1. Language IO
  2. Getting Started
  3. Understanding the Machine Translation Pipeline

Articles in this section

See more

Large Language Models for Translations

Language IO integrates Large Language Model (LLM) capabilities into the translation pipeline. 

While many LLMs with translation capabilities have hit the market, Language IO completed extensive testing to curate the highest quality models for customers with specific translation needs. LLMs are built into Language IO's MT vendor stack alongside previous NMT models, that all passed quality tests of similar rigor. Like NMT models, LLMs have their own strengths and capabilities that vary by language pair. For this reason, it is best to have multiple models available for your use case, and a solution that selects the best model for each transaction dynamically.

Superior Translation Quality and Fluency

The introduction of LLM-based engines marks a significant leap in "out-of-the-box" quality.

  • Natural Phrasing: LLMs consistently outperform NMT models in syntax and fluidity, and align more closely with native-speaker expectations.
  • Contextual Intelligence: LLMs demonstrate contextual resilience: they can identify source typos and offer a better management of social registers and honorifics.
  • Localized Formatting: LLMs render dates, times, and currencies more accurately according to specific target locale conventions.

Advanced Glossary and Morphology Handling

LLMs transform how we manage terminology, significantly reducing the manual burden on Language IO teams.

  • Reduced Glossary Maintenance: LLMs understand domain-specific context so well that they often select the correct industry term without the need for an explicit glossary entry.
  • Morphological Adaptation: Unlike NMT models that use literal substitutions, LLMs can adapt glossary terms to fit the grammatical rules of the target sentence with 90% accuracy.
  • Simplified Management: Teams now only need to provide the base form of a glossary term; the model intelligently adapts the output to the grammatical context of the target.

Massive Scale and Formality Support

The introduction of LLMs increases our linguistic reach and stylistic control exponentially.

  • Expanded Language Coverage: Translation support for up to 140 more languages (depending on the LLM) will be made available gradually after the release.
  • Granular Formality Control: LLMs support formality settings in significantly more locales than NMT models (up to 90 additional languages depending on the LLM).
  • Variant Precision: LLMs show superior performance when it comes to respecting regional variants (for example Latin American Spanish or Canadian French).

Operational Efficiency and Policy Consistency

  • Data Security: Language IO's zero-trace agreement remains in place. Customer data is never used to train these models. This ensures full compliance with existing privacy policies.
  • Glossary Fidelity: Testing shows 99% consistency in glossary term imposition. This leads to higher usability scores compared to NMT models.

Custom Prompting

  • In addition to enhanced translation quality, Language IO LLMs offer a powerful Custom Prompting feature. This allows internal teams to override default settings and resolve complex, non-standard customer issues. 
  • By prioritizing targeted prompts over generalized ones, the Machine Translation Quality team gains direct control to experiment and optimize translation quality for the customer without having to wait for lengthy engineering cycles. This eliminates the need for manual workarounds, and reduces internal escalations for unique edge cases. 
  • As a result, customers receive faster, more accurate resolutions tailored specifically to their unique challenges.

Related articles