10 Ways Translation Technology is Increasing Productivity and Quality

Translation tools are more accessible than ever, and users, including translators, can now access machine translation (MT) via web browsers, applications, plug-ins, APIs, and even directly on the user interface of some smartphones. MT can now also be integrated with more systems and applications, including AI-enabled automation.

Based on Slator’s continuous coverage of translation technology research and advancements in the field, there is widespread agreement among natural language processing (NLP) researchers that large language models (LLMs) are significantly improving their translation performance. 

Better training data, improved fine-tuning methods, and expanded context-aware translation are some of the improvements made. While computer-assisted translation (CAT) tools and translation management systems (TMSs) based on neural machine translation (NMT) are still standard in the translation business, they are being increasingly automated and transformed with LLMs.

Enhancing Machine Translation with AI

MT engines have improved in accuracy and nuance, and as LLMs improve as well, tools are increasingly including a mix of NMT capabilities and artificial intelligence (AI) assisted features. A human review, or machine translation post-editing (MTPE), is still recommended as these language technologies evolve to ensure top quality.

These are some of the innovations and features available in NMT and LLMs that can increase productivity and quality in translation:

1. Ability to adapt to different styles. LLMs can be fine-tuned to recognize different writing styles and registers (e.g., formal vs. informal), a critical element in translation. Also, the ambiguities inherent in language are better handled by LLMs, given their broader understanding of context.

2. Adaptable translation memory (TM). An adaptable TM can be continuously improved and fine-tuned based on new translations or specific domains. This allows the NMT plus LLM system to learn from these improvements and adapt translations over time.

3. Context-aware translations. LLMs analyze language patterns and work by predicting the likelihood of a particular word appearing in a specific context. They can also analyze the surrounding text of a sentence, which helps them understand the meaning and intent behind the words. Training with annotated data (data with added labels or tags to follow certain guidelines) also helps LLMs to better understand and apply cultural context in translation.

4. Continued learning. In translation, once a human corrects the target language, the model learns and improves its performance. LLMs can also learn and handle novel content and rare phrases (which can be a challenge for traditional NMT), and can check their own translation output and change or improve it.

5. Enhanced fluency and accuracy. LLMs excel at understanding context and natural language nuances. In combination with curated translation memories and NMT, LLMs can produce translations that are more grammatically correct, idiomatic, and read more naturally. The combined approach can significantly reduce the time and effort required for MTPE.

6. Faster audio-visual translation. Real-time translation of long stretches of spoken language can be challenging for traditional translation models, but LLMs are capable of segmenting audio input into smaller, more manageable chunks, and at great speeds.

7. Glossary integrations. Glossaries can now be used for training and fine-tuning LLMs or can be added to some NMT-based translation tools to improve the target output. LLMs can prioritize terms, which increases consistency and adherence to specific jargon, or across languages and projects. This helps ensure voice and style preservation, along with technical accuracy.

8. Better scalability. Combined NMT and LLM systems can handle much larger volumes of text in a more efficient way. For example, when an approved glossary is used in model training, the LLM is better at handling consistent terminology across documents and projects. It also handles string localization in context better by adhering to specific patterns.

9. Improved domain-specific performance. LLMs can be trained on datasets specific to a particular domain (e.g., legal, medical, technical) and language pair. The specialized knowledge can be leveraged in combination with NMT for improved translations within those domains.

10. MT model customization. It is possible to customize translation models offered by large tech companies, such as Amazon Translate, Google AutoML, IBM Watson Language Translator, and Microsoft Custom Translator. Users can also train LLMs in a myriad of ways for specific purposes.

With the arrival of translation tools that can machine-translate entire documents (e.g. CotranslatorAI) and orchestration platforms that combine multiple translation and translation management tools in a single interface (e.g., Phrase Orchestrator and Blackbird.io), tools are expected to continue evolving rapidly. 

The combination of LLM and NMT technologies can already capture context and idiomatic expressions, perform automatic translation reviews, and approach professional human translation quality levels in dominant language pairs. Work remains to be done for low-resource languages, but an increasing number of researchers are dedicating time and resources to improving models for those languages.

The following table includes some tools that make use of the latest translation technologies mentioned above. This is not a comprehensive list, but a small sample of the features that have become available in these translation tools since 2023 or earlier.