myClass: A Mature Tool for Patent Classification

In this task 2,000 patents in three languages (English, French and German) were to be classified among approximately 600 categories. We used a classifier based on neural networks of the Winnow type. This classifier is already used for similar tasks in professional applications. We tested three different approaches to improve the classification accuracy: the first one aimed at solving the issue of poorly-documented categories, the second one was meant to enrich the overall training corpus and the third one was based on the processing of the corpus' collocations. Although we ranked first in this competition, none of the three approaches mentioned above provided for a clear improvement in classification accuracy; our results were essentially due to the implementation of the classification algorithm itself.

• Type: Conference
• Year: 2010
• Where: CLEF
• Authors: Jacques Guyot, Karim Benzineb, Gilles Falquet
• Keywords: 000 Patents, Classification Accuracy, CLEF 2010, Information Technology, Task 2