中文專利前案檢索模式之成效評估 Evaluation of Prior Art Search Mothods

全球各地的專利申請件數近十多年來不斷的成長，各國專利主管單位皆積極的增聘專利審查官，以應付漸增的數量，並確保專利審查的品質。審查人員在決定核發專利前，都必頇執行前案檢索，以確定是否有先前之專利或技術，與該申請案雷同，做為判斷專利核駁的依據。對產業界而言，在專利申請前或專利訴訟中，也常執行前案檢索，以確保沒有誤觸他人之專利權。因此，專利前案檢索的結果，對當事人有極大的影響。國外在此方面的研究，已有數年的投入。但針對中文的專利，則付之闕如。本研究以臺灣的中文專利為對象，進行五種檢索模式的成效比較。此五種模式包括：人工布林檢索、系統自動檢索、虛擬相關回饋、真實相關回饋，以及人機互動檢索。以24道前案檢索題目觀測，全自動化的虛擬相關回饋幾乎達到半自動化的真實相關回饋的成效水準。另外，由系統提示各種詞彙與人工互動的檢索中，若檢索者不善於運用這些詞彙，則其檢索成效並不顯著，甚至比全自動的檢索效果還差。顯示自動化的檢索技術，已經可以輔助經驗不多或對檢索主題認識不深的使用者，達到更佳檢索效果的境界。 The number of patent applications has increased sharply over the last decades in various countries. Patent offices of these areas have planned or are about to recruit more patent examiners for dealing with the ever-increasing applications. Before issuing a patent, the examiners need to conduct a prior art search in order to know whether the techniques revealed in the application meet the novelty requirement for a patent. On the other hand, any individual or enterprise, before applying a patent or currently involving in a patent law-suit case, will also inevitably conduct a prior art search to make sure they will not infringe other’s patent rights. Therefore, patent prior art search is a high-stake task. Several research activities have been conducted in other countries. However, none has done for traditional Chinese patents. Based on the real-world patent collection, this study compares five models for prior art search, namely manual Boolean search, automatic search, pseudo relevance feedback, true relevance feedback, and human-machine interaction. Evaluated on 24 prior art search items, fully automatic pseudo relevance feedback was found to be able to achieve the effectiveness level of semi-automatic true relevance feedback. In addition, the human-machine interaction based on various terms suggested by the retrieval system performed un-expected low, even worse than the fully automatic models. However, this shows that the automatic retrieval techniques have reached a level that is able to help novice users in promoting their prior art search performance.