Cn Yur Cmputr Raed Ths?

CN YUR CMPUTR RAED THS? Linda G. Means Computer Science Department General Motors Research Laboratories Warren, Michigan ,18090 ABSTRACT This paper describes strategies for automatic recognition of unknown variants of known words in a natural language processing system. The types of lexical variants which are detectable include in- flexional aberrations, ad hoc abbreviations and spelling/typographical errors. This technique is independent of any particular grammar or parsing formalism, and can be implemented as a lexical lookup routine which heuristically prunes and or- ders the list of possible fixes found in the lexicon, then allowing the parser to treat the list of candi- dates as a set of multiple meanings for a polyse- mous word. 1 INTRODUCTION This paper describes a technique for automatic recognition of unknown variants of known words in a natural language processing system. ~Known word" refers here to a word which is in the lexi- con. The types of lexical variants which are de- tectable include inflexional aberrations, ad hoc abbreviations and spelling/typographical errors. The strategies presented here have been imple- mented fully in an English database query system and play a crucial role in a text-understandlng sys- tem which is in the early stages of design. This technique, however, is independent of any par- ticular grammar or parsing formalism, and can be implemented as a lexical lookup routine which heuristically prunes and orders the list of possi- ble fixes found in the lexicon. First, a context-free plausibility assessment is based on a comparison of the structure of each candidate fix with that of the unknown word, and determines the order in which fixes will be considered by the parser. Then, the parsing process can choose among the candidate fixes in the same way that it tests multiple mean- ings of polysemous words for a good syntactic and semantic fit. The use of heuristics to identify the most plausible fixes for a hypothesized ad hoc ab- breviation or spelling error will be the focus of this paper. Unknown words have traditionally been handled by natural language processing systems in the fol- lowing ways: 1. Query the user for a replacement, possibly of- fering a menu of spelling corrections. This strategy will allow correction of misspelled words as well as correctly spelled words which are not in the lexicon, and generally ensures an accurate interpretation by the computer. However, continued interaction of this sort may prove frustrating to a poor typist, and is, of course, unsuitable for a non-interactive natural language processor. 2. Enter into a dialogue with the user to provide a definition for a new word. This strategy requires a lexicon interface based on a met- alanguage which would specify grammatical properties for a word without necessitating an inordinate degree of linguistic sophistication or knowledge of the database on the part of the end user. Although various attempts have been made to design such interfaces I, many outstanding research issues remain, and this approach too requires an interactive environ- ment. 3. Try to infer syntactic and/or semantic …

Content truncated to 500 words