Introducing Google Neural Machine Translation integrated into Memsource for streamlining product manual translation tasks.

Background

WingArc 1st Inc. boasts a cumulative total of 21,000 companies using its form solution "SVF," holding over 50% market share domestically, while its BI solutions "MotionBoard" and "Dr.Sum" are also leading the BI market.

Not only in Japan, but also expanding the possibilities of data utilization with unique products and services, we are accelerating our expansion into global markets such as China, Singapore, and Australia.

The Documentation Services Group of the Product Management Department that you have requested this time is responsible for organizing the documents created by the development team and releasing the product manuals in Web (HTML) and PDF formats.

Since 2012, we have been assisting with English manual translations for overseas expansion, and in July 2017, we held the "Neural Machine Translation Seminar," which marked the start of the Google Neural Machine Translation implementation support project.

Challenges Before Implementation

Lack of human resources for document translation. We want to utilize Google Neural Machine Translation.

Traditionally, product manuals were translated by humans using the translation support system Memsource.

However, the average translation volume is quite large, ranging from 200,000 to 500,000 characters per month, making it difficult to secure human resources for translation. Therefore, we have started to consider the introduction of machine translation for the purpose of improving efficiency.

Specifically, it is a method where machine translation is used for new translation parts that do not match the existing translation memory, and the translator then makes corrections to that translation.

Additionally, when introducing machine translation, we wanted to apply a UI glossary and style, but since the product manuals are not in a standard HTML format and are customized HTML created in-house, there were parts where the standard features of Memsource did not integrate well with machine translation.

[Translation Assets]

・Translation Memory: 120,000 segments

- UI Glossary: A UI glossary exists for each product group

Implementation Process

After participating in the seminar in July 2017, we conducted meetings and email discussions for specification design, and from September, we implemented the following processes.

・Memsource Plugin Development (for Google Neural Machine Translation Integration)

Glossary Application

Style Processing

I attended a seminar on human science and learned that it is possible to customize the output results of machine translation, and I consulted individually. The key was whether we could solve the challenges that arise when integrating Memsource with Google Machine Translation.

Additionally, as an added effect, we were considering whether we could apply our unique glossary and style. We had direct meetings with the developers, and while clarifying the specific requirements, we proceeded by exchanging sample files to check the output results of the machine translation. Even after starting the test operation, we provided feedback on improvement requests. Those that were technically feasible were improved upon.

(WingArc 1st Product Management Department Documentation Services Team Representative)