Incomplete and Fuzzy Conceptual Graphs to Automatically Index Medical Reports

Most of Information Retrieval (IR) systems are still based on bag of word paradigm. This is a strong limitation if one needs high precision answers. For example, in restricted domain, like medicine, user builds short and precise query, like “Show me chest CT images with emphysema.”, and expects from the system precise answers. In such a case, the use of natural language processing to model document content is the only way to improve IR precision. This paper presents a model for text IR that index documents with Fuzzy Conceptual Graphs (FCG). Building automatically a complete and relevant conceptual structure is known to be a difficult task. To overcome this problem and keeping automatic graph building, we promote the use of incomplete FCG. We show how to deal with this incompleteness by using confidence. This confidence is attached to concepts and conceptual relations. As we use FCG as index, the matching process is based on a fuzzy graph matching. Finally, our experiments show that this outperforms classical word based indexing.