Publication details
The selection of electronic text documents supported by only positive examples
| Authors | |
|---|---|
| Year of publication | 2006 |
| Type | Article in Proceedings |
| Conference | JADT'06 |
| MU Faculty or unit | |
| Citation | |
| Field | Informatics |
| Keywords | machine learning text categorization relevance ranking k-nearest neighbors support vector machines positive examples feature selection feature optimization domain-independence |
| Description | The European Commission has a freely accessible news monitoring system called the Europe Media Monitor NewsBrief (http://press.jrc.it/), which is available for all twenty official languages of the European Union, plus some more languages. Among other things, NewsBrief categorizes articles through routing procedures and it alerts users interested in a large variety of different subject domains automatically. In the effort to improve the multilingual categorization and relevance ranking functionality for some complex interest profiles, for which only positive examples are currently available, we implemented a modified k-NN (k-nearest neighbors) algorithm and empirically detected parameters and parameter settings that produce good results for rather different subject areas (news on the EU-Constitution, on Iraq, and on Terrorism). Experiments on this real-life data yielded very satisfying results: a precision of over 90% for a recall of up to 70%. These results were then compared to others achieved with one-class SVM and with SVM that was trained on both positive and artificially generated negative example sets. Efforts are currently underway to incorporate this new functionality within NewsBrief and to make it available to the users. |