Words at work: New directions in natural language processing with Ted Briscoe

Throughout his career, Ted Briscoe has focused his efforts on solving fundamental problems in natural language processing, such as parsing and syntactic and semantic analysis. But Briscoe’s contributions have not only been scholarly. He has also built applications and companies, and today his work continues to sit at the nexus of both theory and practice.

After more than three decades at the University of Cambridge, Briscoe joined Mohamed Bin Zayed University of Artificial Intelligence last year as deputy department chair of natural language processing and professor of natural language processing. His current interests relate to three projects that seek to advance the field of natural language processing and provide benefits to users.

The written word

Earlier in his career, Briscoe worked for a company called Cambridge Assessment that developed an application that used artificial intelligence to grade exams and provide feedback to students. He is continuing this direction at MBZUAI in the form of an educational technology initiative that is designed to support literacy for Arabic speakers in the Gulf who are interested to improve their written Modern Standard Arabic.

The system provides feedback on students’ writing and helps to improve their facility with written Arabic by giving automated and personalized feedback. The system, which is being built in a collaboration with IBM, is responsive to the unique needs of individual users as it processes student writing and then generates questions and activities that are specifically suited to help them improve. The project was recently described in MBZUAI’s annual research magazine, The Node.

Foreign office

Another initiative is a collaboration with Abu Dhabi Global Market, a financial center and free trade zone in the capital. This project seeks to build a question-answer system to assist companies who are interested in establishing an office in Abu Dhabi but need to understand local rules and regulations.

The practice of establishing foreign offices has typically been reserved for corporations that have compliance and legal teams with resources that allow them to stay up to date on the ever-evolving landscape of regulations in different jurisdictions. Not all businesses can afford to hire dedicated legal and compliance staff, however, as paying experts to keep track of mercurial rules and regulations is hugely expensive. The application Briscoe is developing for Abu Dhabi Global Market could help smaller companies establish an office in the emirate more affordably and enter a market that was previously out of reach.

Developing a question-answer system designed specifically for regulations is challenging because the results must be accurate, and the system must be able to continually account for new information.

Question-answer systems are built on large language models that are trained on huge volumes of text. “One of the problems with LLMs is that while they know a lot from reading huge amounts of text, their knowledge is very diffuse, in that they know small nuggets of information which they then plug together probabilistically,” Briscoe explained. It’s this method of building sentences probabilistically that results in LLMs creating sentences that mix fact and fiction, or what’s known as hallucination.

To address this challenge, Briscoe and his team are integrating a method called retrieval-augmented generation, which combines the generative ability of language models with information retrieval systems. With retrieval-augmented generation, when a question is asked, the system conducts a search to find relevant passages in source documents — in this case, they’re often PDFs that are hosted on the websites of government ministries. Those source documents are then fed into the LLM and are used to construct an answer.

“The input to the LLM isn’t just the question, it’s also the passages that have been filtered out from the document collection that have been determined to be relevant to that question,” Briscoe said. “We’re asking an LLM to do a smaller problem in that we are providing it the relevant information and telling it to base its answer on only that information,” and not the whole body of text it has trained on.

“It also helps us respond to the dynamic nature of the information, as rules and regulations often change and the data set for this application isn’t static,” he added. The Copilot feature of Microsoft’s Bing search provides similar functionality, Briscoe noted, in that it provides an answer like a chatbot but will also share the relevant links from which it drew the information.

The idea is that building responses based on specific subsets of material will reduce hallucination and improve the accuracy of the system.

Modeling language and meaning

While less tangible at its current stage, the third direction of Briscoe’s research is perhaps the most ambitious. He is working to develop natural language programs that can not only generate text but also understand meaning in language, an activity they fall short of today. Indeed, hallucination is a symptom of their inability to grasp meaning.

“Although there’s a sense in the research community that what’s wrong with LLMs is that they don’t reason well and hallucinate,” Briscoe said, “they also don’t fully understand the relationship between form and meaning in language.”

As a type of machine learning, language models work according to principles of statistics. When they generate text, they predict what the next word in a sentence should be based on probabilities that are determined by sentences that they have encountered in the past. “LLMs compensate for a lack of real understanding of how meaning is generated in language with a kind of general knowledge,” Briscoe said.

Getting to a passable answer through general knowledge instead of a true understanding of language can happen in interactions between humans as well, Briscoe explained. For example, a tourist using a foreign language may string together a grammatically incorrect sequence of a few words asking a local how to find a restroom. In that situation, the local doesn’t necessarily need to understand in a grammatically correct way what the tourist is trying to say but can glean from the context of the situation what the tourist is trying to find. “In a case like this, the local wouldn’t be processing language in any real way,” Briscoe said, “but would be using a huge amount of background knowledge to try to figure out the tourist’s intent.”

Language models, even the best ones, like OpenAI’s GPT-4, are doing something similar when they give the right answer in that they are making a guess based on context.

To address this shortcoming, Briscoe and his team are initiating a project in which they aim to separate a model’s facility with a language from the knowledge that it has about individual words. They do this by training a model in an artificial language. Under these conditions, the model’s ability to understand how meaning is generated in language becomes clear. This is the case because with an artificial language, the model can’t rely on the “general knowledge that accrues around words,” Briscoe said. “In an artificial language, in order to predict the next token in a sentence, the model must understand the construction of the sentence because it doesn’t have the background knowledge that is the case with languages like English.”

The method shines a bright light on errors that the model makes. By analyzing these errors, Briscoe and his colleagues can identify and catalogue the mistakes. Doing so may help researchers develop new methods for processing language that are more data efficient and can learn full meaning mapping more rapidly, Briscoe said.

“This has implications not just for English, but also for processing all the languages which are in danger of getting left behind by the current generation of technology,” he noted.

The next transformer

The mistakes language models make are manifestations of the underlying shortcomings of the computational architecture that powers them.

The current generation of large language models are built on a neural network architecture known as the transformer, which was first proposed in 2017 by a team of researchers from Google and described in a now foundational paper titled, “Attention Is All You Need.”

Transformers are useful for natural language processing because they can handle what are known as long-range dependencies. This allows them to weigh the importance of each part of a piece of text, regardless of the distance between elements in the text. As a result, transformers can effectively determine relationships between words in a sentence that may be distant but are nevertheless related.

“The transformer is an architecture that very roughly pays attention to everything, in that the relationship between one word and every other word in a sentence is important,” Briscoe explained. “That’s the starting point, the prior, if you like, of a transformer.”

The “paying attention to everything” nature of the transformer is also the reason that it makes mistakes with language that humans wouldn’t make.

While humans can keep track of how words in a sentence that are distant from each other relate, our attention likely isn’t so blunt as it is with the transformer, Briscoe said. “I think we come with a much more specific prior. We are actually much more focused on what to pay attention to in what we’ve heard. We also recode it very quickly into a format where interpreting what’s coming next in a sentence is much more salient than it is in a transformer.”

Briscoe hopes that by gaining an understanding of where the transformer goes wrong, he and his team will be able to gather clues about how to build a new model that is more data and computationally efficient yet can tie attention to context in a meaningful way.