TrademarkML: Predicting the Likelihood of Confusion under Art. 8 (1) EUTMR

Abstract

Over the past years, the number of trademark application has risen continuously. To make it easier for trademark owners to protect their trademarks' territory and for trademark applicants to know if their trademark is likely to be refused, an automated and reliable tool is needed to compare two trademarks and their respective goods and services. Most of the existing services offering a similarity search just consider the spelling of the trademarks’ names, which is insufficient to assess likelihood of confusion. For this reason, this thesis introduces the concept of a trademark management system, TrademarkML. TrademarkML faciliates tasks like trademark monitoring and searching for conflicting trademarks by automatically classifying trademark pairs. TrademarkML employs state-of-the-art methods for extracting meaningful features for the comparison of trademarks in various aspects. Exhaustive feature selection is then used to tune random forests and support vector machines to all feature combinations. TrademarkML achieves an F1-score of 88% for word marks and 81% for figurative marks, a recall of 98% for word marks and 84% for figurative marks, and a precision of 80% for word marks and 77% for figurative marks. Overall, TrademarkML performs better on word mark data by 6.8% on average for each metric. None of the top-performing models uses a feature for measuring the similarity of goods and services. This makes it impossible to correctly predict partially upheld oppositions. A larger dataset and more meaningful features are required to overcome this issue.