This paper describes an experimental version of a conversational, natural language, information retrieval system. The system under discussion deals with gastroenterology, a branch of internal medicine. One interesting point is that this system has been implemented on that venerable tool of AI research, the IBM 370/148, using a local dialect (UWLISP) of LISP 1.5.
The introduction presents a fine review of some expert systems for medicine. Then we see a lengthy discussion of the components of the system. First, there is the Transformation Expert System (TESS) that translates Polish (precisely what one expects natural language to be at the Institute for Informatics of Warsaw University) sentences into a formal representation that expresses their meaning. This is used to process the text of documents such as medical descriptions. The first component of TESS is the Analytical Stage (AS), which uncovers the syntactic structure of an input sentence. This stage uses five ATN nets to handle the complexities of Polish grammar, which is briefly discussed. The second component of TESS is the Interpretational Stage (IS), which provides the semantics in terms of pattern-concept pairs. Some discussion of rule construction is provided. Again, ATN nets are used, as is a semantic dictionary.
The next component is the deduction module to assist in query processing. It is based upon the theory of fuzzy sets and was written in the FUZZY programming language. (Modified to run on the IBM machine, it can be considered an extension of PLANNER.) This allows the system to be extended to cover such notions as “frequently,” “possibly,” “often,” “rarely,” and “predominately.” Deduction is used to find an assertion that has precisely the same form or contains the same description as the query, has a slightly different form, or can be traced through a reason-to-effect chain. The final component is a natural language answer generator. Again, one is confronted with the issue of Polish grammar. An example scenario for a query and response is given in the paper as:
Q: What is the reason [sic] acute pancreatitis?
A: The predominant reason of [sic] acute pancreatitis is cholelithiasis. Alcoholism is often a reason of acute pancreatitis. Hypertension is pancreatic ducts perhaps causes acute pancreatitis.
This paper is a must read for afficionados of information retrieval. The combination of natural language and retrieval can obviously yield much fruit. In addition, the incorporation of fuzzy logic demonstrates how imprecision can and should be handled.
