Robert Ambrogi Blog
Questel, an intellectual property software and services company headquartered in Paris, has released QaECTER, a new AI model designed specifically for semantic patent retrieval.
The company says the model outperforms competing systems, including those that are significantly larger, across every query type, technology domain and jurisdiction tested.
QaECTER is the product of Questel’s in-house AI Lab, which developed the model using what the company describes as novel training methodologies that combine citation-driven supervision with multi-view self-alignment on Questel’s proprietary patent data.
The model is now available within Questel’s Sophia Search tool and its Orbit Intelligence platform.
A Benchmark Built Alongside the Model
To evaluate the model’s performance, Questel’s AI Lab also created Sophia-Bench, a large-scale benchmark for evaluating patent search engines.
The company says the benchmark was built to address what it characterizes as a longstanding gap in the field — the absence of a realistic, practice-oriented way to measure patent retrieval performance.
Sophia-Bench comprises 10,000 patent queries and 75,000 corpus patents spanning 10 years, stratified across eight technology sections under the International Patent Classification system and 12 filing jurisdictions, including the European Patent Office, USPTO, and patent offices in China, South Korea, and Japan.
Results are graded against examiner-cited prior art, and the benchmark tests 12 distinct query types ranging from structured patent fields to AI-generated summaries.
Questel says QaECTER was tested against both Sophia-Bench and an independent external benchmark, and that it outperformed all competing general-purpose and patent-specific retrieval models evaluated, including systems described as 23 times larger.
Designed Around Practitioner Workflows
According to Questel, QaECTER was trained with patent search practitioners in mind. The model learns relevance through citation relationships and what the company calls “multiple complementary views” of each patent, enabling it to handle a range of inputs, from invention disclosures and problem statements to claims, abstracts, and technical descriptions.
Within Sophia Search, the model is paired with multilingual query translation, reranking, and other capabilities aimed at professional patent workflows.
The technology is designed to serve R&D engineers, patent attorneys, examiners and strategic searchers, Questel says.
Kim Gerdes, director of Questel’s AI Lab and also professor of computer science at Paris-Saclay University, co-authored a technical paper on both QaECTER and Sophia-Bench, published on the HAL open science platform: “Citation-Driven Multi-View Training for Patent Embeddings: QaECTER and Sophia-Bench.”
Gerdes said the team is “dedicated to advancing AI in the IP field” and characterized the developments as placing Questel “at the very forefront of modern Deep Learning for information retrieval.”
Availability
QaECTER is currently available in Sophia Search and Orbit Intelligence. Sophia-Bench is being used internally for now, though Questel says it plans to make the benchmark publicly available in the future.
Questel serves more than 20,000 clients and 1.5 million users across 30 countries and has 30 offices worldwide.