Justsystem at NTCIR-5 Patent Classification

Justsystem participated in Patent Classification Subtask at the Fifth NTCIR workshop. This paper overviews our machine learning-based patent application classification system. Straightforward application of Naive Bayes classifier was effective in theme categorization subtask that has a non-hierarchical category structure. In F-term categorization subtask, we regarded the complicated F-term categorization system as a tree with depth 2. We constructed the document classifier based on the Support Vector Machine and classify documents on this tree. Platt’s sigmoid fitting for SVM output was used for the document ranking. We confirmed that this method was effective for this subtask.