Inside the Potential of Large Language Models in Localization at SlatorCon Zurich 2023

LanguageWire Large Language Model

In a panel discussion at SlatorCon Zurich 2023 moderated by Slator’s Head of Research, Anna Wyndham, Roeland Hofkens, Chief Product and Technology Officer at LanguageWire, Fabian Fahlbusch, Head of Content at Busch Vacuum Solutions, and Alex Coope, Director of Globalization at ServiceNow, explored the practical applications and implications of large language models (LLMs).

The panelists emphasized the importance of customization when working with LLMs. More specifically, Hofkens explained that, while these models are “very powerful”, they can be “quite generalistic.” To make LLMs more tailored to specific corporate languages, companies can use their existing linguistic assets, such as translation memories and termbases.

“Use these assets to customize your LLM,” said Hofkens. This customization enhances the accuracy and appropriateness of LLM-generated content, and techniques like Retrieval Augmented Generation and weight customization can be employed for this purpose.

Fahlbusch emphasized the role of linguistic assets in improving translation quality, especially for technical content, as well. “It is very important for us to have our terminology respected,” he said.

Data Security Is Key

Data security and privacy when using LLMs were addressed in the discussion. With the proliferation of data privacy regulations, such as the European Union’s General Data Protection Regulation (GDPR), companies must be diligent in safeguarding customer data. This includes implementing robust data encryption and ensuring data segregation to prevent unauthorized access.

The panelists highlighted that LLM providers are working diligently to meet these privacy and compliance requirements, and organizations must be careful to ensure that these safeguards are in place. Hofkens recommended choosing an LSP that controls the data life cycle, ensures no data retention, uses transient data storage, and operates within the EU for GDPR compliance.

“Finding an LSP that can guarantee you these aspects is crucial” he said. He also mentioned that running LLMs in-house is an option, but it can be challenging due to the considerable size of these models. 

Quality Estimation

Additionally, the panel explored the quality estimation capabilities of LLMs and their impact on traditional translation workflows. Hofkens highlighted that quality estimation with LLMs is “much more powerful” today. This allows for increased use of raw machine translation (MT) output in various cases, reducing costs, and minimizing human review.

Coope and Fahlbusch emphasized that while LLMs can enhance quality and reduce translation costs, it is still essential to maintain a human review, especially for certain types of content, such as new product descriptions, technical manuals, or legal contracts.

Fahlbusch explained that at Busch Vacuum Solutions all translations undergo a human review conducted by local subject matter experts in the respective countries. Looking ahead, the company envisions a future where raw machine translation output will be the norm, with in-country reviews. Human translators will only be involved in reviewing MT for exceptionally important legal translations.

Coope emphasized that LLMs will not replace human translators but instead will be tools that translators can use to enhance and specialize their work. He also believes that translators would play a crucial role in refining the quality of LLMs. He sees LLMs as “off-the-shelf starting points” that companies can fine-tune with the help of translators. However, Coope emphasized that this transformation will not happen immediately and might take 5 to 10 years.

The Real Disruptor

The discussion also touched on LLMs in content generation. Fahlbusch recognized the value of LLMs in generating content like press releases and social media posts but stressed the irreplaceable role of human expertise in specialized and technical content.

For Hofkens content generation is the “real disruptor” brought about by LLMs. He explained that LLMs, when fine-tuned with domain-specific data, can generate drafts that may not meet production quality but can significantly increase the productivity of content creation.

Given that this LLM-generated content will not reach human-level quality, he emphasized that this can present an opportunity for a review step to fact-check and address stylistic errors.  “So, maybe also something for the LSPs and the translator communities to offer,” he said. Hofkens also expects that in the near future content will not be created entirely from scratch, much like the way translation workflows have evolved with artificial intelligence (AI).

In the Q&A session, questions about the risks of prompt injection emerged — i.e. the malicious manipulation of an LLM’s input to generate harmful or misleading content — and the potential misuse of LLM-generated content. 

The panelists highlighted the need for clean data and data validation to mitigate these risks. By carefully curating the input data and incorporating mechanisms for verifying the generated output, organizations can reduce the chances of LLMs producing undesirable or false information.

For those who could not attend in person, SlatorCon Zurich 2023 recordings will be available in due course via our Pro and Enterprise plans.