Gaétan de Rassenfosse Keynote - 9th Annual IP Data & Research Conference

Value of the Information & Strength of the Argument

The talk provides substantial value by presenting concrete, original research applications of generative AI in patent analytics. The argumentation is solid, grounded in peer-reviewed work and practical examples. De Rassenfosse clearly explains the strengths and limitations of each method, particularly the cost-benefit trade-offs of using LLMs for classification. He also demonstrates how LLM-based embeddings and next-token probabilities can yield novel indicators of patent importance and blocking power, arguing convincingly that these capture dimensions not measured by traditional citation counts. The presentation is well-structured, logical, and supported by quantitative results and visual aids.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates high scientific rigor. De Rassenfosse cites his own publications in top journals (e.g., Strategic Management Journal) and references specific tools like Grobid and open datasets on Zenodo. He transparently discusses the methodology, including manual labeling and accuracy metrics (92% overall). The title is accurate: the content directly addresses generative AI and patent analytics, exploring new methods and applications. The talk includes a brief sponsorship segment? No, there is no such segment. The speaker mentions his substack (the patentist) but that is not commercial. The presentation adheres to academic standards, and the speaker acknowledges limitations and future directions.

Titre matches content: keynote on generative AI and patent analytics, covering new methods and applications.

9/10

Presentation by a leading expert (EPFL associate professor) based on original research and peer-reviewed publications. The methods are transparent, using open-source tools and publicly available data. The talk includes specific cases and quantitative results, and references known academic work.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:30Introduction and outline of three ways generative AI changes patent analytics
• 11:00Application 1: moving beyond front-page citations using vector embeddings; survey evidence and similarity analysis
• 23:00Application 2: classifying office action references with LLMs and cost-efficient machine learning pipeline
• 37:00New research: measuring patent scope destruction via LLM next-token probabilities; blocking score and ranking
• 50:00Third element: autonomous agentic workflows and future implications

• The Patentist (substack blog by de Rassenfosse) — Discussed as a source explaining citation-based measures of invention importance
• Grobid (open-source ML library by Patrick Lopez) — Used for parsing and restructuring patent documents
• Sebastian Rago's paper on claim scope and language models — Published in World Patent Information; basis for the blocking score
• Zenodo dataset on office action classifications — Freely available dataset for linking non-IP data