SymText : a natural language understanding system for encoding free text medical data

As a need for access to information grows, the lack of accessible information becomes more evident, particularly in clinical settings. Hospital information systems, electronic medical records, and other efforts to computerize data for improved patient care have shown a need for more structured, or coded, data. As a natural and primary conveyance of clinical data, free text is a rich source of information; however, the lack of structure leaves much of the information hidden to automated systems. A method of transforming free text to useful, coded data is that of applied natural language processing; but this field is still being developed. Theoretical and implementational techniques for accurate processing have been developed and evaluated in a search for better parsing solutions. One question in particular is the relationship and importance of syntax, or structure, and semantics, or meaning, for deciphering language. This research describes the theories and development of SymText, a symbolic text processing system that specifies a formal model for combining syntax and semantics to encode free text. The system is designed to separate inferencing from knowledge so that it can be easily modified and extended both within a domain and to alternate domains of knowledge. It details the use of Bayesian networks as a model for context. This model is constructed for the domain of dictated chest x-ray reports. Within this domain the knowledge modules of the system are tuned through two developmental iterations after the initial development. The system is proven to respond to training with an overall increase in coding accuracy on an independent data set from 48.1% to 86.1% in recall and from 70.3% to 85.6% in precision. We conclude that the applied techniques for understanding natural language are generalizable and a promising beginning for solutions to generating coded data.