A worldwide linked trademark database for IP research

Researchers and policy makers are concerned with many international issues regarding trademarks, such as trademark squatting, cluttering, and dilution. Trademark application data can provide an evidence base to inform government policy regarding these issues, and can also produce quantitative insights into economic trends and brand dynamics. Currently, national trademark databases can provide insight into economic and brand dynamics at the national level, but gaining such insight at an international level is more difficult due to a lack of internationally linked trademark data. We are in the process of building a harmonised international trademark database (the 'Patstat of trademarks'), in which equivalent trademarks have been identified across national offices. We have developed a pilot database that incorporates 6.4 million U.S., 1.3 million Australian, and 0.5 million New Zealand trademark applications, spanning over 100 years. The database will be extended to incorporate trademark data from other participating intellectual property (IP) offices as they join the project. Confirmed partners include the United Kingdom, Canada, WIPO, and OHIM. We will continue to expand the scope of the project, and intend to include many more IP offices from around the world. In addition to building the pilot database, we have developed a linking algorithm that identifies equivalent trademarks (TMs) across the three jurisdictions. The algorithm can currently be applied to all applications that contain TM text; i.e. around 96% of all applications. In its current state, the algorithm successfully identifies ~ 97% of equivalent TMs that are known to be linked a priori (due to shared international registration number). Current estimates indicate that approximately 40% of candidate positive links identified by the algorithm are false positives. However, we expect the proportion of false positives to become far smaller as we continue to improve the linking algorithm. A major part of improving the linking algorithm will involve combining it with a separate machine learning algorithm that we have recently developed, which exhibits very low false positive and false negative error rates. Briefly, the machine learning algorithm includes an image classification neural network that we adapted to match and disambiguate inventor names in patent records. It uses a novel matching technique whereby each pair of inventor records is compared by firstly converting the two records from raw text into an abstract visual representation, or 'comparison image'. The neural network is able to learn important features within comparison images that indicate whether the two inventor records are likely to be a match (both inventor records refer to the same inventor) or non-match (records refer to different inventors). This is done by training the neural network on data that has been manually labelled as match/non-match. Tests on a sub-sample of the labelled data (withheld from the network during training) indicate error rates as low as ~ 1%. We are currently modifying this machine learning algorithm to match trademarks, rather than inventor names. When complete, the internationally linked trademark database will be a valuable resource for researchers and policy-makers in fields such as econometrics, intellectual property rights, and brand policy.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.