Introduction
How Neural Machine Translation (NMT) engines can be customized to handle large volumes of material and client-specific corpora, improving quality while saving time and reducing costs.
One of Propio’s largest clients needed a high volume of regulated material translated on a cyclical basis from English into seven languages: Hindi, Japanese, Korean, Russian, Spanish, Thai, and Vietnamese.
The client required expedited turnaround times and reliable, tested output to ensure the materials were completely accurate for a multilingual audience.
The Challenge
After assessing the situation, it was clear that several hurdles would need to be overcome to successfully translate the high volume of materials and meet the critical deadlines of each cycle.
One difficulty was that the client-specific corpora needed to be translated accurately and quickly. Corpora can include anything from company logos to product names and terminology used exclusively by the client. Other in-market engines were incapable of deciphering not only the corpora, but also more complex, nontraditional languages such as Vietnamese.
To meet the turnaround requirements and ensure the quality of the output across languages and corpora, it was determined that a customer-specific NMT engine would need to be created.
Propio’s dedicated NMT team gathered all necessary industry-specific and client-specific content before completing the following steps: data cleaning and preparation, training, and multiple levels of testing. These phases were completed before the customized engine was released for use within the translation workflow.
The Solution
To create a customized engine, a minimum of 15,000 clean segments is needed. Because Translation Memories (TMs) usually contain a significant number of repetitions and segments consisting only of figures or unusable text, approximately 30,000 TM segments provide a good starting point.
Propio’s NMT team followed the standard engine-customization process:
- Identify and extract eligible existing training material
- Perform a corpora-cleaning process using proprietary and third-party technologies
- Generate an acceptable set of clean, usable, and aligned segments to train the customized engine
Proven Results
Propio compared the BLEU scores of in-market engines with the NMT engine custom-built for the client. The results illustrated a significant quality disparity between in-market and customized NMT engines.
In live production, the customized engine achieved the following BLEU scores:
The increase in BLEU scores was significant and allowed Propio to elevate more difficult languages, such as Russian and Thai, to an appropriate quality level for production use.
In addition to the quality improvements, the customized NMT engine delivered significant client benefits, including reduced turnaround times, a reduction of more than 38 percent in time to market, and overall cost savings.
