Developing Tech2Vec: A new embedding approach of technology information using a triple layer

The recent increase in the number of patent applications highlights the urgent need for an effective embedding technique to automatically analyze enormous patent datasets. Extensive research is being conducted on the application of high-performance artificial intelligence (AI) technology to enhance patent analysis tasks. However, these studies do not consider various types of data. Instead, they examine technological information from a single perspective, such as technological terminology, patent functions, and goods. To cover all aspects, namely, technological system, function, and technology, a technological information analysis model that exploits both structured and unstructured data from previous patent filings is required. Therefore, this study proposes a new embedding approach called Tech2Vec to conduct function-oriented patent searches that can use the function and technological information of patent documents. More precisely, various types of technological information included in patent applications are organized into a triple layer, that is, the system, function, and component layers; vectorized layer by layer; and concatenated into a single technology vector. For example, by leveraging the patents and papers of three sectors, namely electric vehicles, displays and industrial robots, Tech2Vec is effectively applied and mapped to the technological latent space. Additionally, a function-oriented patent search is performed by comparing the query vectors entered by a user in natural language rather than the search query format. This study may be used as a reference for a range of technology management activities, such as document categorization, technological opportunity identification, and technology evolution analysis.