以連結特性為基礎之專利分類研究

Abstract

在科技發展與日漸進的情況下，專利的數量也隨著企業彼此的攻防而增加。然而，專利審核單位或專利分析工作者皆需面臨專利分類的問題。專利分類是個專業性質高且耗時費力的工作，先前許多研究提出利用機器學習的方式來解決專利分類的問題，以達成專利分類的自動化。傳統的專利分類使用文字探勘技術針對專利內文進行自動分類，效能不盡理想。專利文件中仍有許多寶貴的資訊可供運用以提昇自動分類效能。 本研究實作結合以連結特性為基礎之專利分類混合方法，利用專利文件的引用網路及共引用的特性，設計數種核函數，並使用支援向量機進行專利文件自動分類。最後，本研究進行實驗評估，以比較各分類方法在專利分類之成效。

Classifying patents manually is a costly and time-consuming work. As more and more patents come into existence, effective and automatic patent classification is necessary. Traditional text-based document classification uses term vectors to conduct classification based on patent content. However, the performance of text-based approach is not effective. Patent data contains useful information such as patent citations that can be explored in patent classification. In this study, we implement and compare several hybrid patent classification methods according to the link-based features of patents, which exploit references and reference-by features of patents. The link-based features form co-citation and citation network of patents and can be used to improve the effectiveness of classification. The experiment result shows that the hybrid methods based on link-based features can improve the performance of patent classification.