A Terminological Simplification Transformation for Natural Language Question-Answering Systems

A Terminological Simplification Transformation for Natural Language Question-Answering Systems David G. Stallard BBN Laboratories Inc. I0 Moulton St. Cambridge, MA. 02238 Abstract A new method is presented for simplifying the logical expressions used to represent utterance meaning in a natural language system. 1 This simplification method utilizes the encoded knowledge and the limited inference-making capability of a taxonomic knowledge representation system to reduce the constituent structure of logical expressions. The specific application is to the problem of mapping expressions of the meaning representation language to a database language capable of retrieving actual responses. Particular account is taken of the model-theoretic aspects of this problem. 1. Introduction A common and useful strategy for constructing natural language interface systems is to divide the processing of an utterance into two major stages: the first mapping the utterance to a logical expression representing its "meaning" and the second producing from this logical expression the appropriate response. The second stage is not neccesarily trivial: the difficulty of its design is signifigantly affected by the complexity and generalness of the logical expressions it has to deal with. If this issue is not faced squarely, it may affect choices made elsewhere in the system. Indeed, a need to restrict the form of the meaning representation can be at odds with particular approaches towards producing it - as for example the "compositional" approach, which does not seek to control expression complexity by giving interpretations for whole phrasal patterns, but simply combines together the meaning of individual words in a manner appropriate to the syntax of the utterance. Such a conflict is certainly not desirable: we want to have freedom of linguistic action as well as to be able to obtain correct responses to utterances. This paper treats in detail the particular manifestation of these issues for natural-language systems which serve as interfaces to a database: the problems that arise in a module which maps the meaning representation to a second logical language for expressing actual database queries. A module performing such a mapping is a component of such question-answering systems as TEAM [4], PHLIQA1 [7] and IRUS [1]. As an example of difficulties which may be encountered, consider the question "Was the patient's mother a diabetic?" whose logical representation must be mapped onto a particular boolean field which encodes for each patient whether or not this complex property is true. Any variation on this question which a compositional semantics might also handle, such as "Was diabetes a disease the patient's mother suffered from?", would result in a semantically equivalent but very different-looking logical expression; this different expression would also have to be mapped to this field. How to deal with these and many other possible variants, without making the mapping process excessively complex, is clearly a problem. The solution which this paper presents is a new level of processing, intermediate between the other two: a novel simplification transformation which is performed on the result of semantic interpretation before the attempt is made to map it to the database. …

Content truncated to 500 words