基於專利技術功效矩陣之科技機會探索方法

Abstract

有效探索目標科技領域即將開始發展的技術，不僅能協助政府規劃國家科技政策，也使公司在面對科技變化時有較長的應對時間、得以增加競爭優勢。基於效率與效能的考量，許多研究致力於增加自動化科技機會探索的可能性。在這當中，專利文件因與科技高度相關且數量豐富，加上格式明確有助於分析，經常作為科技機會探索研究的資料來源。過去基於專利的科技機會探索研究，主要可分為兩個主要的方向－專利地圖分析法與型態矩陣(morphology matrix)分析法。前者利用文字探勘技術將專利文件投影至二維地圖，並挑選地圖中的空洞當作科技機會；後者則利用專家為特定科技領域建立的型態矩陣，將未被實現的狀態組合視為科技機會。然而，過去的研究雖皆為自動化科技機會探索帶來許多可能性，卻也存在著侷限。專利地圖分析法缺少有意義的軸座標，需仰賴專家為空洞做進一步的分析；型態矩陣分析法則於定義矩陣時須投入大量專家人力。為此，本於技術功效矩陣富明確科技意涵，且較型態矩陣具自動建置可能的特性，本研究提出基於技術功效矩陣的科技機會探索方法，並分為兩階段呈現。第一階段藉技術與功效詞的特性萃取特徵詞、利用情境式(context-aware)文件分群技術將已知技術/功效類別描述擴充以協助取出有意義的特徵詞，並以調整後的k-means將專利於技術層面及功效層面各自分群，形成技術功效矩陣。第二階段利用技術功效矩陣，將當期尚無專利的格子(矩陣上的一種技術-功效配對)視為可能機會，並藉助矩陣取出預測子(predictors)分辨該機會真偽。第一階段中，我們證明提供類別描述並加以擴充能增加矩陣建置的準確性(precision)；第二階段則藉由呈現不同類型預測子、不同特性格子的可預測性，顯示此科技機會探索方法的可行性。

Proper technological opportunity discovery (TOD) brings both governments and companies longer lead-time to prepare for changes, and thus yield competitive advantage. In order to efficiently and effectively discover technological opportunity (TO), some prior studies dedicate to increase the feasibility of TOD automation, and use patents as their objective data source. Although these studies all shed light on the possibility for automated TOD, they still suffer from such limitations as heavily relying on domain experts. Considering technology effect matrix (T/E matrix) as an organized representation of a target technological field, which is more easily constructed than morphology matrix, we propose a T/E matrix based method for TOD. This study is decomposed into two phases, one is to construct a T/E matrix with little experts’ involvement and the other is to automatically identify unoccupied grids on a T/E matrix to be true or false TOs. In the first phase, we extract features based on special characteristics of technology and effect terms, select features with context-aware document clustering technique, and cluster patents by an adjusted k-means method. We verify that the cluster precision can be increased by our method, and even escalated when adopting different expansion corpora. In the second phase, we regard unoccupied grids on T/E matrix as candidate TOs and use predictors to help classify grids into true or false TOs. We figure out that diverse effectiveness occurs when applying different types of predictors and when the to-be-predicted grids have different characteristics, and disclose the feasibility of our proposed method.