The jikitou biomedical question answering system: Facilitating the next stage in the evolution of information retrieval.

摘要

In clinical and biomedical settings researchers often use specialized search engines to acquire answers to technical questions or to verify experimental results through the use of peer reviewed scientific literature. The outcome of such queries typically results in the reading and scanning of multiple Web pages and documents. Information retrieval is the science of retrieving relevant items; question answering (QA) is a specialized type of information retrieval with the aim of returning precise short answers to queries posed as natural language questions. In this dissertation I describe and discuss a QA system named Jikitou (www.jikitou.com), which creates a dialog with the user that mimics human interaction and utilizes multiple search agents to answer biomedical questions. Jikitou's modular design allows for easy modification and evolution of core components. An evaluation system has been devised to aid the evolution, which allows for the quick and systematic comparison among different algorithms for finding relevant answers. The system's architecture is composed of four subsystems: knowledge base, question analysis, answer agents, and user interface. Multiple software agents find possible answers to questions, and the most relevant are presented to the user. Additional relevant information is presented to the user establishing a kind of dialog with the user to obtain feedback to refine the query. Answers are automatically marked up and linked to semantically relevant content in other databases. The additional information is presented in a popup window that appears when a marked term is clicked. There is a lack of systems that allow the user to establish context, take advantage of multimedia information resources, and utilize both to return the appropriate answer. Jikitou addresses these current requirement gaps in biomedical question answering, namely by, incorporating multimedia information through the HyperGlossary and having the ability to interact with the user through query refinement. Jikitou returns answers to biological questions rather than lists of documents, which reduces the need to read entire documents. In addition to addressing current gaps, the system demonstrates an architectural framework that can continually evolve, maintaining itself as a valuable tool to researchers not only for question answering but also for other information retrieval needs.