[4Q3-GS-9] Natural language processing, information retrieval: Social problem aid

[4Q3-GS-9-03] Claim-based patent map by deep learning

〇Tadashi Tsubota Tsubota¹, Yuichi Miyamura¹, Tomotake Kozu¹ (1. Deloitte Touche Tohmatsu, LLC)

Keywords:LSTM, Sentence embedding, Patent map, Claim

Patent data is generally useful for companies to develop their business strategies based on technology trends. Although patent similarity estimation is a critical step for analyzing technology trends, previous methods relied primarily on unsupervised sentence vectorization in which manual laboring (e.g., thesaurus definition) was needed. Here we introduce a new approach for obtaining embedded vectors of patent claim sentences based on recurrent neural network. The network is trained by a newly developed task to discriminate whether a claim-pair is similar or not. We demonstrate that the discrimination task can be solved by the network with high accuracy, and that the patent technology map created by claim-sentence vectors derived from the trained model clearly separates multiple technology fields included in the patent dataset of interest, without manually elaborating thesaurus and stop-word lists.