A natural language processing system to assess user needs in information retrieval

An important goal of Biomedical Informatics is the delivery of information to health care providers that can positively influence the health care process. The biomedical literature is a significant source of such information but it is difficult to access at the point-of-care. Information retrieval (IR) systems can deliver this information, but do not assist the clinician in understanding his or her information need. Our new approach to clinical IR specifies the information a clinician requires before formulating IR queries by analyzing relevant text documents with a new Medical Language Processing (MLP) system.; We have designed a MLP system that can structure natural language for both representing contextual information and for conceptual IR. The system was designed specifically to require the minimum amount of human engineering. The system uses existing language resources, specifically the Unified Medical Language System's definitions for semantic types and relationships, to structure text into conceptual graphs. Our MLP system differs from existing MLP systems through its use of computerized learning techniques. Our system is capable of generating its knowledge structures solely from positive training examples. Furthermore, the system is robust and will never fail to generate a parse and therefore will always be able to pass a parse along to an IR system.; The system uses a machine learned syntactic grammar as the foundation for a semantic one. Our results showed that using syntactic analysis prior to semantic parsing improved the system's ability to identify semantic relationships. This MLP system has been integrated into an existing information needs identification system. The system is able to generate information needs based on selected user text. This shifts the burden of articulating information needs away from the user; the user need only select a region of interesting text and the system will generate reasonable information needs based on that selection. When compared to human generated graphs, the system as a whole performed comparably in identifying user needs, although it still requires clinical evaluation to confirm that the generated needs are can be utilized in IR tasks.