Search

We describe the development of an advanced retrieval engine to search information in libraries of chemical patents. We use the international TREC (Text Retrieval Conferences) evaluation framework to develop and assess an original search strategy combining a standard vector-space model, a network of co-citations between patents, and a strategy of standardization/expansion of chemical named-entities based on natural language processing. Our basic engine obtains an average accuracy of 0.067. The most significant precision gain is provided by the use of co-citations (+168%), while other contents, in particular ICP codes, do not improve retrieval effectiveness of the engine. The official TREC performance of our engine (ranked #1, with a mean average precision approaching 20%) emphasize the role of document contents as opposed to technological expertise in document retrieval. [ABSTRACT FROM AUTHOR]