PUB - Publikationen an der Universität Bielefeld

Background: Electronic Patient Records (EPRs) contain vast amounts of information in the form of texts, such as letters sent from a hospital to the patient’s local doctor (GP). The ability to extract data reliably from these texts would yield advances both in medical research and in individual patient care. In developing effective extraction methods, a crucial step is to create corpora that align samples of the relevant texts with a formal encoding of their meanings. To compile such corpora by human annotation is expensive and time-consuming; the more the task can be automated, the better. Results: We propose a solution in which the results of an extraction algorithm are corrected using a transparent and fully reliable Natural Language Interface. Information Extraction records are transcoded to a Description Logic A-box and presented through a generated ‘feedback text’, which a medical expert can edit using the Conceptual Authoring technology. Conclusions: At present we achieve a weak alignment between the encoded meaning and the original text (i.e., a set of records for the whole text); the next step is to achieve a strong alignment that links each phrase in the original text with the entity in the A-box that encodes its meaning. This technology has a wide range of potential applications, for instance as an editing tool for the Semantic Web.