TORUS - A Natural Language Understanding System For Data Management

This paper describes TORUS, a natural language understanding system which serves as a front end to a data base management system in order to facilitate communication with a casual user. The system uses a semantic network for "understanding" each input statement and for deciding what information to output in response. The semantic network stores general knowledge about the problem domain, in this case "student files" and the educational process at the University of Toronto, along with specific information obtained during the dialogue with the user. A number of associated functions make it possible to integrate the meaning of an input statement to the semantic network , and to select a portion of the semantic network which stores information that must be output. A first version of TORUS has been implemented and is currently being tested. 1. Introduction This paper describes the scope and current status of the TORUS (TORONTO Understanding System) project. The main goal of the project is to develop a methodology for designing and implementing "Understanding" Data Base Management Systems (UDBMSs), i.e., systems which have available (a) A data base stored in and maintained by a data base management system (DBMS), (b) A natural language understanding system (NLUS), serving as a front end to the DBMS, which is given knowledge about the data, can understand and respond to simple English sentences and can engage the user in a dialogue. Such systems can obviously play a very important role in the future in making data bases available to casual users who have neither the time nor the patience to learn a programming or query language so that they can communicate with the computer. As a first step towards achieving the basic goal of TORUS, we have designed and implemented a prototype UDBMS, using a data base of graduate student files. Two of the most important methodo-logical decisions we made are given below: (a) We have fixed the type of DBMS available to the UDBMS to be a relational system along the lines proposed by Codd [1]. This decision enables us to be more explicit about the NLUS-DBMS interface and also makes our work easier since relational DBMSs offer a more powerful set of "higher level" commands than other types of DBMSs. (b) We have decided to base the "understanding" capabilities of the NLUS on semantic nets, i.e., directed labelled graphs whose nodes represent conceptual objects and whose edges represent …