Facebook Inc. has designed a new artificial intelligence framework it says can create more intelligent natural language processing models that generate accurate answers to questions without being constantly retrained.

Natural language processing, or NLP, refers to the process of teaching computers to understand how to interpret and manipulate human language. It’s one of the oldest challenges in AI research, and it has come a long way already, with models that can fine-tuned to perform lots of different tasks, including analyzing medical texts or responding to customer inquiries.

Facebook says the biggest challenge in NLP today is creating a model that can research and contextualize what it hears and reads. But today it announced it has made “substantial progress” in this area with its new Retrieval Augmented Generation architecture, which is being released as open-source software as part of its Hugging Face transformer library of trained models that can be used for a variety of natural NLP tasks.

In a blog post, Facebook researchers Sebastian Riedel, Douwe Kiela, Patrick Lewis and Aleksandra Piktus said the RAG architecture is an “end-to-end differentiable model” that combines Facebook AI’s dense-passage retrieval system with its Bidirectional and Auto-Regressive Transformers sequence-to-sequence model generator.

The RAG architecture is essentially a kind of knowledge extractor that can generate answers to questions posed to it simply by reading documents available on the internet. It can do so even when the documents it reads only provide clues to the correct answer, without stating it verbatim, Facebook’s researchers said.

“We obtained very strong results on NaturalQuestions, CuratedTrec, and WebQuestions with RAG, demonstrating that state-of-the-art machine reading performance can be achieved with a generative, rather than extractive, reader,” they wrote.

The researchers said RAG excelled when it came to knowledge-intensive “Jeopardy” questions, thanks to what they believe is its ability to synthesize a response using disparate pieces of information drawn from a number of sources.

“With RAG, we control what it knows simply by swapping out the documents it uses for knowledge retrieval,” the researchers said. “We tested this behavior by replacing our original Wikipedia data set with an older one and then asking questions like ‘Who is the prime minister of Iceland?’”

RAG’s responses showed that it was able to adjust its answers based on the new data set. This capability should be invaluable for applications such as AI agents, which need to be able to access vast quantities of information and also determine what is the correct information, Facebook’s researchers said.

That’s a problem for today’s existing trained models since they require constant retraining in order to keep themselves up to date. With RAG, it’s possible to create more adaptive NLP models that can bypass the retraining step by accessing and understanding more up-to-date information.

“We see broad potential for RAG, which is why we released it today as a component of the Hugging Face transformer library,” the researchers said. “With RAG’s inclusion, we believe the community will be able to apply retrieval-based generation to the knowledge-intensive tasks we already explored and some we haven’t even imagined yet.”

Constellation Research Inc. analyst Holger Mueller told SiliconANGLE that AI is all about the battle for the future of automation, and in the case of human interaction that means better understanding of natural language.

“Natural language models are large and complex, and retraining them because the world moves on both from a language and context perspective is slow and expensive,” Mueller said. “Facebook’s RAG contribution to open source avoids the need to retrain models. This makes it very compelling for AI developers building next-generation applications.”

Image: fullvector/Freepik