#Japan #Tokyo #Yaraku #Translation Technology #Quality Estimation

Introduction

Yaraku Corporation, a Tokyo-based company specializing in generative AI-powered translation tools, recently celebrated a significant achievement as their Natural Language Processing (NLP) team successfully published a research paper in a renowned domestic journal, "Machine Translation," issued by the Asia-Pacific Association for Machine Translation (AAMT). This paper, titled "Quality Estimation Reranking for Document-level Translation," discusses an innovative approach integrating multiple translation options evaluated by a separate AI model. The findings highlight how this method effectively increases the naturalness and consistency of translations across entire documents.

Background and Novelty of the Research

The focus of the published paper is on a process known as Quality Estimation (QE) reranking, which allows for the selection of superior translations from a pool of multiple candidates produced by translation AI. Traditionally, QE has been applied to single sentences, limiting the scope of evaluation. However, this research marks a notable advancement by conducting comprehensive assessments that encompass entire documents, recognizing the interconnectedness of multiple sentences. Various evaluation methods were utilized, including:

• - Context Understanding: Using the SLIDE model, an extension of the Comet family for document-level evaluation.
• - Naturalness Assessment: Leveraging Comet, a neural network-based evaluation tool to judge fluency and coherence of translations.
• - Evaluation Through Large Language Models (LLM): Implementing GEMBA-DA as a comparative measure.

The study revealed that an increase in the number of translation candidates significantly correlates with an improvement in quality, showing continued enhancement up to 32 candidates. This suggests a promising avenue for introducing QE reranking into translation engines to boost performance.

Significance of the Research

While the study does not directly correlate with the current capabilities of Yaraku's translation tool, it substantiates the technical direction pursued by the company. The study supports:

• - Improving overall naturalness and coherence of document-level translations.
• - Designing systems to choose the optimal translations from a range of alternatives.
• - Employing objective, data-driven approaches rather than subjective judgment to enhance translation quality.

Moreover, Yaraku continues to evolve their machine translation (NMT) offerings, integrating LLMs like GPT and Claude to deliver more contextually accurate and natural translations.

Implications for Yaraku Translation Users

For users of Yaraku's translation services, the results of this research hold significant implications:

• - Validation of Technical Direction: The academic foundation for Yaraku's emphasis on improving document-level quality is reinforced by this study.
• - Foundation for Continuous Product Improvement: By developing the translation model based on scientific validation, a solid foundation for future quality enhancements has been established.
• - Enhancing Consistency in Corporate Documents: The research illuminates pathways to improve inter-sentence consistency in essential documents, such as contracts, manuals, and investor relations materials.
• - Transparency in Data-Driven Development: This demonstrates a commitment to advancing quality through objective metrics rather than personal intuition, aiding in the evaluation of service reliability.
• - Long-term Assurance for Users: The ongoing investment in translation technology and research infrastructure provides users with confidence regarding the long-term viability of the service.

Overview of the Paper

• - Title: Quality Estimation Reranking for Document-level Translation
• - Journal: Machine Translation, No. 83, published on November 29, 2023 (Publisher: Asia-Pacific Association for Machine Translation)
• - Main Content: The assessment of QE reranking at the document level, reevaluation of NMT-generated candidates through QE, comparisons of various methodologies like SLIDE, Comet, and GEMBA-DA, confirmation of quality enhancement through automatic evaluation, and computational efficiency for practical applications.
• - Authors: NLP Team at Yaraku Corporation: Krzysztof Mrozinski, Minji Kang.

For further details, you can explore the AAMT Journal, where the relevant paper is published on page 50.

Company Overview

• - Company Name: Yaraku Corporation
• - Location: 16th Floor, Link Square Shinjuku, 5-27-5 Sendagaya, Shibuya, Tokyo
• - CEO: Yu Sakani
• - Business Focus: Development and provision of "Yaraku Translation," a translation support tool utilizing generative AI.
• - Official Website: https://www.yaraku.com