A Swinburne team is analysing text data from Australia’s patent office to predict the evolution of technologies.
Associate Professor Kai Qin and his team work at Swinburne’s new Intelligent Data Analytics Lab, part of the Swinburne Digital Research Innovation Capability Platform.
The data scientists bring deep neural-network machine-learning algorithms and rapidly maturing natural language processing technology to government agency, IP Australia. Their hope is to develop software that scans the agency’s past 20 years of patents as well as scientific article texts, including specifications and claims.
“Our model will tackle the natural language processing needed to analyse the patent data and any associated scientific articles,” said Qin. “But it will be the modelling parameters that are hardest to identify.”
This project follows a successful 2017 Swinburne collaboration with IP Australia and the University of Melbourne. A Swinburne group created a world-first trademark database, TM-Link, linking trademark application numbers across countries. This provided insights into the foreign trade interests of Australian businesses, by showing how trademarks are used in different markets, while also opening research into trademark trends.
For Qin’s patent project, analysing information regarding failed technologies, as well as success stories, could indicate which technologies the country should invest in. On a broader scale, the work could reveal why scientific knowledge progresses in certain directions and triggers for faster or slower growth across fields.
These results, funded by an ARC Linkage Project grant, could ultimately improve IP Australia’s database search, reveal new technologies and potential collaborators for business analytics companies, and help academic economists to understand how knowledge travels and accumulates.
Tracking Trademarks with TM-Link
A 2017 collaboration between IP Australia, Swinburne and the University of Melbourne resulted in TM-Link, the first trademark database of its kind.
The databases’ neural network was designed to identify equivalent trade marks in different jurisdictions, assigning a common identification marker by considering similar trademark text, applicant names and classes.
The data already includes more than 10 million entries and helps track trademark use across regions. It confirms its finding based on text and imagery, using state-of-the-art imaging technology.
A trademark labelling game also improves the algorithm’s training data by asking users to confirm a match between suggested pairs of trademarks.