Here is my opening statement at the AI FOR DEVELOPING COUNTRIES FORUM Winter Summit 2025 at the United Nations Office at Geneva, Switzerland:

‍

"I am Director of LangOps, or Language Operations, at the Philadelphia Church of God and Advisory Board Chair at the LangOps Institute.

I have worked in the translation/interpretation industry for 20 years and traditionally translation is a tedious and expensive process. This is because language operations needs to address the complexity between human resources, tech pubs, DevOps, marketing, customer care and legal and regulatory requirements, the combinations of which I refer to as a multi-lingual 'language factory'.

This complexity, and associated cost, is why only a small percentage of content is translated and therefore only a small percentage of content is accessible, especially in developing countries. The localisation industry has been working with AI for years, before it was hyped as AI, and there are clearly unprecedented opportunities due to the valuable high quality multilingual data provides, but also limitations and threats.

As an example, when we analysed the translation of more than half a million words in the news domain, in EN-ES, a more common language pair, we found that about 1/3 of the generic neural machine translation was untouched by human translators. Another 1/3 needed light edits and another 1/3 needed heavier edits. That means 1/3 of generic neural machine translation was accurate. We are now making significant progress in making the second 1/3 available by applying light edits through context-aware automatic post editing by utilising fine-tuned large language models and advanced prompting.

Then remains solving the final 1/3. Unfortunately, this final 1/3 tends to be the most important 1/3 of content, without which the most important meaning is lost.

This is where human-at-the core workflows are essential for most high value content. Our current research is working to solve for that through domain-specific structured knowledge in the form of knowledge graphs, utilising semantic techniques such as multilingual ontologies and language-specific taxonomies, to transform unstructured knowledge into structured knowledge.

These knowledge graphs will then be used in Retrieval Augmented Generation workflows. If you are interested in solving for multilingual communication, at scale, speed and quality, please come and talk to me."

It was pleasure to present this with our moderator Sachin Chauhan and along side Dr Nici Sweaney, Senthil M. Kumar, Rupert Douglas-Bate and Arun Raste. They are working on some incredible projects.

Grateful credit to Adam Bittlingmayer (ModelFront) for providing the general assessment of the thirds above; knowledge graphs research with Prasad Yalamanchi (Lead Semantics) and the term 'language factory' comes from Jochen Hummel (ESTeam AB).

‍

Contact us to talk about consultation services