural language interpretation systems, and it is one which becomes increasingly serious as the syntactic coverage of a system increases. A great deal of research in NLP has focussed on how to avoid or alleviate the problem, and several differ- ent (and reasonably complementary) approaches have been devised. Four broad classes of ap- proach can be distinguished. The oldest approach involves devising structural heuristics for disam- biguation; these are often psycholinguistically- inspired, such as the minimal attachment strat- egy of Ferreira and Clifton (1986). Another class of approaches rely on the deployment of world or situation knowledge to assess alternative inter- pretations of a sentence. These approaches re- quire interpretation systems which generate fairly rich semantic representations for sentences, along with large knowledge bases of facts about the do- main of interpretation, along with sophisticated inference mechanisms to operate on them. Limi- tations on the size of knowledge bases and on the speed of theorem-provers mean that this approach is often hard to scale to real applications. (How- ever, there are a number of promising approaches for tackling the scalability problem; see for in- stance Bos (2001), Hobbs (1993), Stone (1998).) But by far the dominant approach to syntactical ambiguity resolution at the moment involves the use of statistical techniques in one form or an- other. In particular, using statistical grammars which take into consideration lexical dependen- cies (see e.g. Collins (1996), Goodman (1998), Magerman (1995)) has proved a very sensitive way of capturing the kind of grammatical contexts that particular words or word combinations typi- cally appear in; this information can be effectively used to decide between alternative parses. However, there remain some species of ambi- guity which are hard to resolve using statistical parsers. It frequently happens that at least two of the alternative readings of a sentence are rel- atively frequently attested in the corpus used to train the parser. Some of the classical illustrations of syntactic ambiguity can be used to make the point. For instance, there are two readings of the following well-known sentence: (1) The man saw the girl with the telescope. In one reading, the PP with the telescope attaches to the NP the girl, while in the other, it attaches resolved sentence DRS Saliency list stack DRS Context DRS Global DRS text …

Content truncated to 500 words