Reduced translation costs

Greater productivity and the ability to generate content at greater scale

Improved translation quality — text that appears to have been written by a human

Enhanced customer experience

New opportunities to enter more markets

How do generative AI and LLM systems work?

Generative AI and LLMs like GPT are AI engines that have learned how humans write text. They’ve been trained on the massive corpus of the internet. Give the model an input, and it will produce the most plausible output from its extensive training.

What type of content is best suited for LLMs?

It is appropriate to use LLMs for all types of content when engaging in the following tasks: editing, error correction, management of style and tone, terminology adherence, and clarity of expression. 

Use LLMs when your content requires fluency to produce reach and engagement. Using LLMs for content requiring precise translations is riskier as this technology can produce inaccurate information.

How do LLMs compare to Neural MT (NMT) engines?

We discovered that OpenAI’s GPT-4 model can produce better translation results than Yandex in certain situations for the English-to-Chinese language pair. This achievement is a significant milestone. However, GPT-4 doesn’t yet deliver the same outcomes as the established NMT engines (Microsoft, Yandex, Google, DeepL, and Amazon) and can be 12-50 times more costly.

Older GPT models, such as ChatGPT and GPT-3, are less advanced than GPT-4, and their translation performance is inferior to the major, specialized NMT engines.  We found some issues with agreement and concordance of gender that are not present in NMT outputs.

Is generative AI technology ready for widespread professional translation?

LLMs are in great demand, but they cannot yet cater to the volumes of content needed for industrial-scale localization. We expect this situation to change, but the timing is uncertain. Until such time, it will take longer to process translations and generally cost more to use LLMs for translations instead of NMT engines.

Is it safe to use LLMs from a security standpoint right now?

It is imperative to review and understand the terms of use and privacy policy of the LLM you intend to use, especially free offerings. Typically, when using free technology, your information is the price you pay. 

The National Cyber Security Centre  outlines some risks and recommends that you refrain from providing sensitive information to public LLMs and avoid submitting queries to public LLMs that could harm your company should they become public.   

Lionbridge goes to great lengths to protect its customers’ information when using automated technology.

Should I select LLMs or NMT engines for my translations? 

The pace at which LLMs are improving suggests these AI systems are advancing the Natural Language Processing (NLP) field and will be part of a new paradigm as the Neural MT paradigm ends. Nonetheless, it is too early to entirely dismiss the major MT engines for automated translation. 

During this transition period, we advocate using a hybrid model combining the benefits of both NMT and LLM engines rather than selecting one model at the exclusion of another. 

A promising area of our research demonstrates the benefit of a hybrid model that uses an NMT engine for translations and then leverages an LLM engine for post-editing. This approach improves the fluency of the NMT output since LLMs have the uncanny ability to produce text that appears to have been created by humans.

What can we expect to happen when generative AI matures?

Generative AI has the potential to resolve some longstanding quality issues associated with Neural MT engines, including its inability to attain and consistently achieve the right formality level, tone, or handle negation.

As generative AI and LLMs mature, expect:

A significant leap in MT quality, including enhanced workflow automation

Increased global content generation and output 

Greater adoption, and 

Enhanced customer experience through MT use

What type of generative AI services can Lionbridge provide right now? 

Lionbridge can — and has — provided the following LLM-related services for its customers:

Multilingual text generation / transcreation

Prompt engineering / multilingual prompt engineering 

Prompt creation, translation, transcreation, translation review, analysis, and testing

Response evaluation and validation

Diversity, equity, and inclusion initiatives, including identification of stereotypes, biases, or problematic content in the training corpus, prompts, and responses

Reinforcement learning from human feedback (RLHF)

LLM evaluations

Data-related services, such as data annotation and data cleaning for LLMs

Multilingual asset optimization

Model fine-tuning / customization

Workflow automation 

Our Research & Development (R&D) team continuously investigates new ways to deploy generative AI as it evolves.

Will generative AI replace the need for Language Service Providers or professional translators?

No, generative AI will not replace Language Service Providers (LSPs). Since its inception over 25 years ago, Lionbridge has embraced, become an expert on, and fully leveraged emerging technologies to provide top-notch language services. Our ability to adapt to technological change has driven our growth, longevity, and success. When assessing LSPs, it is imperative to investigate the supplier’s ability to capitalize on evolving technologies.

During this transition period, it may be premature to change the usage of your current MT engines drastically; our vast MT experience makes us suited to help you identify the right moment to perform that change and provide guidance on execution.