The TICC: Parsing Interesting Text.

TIC : PARSING INTERESTING TEXT. David Allport School o/Cognitive Sciences University of Sussex, Falmer, Brighton BN1 9QN davida%ulc ,,c.mssex.cvaxa@cs -clac ad¢ ABSTRACt This paper gives an overview of the natural language problems addressed in the Traffic Information CoHator/Condenscr (TICC) pro- jeer, and describes in some detail the "interesting-corner parser" used in the TICC's Natural Language Summariser. The TICC is designed to take free text input describing local traffic incidents, and automatically out- put local traffic information broadcasts for motorists in appropriate geographical areas. The "interesting-corner parser uses both syn- tactic and semantic information, represented as features in a unification-based grammar, to guide its bi-directional search for significant phrasal groups. 1. INTRODUCTION The overall goal of the TICC project is to show the potential benefits of automati- cally broadcasting local traffic information. Our target system, dealing with traffic incidents in the Sussex area, is to be com- pleted by September 1989. The project forms part of the Alvey Mobile Information Sys- tems large-scale Demonstrator. The Natural Language Summariser com- ponent of this system is being developed at Sussex University. Its function is to accept a series of free text messages describing traffic incidents, and to extract from these messages any information that might be relevant for broadcast to other motorists. The Natural Language Summariser is designed to work in a restricted domain, and only needs to solve a subset of the problems of text understanding. The TICC's output messages are short and very simple assem- blies of canned text, posing no significant natural language generation problems. Our main concern is that the messages should be useful to motorists, i.e that they be reliable indications of the state of the roads at the time they are broadcast. Programs such as METEO [Chevalier et al. 1978] have demonstrated that in a res- tricted domain with a restricted sub-language, automatic information broadcasts can be use- fuL Programs such as FRUMP [De Jong 1979, De Jong 1982] have also demonstrated that expectation-driven analysers can often successfully capture the gist of free text. However, the top-down depth-first confirmation of expectations based on sketchy scripts, ignoring most of the input structure, can lead to serious misinterpretations [Ries- beck 82]. Our concern for accuracy of interpretation has led us to a processing stra- tegy in which the Natural Language Sum- mariser analyses the input text at a far greater level of detail than is given in the output messages, so the system "knows more" about the traffic incidents it is describing than it says in its broadcasts. Our parser uses both syntactic and semantic information to guide its search for phrases in the input that might be directly or indirectly relevant to motorists, and explores alternative possible interpretations bottom-up using an active chart [Farley 1970, Kay 1973]. This is an ongoing research project, and we do not claim to have solved all the problems involved in developing a successful system yet. The current paper considers the particular natural language problems we are addressing and describes the "interesting- corner …

Content truncated to 500 words