As the world’s second-largest supplier of automotive parts, DENSO produces enormous quantities of textual data on problems and procedures at dozens of manufacturing plants around the globe. Analyzing all this data, which has been gathered in notes written in multiple languages in ever-changing production environments over many years, is an intimidating challenge for the Kariya, Japan-based firm. DENSO is turning to artificial intelligence technologies to better leverage its knowledge bases with this unstructured textual data, teaming up with Luminoso Technologies, a natural language understanding company in Boston. The auto parts maker had found that traditional software approaches to analyzing such data on this global scale are imprecise and difficult to keep updated. Moreover, these software platforms struggle to understand meaning and context in Japanese and other languages. “Luminoso’s natural language analysis software offers a feature-rich, high-quality solution that provides us with a comprehensive understanding of the issues surrounding the facilities,” says Masahiro Saito, director of DENSO’s production engineering division. His group has now begun analyzing hundreds of thousands of maintenance reports from all DENSO factories in Japan.
Luminoso’s natural language analysis software offers a feature-rich, high-quality solution that provides us with a comprehensive understanding of the issues surrounding the facilities.
Spun out of the MIT Media Lab, Luminoso is an MIT Startup Exchange (STEX) company and was chosen for the first group chosen for the STEX25 startup accelerator program. Luminoso first connected with DENSO, an MIT Industrial Liaison Program (ILP) member, via ILP. “The successful collaboration between Luminoso and DENSO is just the type of outcome we’re trying to promote and foster,” says Karl Koster, MIT executive director of corporate relations. Natural language learning, deep and wide Luminoso’s software combines deep learning and natural language processing to help companies rapidly and accurately understand the concepts within their unstructured textual data—without requiring massive sets of training data. The software traces its roots back to the Media Lab’s Open Mind Common Sense project, kicked off in 1999. Common Sense was a pioneering crowd-sourced project aimed to help computers understand the meanings of words that people use and the relationships between them. Common Sense evolved into ConceptNet, an ongoing global open data project, which laid the foundation for Luminoso’s core software. “We started out being incredibly multilingual, because the data that we collected at MIT for the Common Sense project was inherently multilingual to start with,” says Catherine Havasi, co-founder of both Common Sense and Luminoso. “At MIT, we partnered with universities in Brazil and in Japan as well, and we worked with additional Media Lab companies to build the Japanese system.” Luminoso’s software now supports 15 national languages, with close connections between them. “We are probably the first company to commercialize transfer learning, which takes something in one domain or language, and easily refines it or trains it into a different language,” Havasi says. “Most of what we do is take big knowledge about how the world works in general and specialize that knowledge to a particular domain, say car manufacturing,” she says. “That specialization becomes really critical, and being able to do it in multiple languages is incredibly important.” Proof on the production lines Havasi presented Luminoso at an ILP meeting in Tokyo in 2015, emphasizing both the power of its natural language platform and the ability to provide services as well in Japanese as in English. DENSO experts immediately began a discussion. “They had a great technical understanding of the problem of understanding their specific data, and they were taking the time to understand what we were bringing to the table,” Havasi says. “After we made the connection, we realized that this was a real use case for our software.” In the first follow-up meeting between the two firms, DENSO provided about 20,000 sample records for Luminoso to examine, recalls Kohei Nakamura of the DENSO Factory Internet of Things team. “Luminoso immediately showed the analysis results on the spot,” he says. “The accuracy of understanding the meaning of words without training data was higher than that of other natural language technologies.” “We were shocked that Luminoso could understand words specific to our company (such as our subsidiary name, DENSO Wave) as one word or concept,” he adds. Previously, the company needed to manually annotate such company-specific words.
We were shocked that Luminoso could understand words specific to our company (such as our subsidiary name, DENSO Wave) as one word or concept.
Conventional text-analysis methods eventually could have achieved this higher level of accurate categorization, Nakamura says. But looking to the future, and the need to integrate data from a much broader set of information sources, DENSO believed that those more traditional methods would not make the grade. In 2017, DENSO and Luminoso began a proof of concept project to learn from DENSO’s historical non-structured textual data on manufacturing equipment maintenance. This textual data is entered by maintenance workers following up on issues on production lines. They write a short report, and then after they figure out a fix for an issue, and the effects of the fix, they update the note. Throughout this process, the maintenance workers describe the issues rapidly and without worrying about exact word choice. Some might write “robot” and others “RB”, for instance. “So how can you assume you know exactly how people will talk about everything?” says David Smith, representative director at Luminoso Japan. “But Luminoso does that, because we look at every word independent of its meaning and we relate it to other words and their concepts.” In one case, Luminoso’s software let DENSO engineers responsible for machine design instantly view information across more than a million documents, aiding their efforts to create highly productive new machines. Among the efforts, DENSO plans to use the software to explore machine breakdowns, identify the root causes of breakdowns that require long hours for recovery, and find ways to change the design of equipment or optimize schedules for machine maintenance. The natural language software offers great potential for many future applications, says Nakamura, especially as DENSO gains more experience with the tools and their results. It also will open up opportunities for DENSO to move ahead into more complex analyses that can incorporate structural data and multilingual integration, he says.