構造的関連性学習を用いた論文特許検索

抄録

We propose a method to automatically associate documents from different domains such as scientific paper and patent. The proposed method enables cross-domain academic search on the scientific data. Borrowing ideas from multi-task learning and structural correspondence learning, our approach automatically identifies correspondences among the words from different domains using a small number of so-called concepts. Our method models the correlation between the concepts and all other words by training linear classifiers on the documents from different domains.