CommercialAI/MLGlobalLinkTechTranslation/Language

TransPerfect MT Earns Top Marks at International Machine Translation Competition

TransPerfect MT Earns Top Marks at International Machine Translation Competition

As machine translation (MT) technology continues to evolve, it’s increasingly important to look beyond the commercial sphere for the latest in innovation. At TransPerfect, we place special emphasis on the importance of engaging with the research communities specializing in MT—which is why we actively participate in academic conferences, collaborative projects, and translation competitions.

Recently, TransPerfect stood out among the competition and achieved high marks ahead of the Sixth Conference on Machine Translation (WMT21). The full event is scheduled for November 10–11.

Sixth Conference on Machine Translation and Competition

The event will feature presentations of scientific papers on a range of topics related to MT. Included in the event was a shared task translation competition held prior to the main conference.

In the shared task competition, research groups were invited to tackle specific translation challenges from an array of themes. These themes included news, biomedical, European low-resource languages, and more.

Using proprietary data or data sets provided by the WMT organizing committee, participants designed MT-based solutions to the challenges in the shared task description. The results were published on the conference website.

TransPerfect’s Winning Strategy

TransPerfect leveraged its specialization in MT for content from life sciences and clinical research organizations in the biomedical shared task. This task aimed to use automatic and human evaluation methods to determine how well different MT systems translate abstracts and summaries of proposals for animal experiments.

This year’s shared task addressed eight language pairs; TransPerfect participated in the English into Spanish and Spanish into English combinations. The results, published on August 5, demonstrated the power and precision of our MT models. We placed first for the English into Spanish language pair and second for the Spanish into English language pair.

Competition Results

Translated texts produced by participant MT systems were evaluated using the BLEU score, a metric commonly used for assessing MT output quality. The biomedical texts translated from English into Spanish by TransPerfect MT achieved a notably high maximum BLEU score.

Our results placed us well above the nearest competitor’s score. Further, the maximum score for the Spanish into English language pair placed us just behind the top competitor’s score, earning us a close second-place finish.

The first- and second-place rankings reflect our expertise in life sciences, along with our robust MT engines and the quality of the data used to create and regularly update them.

Other Recent MT Successes

Our results at the 2021 WMT translation competition were similar to our earlier success this year at the Eighth Workshop on Asian Translation (WAT 2021), where we participated in the patent translation shared task.

For this competition, which was aimed at evaluating the quality of machine-translated texts from the patent domain, task-specific data sets were provided. A single base transformer model was created for each of the English into Japanese and Japanese into English language pairs. Through this model, the output quality of our translations earned one of the top four placements for both language combinations.

In contrast to other participants, our results in the WAT competition were particularly notable because they didn’t employ sophisticated engine training techniques. This underscores the high quality of TransPerfect MT systems for translating legal and patent content. It’s a promising result for an exercise aimed at benchmarking the performance of our models against the top research systems in the field.

Key Takeaways

The shared tasks competitions are important points of access to cutting-edge knowledge and research in the field of machine translation and natural language processing. Additionally, the shared task competitions offer a unique opportunity to evaluate how our MT stacks up against other systems.

Our participation in these competitions challenges us to innovate and expand the way we approach MT, and the results serve as a reflection of our commitment to continuous improvement of our MT solutions.

To learn more about how this commitment also extends to the technology and services we provide to our clients, please reach out to us here.