---

A Case Study

The client

Accuray Incorporated is a pioneering radiation oncology company that specializes in the development, manufacture, and sale of innovative radiation therapy treatment solutions. The company is renowned for its commitment to expanding the curative power of radiation therapy, helping clinicians treat patients more efficiently and effectively.

The Accuray portfolio includes innovative technology like the CyberKnife and TomoTherapy Systems, which have set new standards in the field of cancer and neurological disease treatment.

Accuray has an expansive global reach, continually striving to make treatment shorter and more personalized. Their ground-breaking approach to radiation therapy ultimately aims to enable patients to live longer, better lives.

Since 2018, Accuray has been building a professional collaboration with Argos. Argos’s key services include precisely translating technical documentation and software strings, catering to over 25 languages globally.

The challenge

Accuray faced a significant challenge in maintaining the quality of their translation memories (TMs), which they heavily rely upon. Over the past two years, they achieved an impressive average of 80% reuse from these TMs.

However, in collaboration with Argos, potential quality issues were identified within these legacy translation memories. Accuray’s commitment to high standards led them to confront this issue head-on.

Accuray needed a comprehensive clean-up of its TMs in five of its most frequently used target languages: French, German, Italian, Spanish (Spain), and Portuguese (Brazil).

The intended outcome

Accuray’s primary objective was to enhance the overall quality and efficiency of their TMs in the target languages through a comprehensive cleanup project. This involved removing outdated content to streamline the TMs, thereby improving their functionality.

Additionally, Accuray aimed to use the refined TMs as a foundation for training custom machine translation (MT) engines.

This strategic move was intended to generate further cost savings over time by leveraging high-quality MT.

What we did

Argos developed a custom solution based on AI technology that evaluates each segment using a trained large language model (LLM) and multilingual vector distance. Using a self-hosted machine learning instance, we can process the data within a secure environment. The application provides an interface that displays at a glance a “TM Health” view for each translation memory. Using this approach, we align with Accuray’s goals.

Initially, a metadata-based method was used to identify and eliminate outdated and unused portions of the translation memories, removing 50% of the total TM volume with acceptable leverage loss. This step allowed us to focus on non-legacy segments, avoiding unnecessary costs.

Accuray also reviewed their existing glossary terms and added over 700 new ones. To expedite the process, the AI TM cleanup was divided into two phases. The first phase addressed issues detected by AI, consistency checks, and targeted regular expressions. The second phase began when the new glossary was ready and tackled false positives in glossary checks.

Upon completion of both phases, the cleaned TM was merged with the delta from parallel translation projects. The cleaned TMs were then used to fine-tune MT engines for improved output quality, which was confirmed through blind human evaluation tests.

The MT engine is currently updated monthly with new TM content, with performance tracked in real time to prevent any deterioration in the TMs. The project highlights the importance of a custom approach, clean terminology for MT fine-tuning, and real-time performance tracking.

How it’s going

The implementation of the custom solution yielded remarkable results, particularly in the improvement of MT.

A blind human evaluation was conducted to assess the quality of the untrained and trained engines. The untrained engine scored 3.8 out of 5, translating to a 76% quality rating. However, after training, the engine’s score soared to 4.6 out of 5, indicating a substantial increase to 92% quality. This represented a 16% quality increase in MT content output, a testament to the efficacy of the solution implemented.

Furthermore, there was a significant reduction in terminology issues by 77%, demonstrating the solution’s effectiveness in enhancing the precision of translations. These statistics reveal the profound impact of the implemented solution on improving the quality and accuracy of MT at Accuray.

Callout stat: 77% reduction in terminology issues