Model-based Analysis of Messages about Equipment

Model-based Analysis of Messages about Equipment Ralph Grishman, Tomasz Ksiezyk, and Ngo Thanh Nhan Department of Computer Science Courant Institute of Mathematical Sciences New York University ABSTRACT The aim of PROTEUS -- a system for the analysis of short technical texts -- is to increase the reliability of the analysis process through the integration of syntactic and semantic constraints, domain knowledge, and knowledge of discourse structure. This system is initially being applied to the analysis of messages describing the failure, diagnosis, and repair of selected pieces of equipment. This has required us to develop a detailed model of the structure and function of the equipment involved. We focus in this paper on the nature of this model and the roles it plays in the syntactic and semantic analysis of the text. 73 I. Introduction Considerable progress has been made in developing systems which understand short passages of technical text. Several prototypes have been developed, for such domains as patient medical records [Sager 1978], equipment failure reports [Marsh 1984], and intelligence messages [Montgomery 1983]. Except for very narrow domains such as weather reports, however, none of these systems seem tO be robust enough for operational use. Typical success rates - where any are reported - are in the range of 70 to 80% of sentences correctly analyzed; substantially better rates are very hard to obtain, even with careful system tuning'. Our objective in developing PROTEUS (the PROtotype TExt Understanding System) is to see if this rate can be substantially improved for a domain of moderate complexity. In order to achieve this improvement, we must bring to bear on the language analysis task the various syntactic, semantic, and discourse constraints, along with a fairly detailed knowledge of the domain of discourse. Our system is initially being applied to equipment failure reports ("CASREPs") for selected equipment on board Navy ships (initially, the equipment in the starting air system); a sample message is shown in Figure 1. In this case, the domain knowledge is the knowledge of the structure and function of these pieces of equipment. In this paper we first present an overview of the PROTEUS system. We then focus on the domain information: how it is represented, how it is integrated with the language processing, and how it serves to resolve ambiguities in the input text. 2. Prior work New York University has been involved in the automated analysis and structuring of technical text for over a decade. Most of this work has been on medical records [Sager 1978, Hirschman 1982], but we have also been involved with the Naval Research Laboratory on a system for CASREP's [Marsh 1984]. These systems used domain-specific frame-like target structures, and employed selectional constraints to weed out bad parses, but did not incorporate detailed domain models. Our experience with these systems - in particular, the difficulty of obtaining success rates (% of sentences correctly analyzed) much above 75% - led us to our work on PROTEUS. The use of detailed domain models in language processing systems …

Content truncated to 500 words