Most enterprise customers possess extensive amounts of valuable data within their translation memories (TMs), accumulated over years of translation work. These TMs are invaluable in providing pre-existing translations for similar or repeated segments. However, a common challenge arises with “Fuzzy Matches,” where a translation is close to the stored segment in the TM but not an exact match. Although most of the information remains correct in a Fuzzy Match, translators still need to manually adjust the parts that differ, which can be labor-intensive.

Faced with this, some clients may bypass Fuzzy Matches altogether and rely exclusively on Machine Translation (MT) to produce potentially correct segments. While MT can generate fully accurate output in some cases, this approach comes with significant risks and costs. First, using MT without leveraging the valuable data stored in TMs can increase MT service consumption, driving up costs. More critically, ignoring TMs risks introducing inconsistencies, as MT systems may not align with the established terminology and stylistic preferences contained in the TM. This can lead to costly quality control issues, undermining the consistency and reliability of translations across projects.

Dealing with high Fuzzy Matches presents a unique challenge:

• Translators must invest time in fixing portions of the Fuzzy Matches that differ from the TM, even though most of the content is correct.
• Relying solely on traditional MT systems can sometimes overlook valuable TM data, leading to terminology errors and inconsistencies with previous translations. This can make it harder to maintain consistency across content, complicating the user experience and the ability to refer back to previously translated material.
• Maintaining consistency between past translations and new segments is critical to ensure brand voice, industry-specific terminology, and linguistic quality, but this is hard to achieve without human intervention.

The ideal solution would intelligently correct only the differing parts of the Fuzzy Match while preserving the high-value content already stored in the TM.

We developed a custom AI solution designed specifically for handling Fuzzy Matches, optimizing the translation process while retaining consistency with enterprise TMs. Here’s how it works:

• The AI uses high Fuzzy Matches as a starting point, focusing on the segment portions that need adjustment while leaving the correct parts intact.
• It references other matches in the translation memories and customer-specific terminology databases to automatically repair the mismatched portions of the Fuzzy segment.
• By modifying only the required sections, the AI ensures that the final segment closely aligns with the TM without introducing inconsistencies or deviating from customer preferences.

The AI-driven process reduces the need for human intervention in correcting high Fuzzy Matches, while still delivering translations that are accurate, consistent, and aligned with prior work.

This AI feature has been tested across various languages, including German, Spanish, French, Italian, and Chinese. Early results show promising improvements in translation quality:

• Between 40% and 80% of Fuzzy Match segments generated from TMs see improvements due to AI intervention, reducing the post-editing workload for human translators.
• Importantly, the AI feature has not introduced significant errors or hallucinations, meaning that it improves overall translation quality without generating new problems.

However, there is still room for improvement. Inconsistent performance across different test sets - despite using the same AI prompts - indicates that the feature is not yet ready for broad implementation. Although most segments are improved, performance variability means the feature requires further refinement before it can be fully rolled out to all customers.

Our goal is to ensure that when this AI feature becomes available to all customers in WordsOnline, it will provide robust, reliable results and deliver the highest quality translations possible.