Denecke and Bernauer ARTIFICIAL INTELLIGENCE IN MEDICINE 2007
Citation
Kerstin Denecke and Jochen Bernauer. 2007. Extracting Specific Medical Data Using Semantic Structures. Artificial Intelligence in Medicine, LNCS 2007 Volume 4594/2007, 257-264.
Online version
Summary
The paper presents a medical Information Extraction system which is based on automatic generation of semantic structures for free-text. Mining of discharge summaries for information on the hospitalization of a patient dataset is a gold standard comprising 50 discharge summaries from a surgical department in a hospital.
The key method of this system is building semantic structures from texts. A conceptual graph-like representation is produced for each sentence of a text, and then these semantic structures are exploited to extract information. The system is composed of three major modules: a text-preprocessor, a semantic interpreter, and an extractor. In the text-preprocessor, sections are detected and classifies, and texts are parsed. This results in tagged texts with dates, quantities and dosage specifications, and expressions with special meaning like "exclusion of" and "evidence of". In the semantic interpreter, they used a terminology of medical terms (Wingert Nomenclature (WNC)) Each sentence is mapped to a semantic representation consisting of concepts and relations. In the extractor, according to manually defined rules, information is extracted from the text itself, and its semantic representation in order to fill in templates.
limitation of current approaches current approaches to medical NLP are often limited to a certain medical domain and their construction is fairly complicated. Also limited to a certain language, especially in English.
a method for automatically mapping natural language text to semantic structures.
The advantage of this system is that the components that might be adopted for processing texts in another language because it represents texts as semantic structures.
The system achieved 81-95% precision and 83-97% recall of performance on template filling tasks. The main reasons of error were unknown words, wrong paragraph detection, missing trigger words, and processing needs additional knowledge.
