NTT will utilize ”tsuzumi”, a lightweight large-scale language model (LLM) with world-class Japanese language processing performance, to provide it as a commercial service from the NTT Group in March 2024.
tsuzumi has four main characteristics.
One is that it supports not only Japanese but also English and other languages, with Japanese in particular boasting world-class performance in various benchmarks.
The second is high cost performance. Although it has the same performance as Open AI’s GPT-3, it can achieve high performance with less power consumption and less GPU.
Third, general-purpose AI does not provide sufficient performance when a lot of technical terms are involved, such as in the digital domain or software development, but tsuzumi can be customized to suit the industry.
Fourth, it is also possible to read charts and tables.
　
NTT has been conducting wide-ranging trials both internally and externally since October of this year.
For example, they are collaborating with Kyoto University Hospital and others to promote data structuring of electronic medical records. Although medical institutions are making progress in introducing electronic medical records, the way electronic medical records are written differs depending on the doctor and hospital, making it difficult to share and analyze medical record data as a whole. Using tsuzumi to convert unstructured data into structured data will speed up drug development and advance personalized medical care.

Tokyo Marine & Nichido is promoting its use at its contact centers.
The company has over 10,000 operators supporting users. The operator listens to the circumstances of the accident and injuries over the phone, organizes the details of the response after the call ends, and inputs the necessary information into the system. It is said that after-call work takes up 800,000 hours a year. Until now, efforts have been made to reduce time through voice mining, etc., but by combining it with tsuzumi, progress will be made in summarizing and organizing the content of responses, and it is expected that the amount of after-call work will be reduced by more than 50%.

NTT DATA for software development. The code required for software development is intellectual property and highly confidential, but by using the lightweight tsuzumi, code and test scenarios can be automatically generated in a secure environment such as on-premises. This can significantly shorten the development period.
It is possible to automatically generate business definitions and development, operation, and maintenance documents appropriately, leading to improved quality of software development.

NTT East uses it at its support center. NTT East’s support center is staffed by more than 1,000 operators, but support for handling communication equipment covers 3,500 types of equipment and a total of 200,000 pages of manuals, which inevitably leaves users waiting. As a result of verification in a test environment, the time spent waiting for users is expected to be reduced by 90%, and the amount of time required to manually create frequently asked questions and answers based on response history is expected to be reduced by 80%.