LCI-INSA Linguistic Experiment for CLEF-IP Classification Track

We present the experiment the LCI group has performed to prepare our submission to CLEF-IP Classification Track. In this preliminary experiment we used a part of the available target documents as test set and the rest as train set. We describe the systems AGFL used for extracting these triples and the LCS used for classification by the Winnow algorithm. We show that the use of linguistic triples in place of bags of words improves the accuracy, as well as using the names and addresses of the applicants. we found that using the complete descriptions as bags of words does not really perform better than using only abstracts and titles. Some simple mathematics show that the official measures are redundant and that R@N should be used to evaluate a ranking, P@1 to evaluate routing and that the usual precision, recall and F1 should be used on the results of a real classification, that is a selection of the classes performed internally by the classifier.