Patent Map Generation Using Concept-Based Vector Space Model

This paper proposes a patent map generation system using concept-based vector space model and presents evaluation results from the NTCIR-4 patent feasibility study (FS) task. The concept-base is a knowledge base of words, which expresses each word as an associated vector. The word vectors are computed based on word co-occurrence in a target document set, therefore, the word vectors reflect target documents’ characteristics. Each document in the target document set is expressed as a vector that is composed from vectors associated with words included in the document. The word vectors and document vectors are positioned in an identical vector space and relevant degree between any two words and/or documents can be computed as a cosine coefficient of two vectors. Taking advantage of this model, problems sections and solutions sections of patent documents are expressed as vectors, then, they are clustered and the label word for each cluster are chosen from words which gives high cosine coefficient to the center of gravity of the cluster. A trial of generating patent maps for NTCIR-4 patent FS task topics using the system has been done. Comparing with human-generated patent maps, the system provides fairly good accuracy of clustering of target patents but poor accuracy of cluster labeling.