Generating English Paraphrases From Formal Relational Calculus Expressions

GENERATING ENGLISH PARAPHRASES FROM FORMAL RELATIONAL CALCUq.US EX~RESSIONS A.N. De Roeck and B.G°T. Lowden University of Essex Wivenhoe Park Colchester -. United Kingdom O. ABSTRACT This paper discusses a system for producing English descriptions (or "paraphrases") of tile content of formal relational calculus (RC) formulae expressing a database (DB) query. It explains the underlying design motivations and describes a conceptual model and focus selection mechanism necessary for delivering coherent paraphrases. The general paraphrasing strategy is discussed, as are the notions of "desirable" paraphrase and "paraphrasable qnery". Two examples are included. The system was developed and implemented in Prolog at the University of Essex under a grant from ICL. I. INTRODUCTION. Querying databases (DB) is often problematic for casual users, who cannot be expected to master fully the art of expressing themselves in a query language. Much has been said about how natural language (NL) front ends (FE) would help them, but NLEEs create problems of their own. In mapping from an ambiguous NL to an unambiguous formal one they mast make a number of decisions regarding interpretation of the input which remain outside the user's control° This may introduce new misconceptions which cannot always be detected from the format of the retrieved information. [13[ One solution to the problem is to present the casual user with a NL description of what his query has been taken to mean. lle can then verify whether its interpretation corresponds to what he intended to ask. This paper describes such a descrlption generator, or "paraphraser" as we will call it. 2. DESIGN PRINCIPLES. The system is intended to work alongside NEL (formerly QPROC [14]), an NLFE which is a current research project at ICL. NEL uses "Descriptions and Qualifiers" (D&Q) as an intermediary representation (IR) in mapping English sentences on Querymaster (QM - an ICL query language). However, IR such as D&Q which are to an extent linguistically motivated can often represent queries which cannot be evaluated against the DB. Since the aim is to inform the user about how his query has been ultimately understood, a text deriw~.d from snch IRs may be misleading. This influenced our choice of QM as the input to the paraphraser~ which results in a system which will equally benefit users who have no access to a NLFE. The QM question is translated into the RC before paraphrasing which makes the system independent of QM syntax. Since a trivial mapping exists between all relationally complete query languages and the RC, this step enhances the system's portability. The paraphrase must be grammatical, and this can best be ensured by reference to a linguistic theory which can contain a grammar. Our choice of Lexical Functional Grammar (LFG - [9,7]) as an underpinning :is to some extent arbitrary, as most implementable generative theories will accomodate our needs. Yet, the high degree of stratification in the theory was thought an advantage as it predicts which linguistic information must be calculated at every stage in the process. This feature gains importance relative to …

Content truncated to 500 words