| A natural language (NL) system called SEEGER (Sleep {dollar}underline{lcub}rm EEGer{rcub}{dollar}) has been developed to retrieve information from a sleep database without requiring any database expertise or computer programming knowledge of the operator. A new knowledge-based structure for building a NL system has been developed and designed into SEEGER. It provides a more convenient method for incorporating domain knowledge than the previous approaches to building NL systems. In addition, it provides a more efficient database mapping method for those NL requests that rely on bottom-up parsing more often than top-down parsing. SEEGER handles ungrammatical input, makes inferences, and disambiguates some complicated word senses. It engages in an interactive dialogue to correct spelling errors, to clarify ambiguous or incomplete queries, and to enter a synonym when a word is not in the lexicon. The database is then retrieved by retrieval query created from SEEGER. The sleep database has been designed to contain information about subject record, channel recording, epoch summaries of waveform occurrences in EEG/EOG/EMG, sleep stage parameters, and night parameters. It has been implemented with the commercially available dBASE IV database management system.; SEEGER is now in the experimental stage, and it was evaluated using four users familiar with sleep data analysis. It gave correct answers at the performance level of 63%, achieving the preliminary design goal. The entire program has been written in about 10,000 lines of LISP. The system performance was analyzed in terms of the elements of the syntactic variations, semantic complexities, and interactive dialogues based on the test results. These tasks show that future improvements can lead to a sleep NL system performing at the level of a data processing technician, but it would take at least two man-years of effort, assuming the developers already possess some background in artificial intelligence programming. SEEGER has been implemented in a personal computer; it requires 6 megabytes of system memory and 15 megabytes of hard disk storage for the LISP environment and dBASE IV. Finally, the guidelines for future development are suggested. |
