In an era of rapidly evolving technological advancements, Questel, a Paris-based intellectual property software and services company, has announced the launch of QaECTER, an AI model developed for semantic patent retrieval. The company claims that this innovative model excels across various metrics, outperforming competing systems even when they are significantly larger in scale.
The QaECTER model is integrated into Questel’s Sophia Search tool and the Orbit Intelligence platform. Both platforms are utilized by research and development engineers, patent attorneys, examiners, and strategic searchers. The model is built from Questel’s in-house AI Lab, utilizing novel training methods that combine citation-driven supervision alongside multi-view self-alignment on proprietary patent data.
To accurately measure the QaECTER model’s effectiveness, Questel developed Sophia-Bench—a comprehensive benchmark aimed at evaluating patent search engines. Sophia-Bench stands as a response to the industry’s need for a realistic metric to assess patent retrieval performance. It features as many as 10,000 patent queries and 75,000 patent corpus spanning a decade.
Results from Sophia-Bench are measured against examiner-cited prior art, involving 12 distinct types of queries including structured patent fields and AI-generated summaries. Significantly, Questel asserts that QaECTER not only excelled in these tests but also surpassed other general-purpose and patent-specific models—including those described as many times larger.
The model was tailored specifically to meet the workflows of practitioners in patent searching. It understands relevance through citation relationships and leverages multiple views of each patent, encompassing inputs from invention disclosures to problem statements, and more.
A technical white paper on QaECTER and Sophia-Bench was co-authored by Kim Gerdes, director of Questel’s AI Lab and Professor of Computer Science at Paris-Saclay University. The paper, titled “Citation-Driven Multi-View Training for Patent Embeddings: QaECTER and Sophia-Bench,” is accessible through the HAL open science platform.
Currently, QaECTER is available for use within the Sophia Search and Orbit Intelligence platforms. Though presently utilized internally, there are plans to make Sophia-Bench accessible to a public audience in the future. Questel’s client base spans over 20,000 clients, with 1.5 million users spread across 30 countries and equipped with 30 offices globally.
For more insights into QaECTER and its place within the industry framework, consult the original article.