AI-SDV 2022: New Insights from Trademarks with Natural Language Processing Alexander Lehmann (Canadian Intellectual Property Office, CA)
0 likes•875 views
This document discusses using natural language processing on trademark text data to gain insights. It presents research on how trademark activity changed during COVID-19, detecting emerging trends in trademarks over time, and classifying trademarks by industry. The research uses techniques like topic modeling and deep learning classifiers to analyze trademarks and identify patterns. The analysis of trademarks can provide economic indicators and reveal where businesses are focusing their innovation and market presence.
AI-SDV 2022: New Insights from Trademarks with Natural Language Processing Alexander Lehmann (Canadian Intellectual Property Office, CA)
- 1. New Insights from Trademarks with Natural Language Processing Alex Lehmann CanadianIntellectual Property Office
- 2. Acknowledgement CIPO Economic Research & Strategic Analysis team • Gray Barski • Mohammad Dehghan • Dakshdeep Singh • Xiang Zhao 2
- 3. 3 Outline 1. Trademarks as economic indicators 2. COVID-19 and trademark activity 3. Detecting emerging trends 4. Classifying trademarks by industry
- 5. What is a trademark? • Form of intellectual property (IP) • Means of identifying a firm’s products/services • Name, logo, symbol, etc. 5
- 6. Economic Importance • Trademarks help define a firm’s brand • Leading indicators for market activity • Registration takes time and resources CONCLUSION: Trademarksurveillance shows where firms are interestedin establishing and maintaining a presence in the market. 6
- 7. Text Mining in Trademarks • Trademark applicants must designate goods and services for their mark • Lines of business for the brand 7
- 8. Study Data • 276,271 registered trademarks • Filed 2012-2022 • Available via IP Horizons • Formatted • Lemmatized • Stop words removed 8 “Providinga web-based portal for an array of benefits” “provide web based portal array benefit”
- 10. Research Question • GDP and trademarks are usually correlated • Pandemic GDP went down, but trademarks went up • Mostly first-time trademark filers QUESTION:Which lines of business were driving the trademarkboom? 10
- 13. Regional Variation • How did different regions’ medical suppliers respond to the pandemic? • Estimate distributions of “mask” emphasis in trademarks 13 Regional Prevalence for “mask” topic
- 14. First-Time Filers’ Lines of Business 14 First-time filers Establishedfilers
- 16. Motivation • Emerging markets offer great opportunity for Canadian business, especially small and medium enterprises (SMEs) • Early identification of emerging markets for innovation is a crucial component of building IP awareness and advancing innovation 16
- 17. Emerging Topic Model Text Mining (per-period) 1. Embed trademark text using machine learning. 2. Find clusters of similar trademarks. 3. Represent clusters with most relevant words. Modeling Evolution (overall) 1. Calculate similaritybetween clusters in each time period. 2. Consider similarclusters “linked”. 3. Tracking links tells us how lines of business appear, change, and terminate. 17
- 18. Topic Evolution 1. Continuation 2. Emergence 3. Splitting 4. Merging 5. Termination 18 Initial topic Continuation Emergence Splitting Merging Termination
- 19. Evolution of a Line of Business 19
- 21. Motivation • North American firms have a classification under the North American Industry Classification System (NAICS) • We want to know when firms move/expand into a new industry • Trademarks could be convenient leading indicators for this activity 21
- 22. Classification Methodology 22 Collection of documents (“corpus”) Embedding Model (BERT) Deep learning classifier NAICS code for each document “221113”
- 23. Preliminary Performance • Subsector identification task: • F1 score: ~0.77 • Accuracy > 97% in certain sectors 23 221113 Sector (Utilities) Subsector (Utilities) Industry group (Electric power generation) Industry (Electric power generation) National sector (Nuclear electric power)
- 24. 24 Conclusion
- 25. Conclusion • Trademarks are useful indicators for innovation and drivers of economic change • Text mining and natural language processing are powerful tools for trademark surveillance • There are myriad opportunities to apply these techniques to answer important questions about innovation 25
- 26. Thankyou! Alex Lehmann BusinessImprovement Services CanadianIntellectual Property Office [email protected] 26