tailieunhanh - Báo cáo khoa học: "INTERPRETING NATURAL LANGUAGE DATABASE UPDATES"

Although the problem of querying a database in natural language has been studied extensively, there has been relatively little work on processing database updates expressed in natural language. To interpret update requests, several linguistic issues must be addressod that do not typically pose difficulties when dealing exclusively with queries. This paper briefly examines some of the linguistic problems encountered, and describes an implemented system that performs simple natural language database update | INTERPRETING NATURAL LANGUAGE DATABASE UPDATES s. Jerrold Kaplan Jim Davidson Computer Science Dept. Stanford University Stanford. Ca. 94305 1. Introduction Although the problem of querying a database in natural language has been studied extensively there has been relatively little work on processing database updates expressed in natural language. To interpret update requests several linguistic issues must be addressed that do not typically pose difficulties when dealing exclusively with queries. This paper briefly examines some of the linguistic problems encountered and describes an implemented system that performs simple natural language database updates. The primary difficulty with interpreting natural language updates is that there may be several ways in which a particular update can be performed in the underlying database. Many of these options while literally correct and semantically meaningful may correspond to bizarre interpretations of the request While human speakers would intuitively reject these unusual readings a computer program may be unable to distinguish them from more appropriate ones. If carried out they often have undesirable side effects on the database. For example a simple request to Change the teacher of CS345 from Smith tò Jones might be carried out by altering the number of a course that Jones already teaches to be CS345 by changing Smith s name to b Jones or by modifying a teaches link in the database. While all of these may literally carry out the update they may implicitly cause unanticipated changes such as altering Jones salary to be Smith s. Our approach to this problem is to generate a limited set of candidate updates rank them according to a set of domainindependent heuristics that reflect general properties of reasonable updates and either perform the update or present the highest ranked options to the user for selection. This process may be guided by various linguistic considerations such as the difference between ưansparent and