UCL academics win Best Paper at Cognitive Linguistics Conference

UCL researchers have won one of the best paper awards at the 2024 Cognitive Modelling and Computational Linguistics (CMCL) conference.  

Their paper, “How can large language models become more human?”, focuses on improving large language models (LLMs) by enhancing their ability to process complex language structures, making them more human-like. 

The research, led by Professor Mehrnoosh Sadrzadeh (UCL Computer Science), Daphne Wang (Quandela), and UCL Linguistics professors Wing Yee Chow, Richard Brehany, and Miloš Stanojevic (Google DeepMind), addresses a major challenge in LLMs.  

While models can predict individual words and phrases, they often struggle with more complex sentences, such as “garden path” sentences. These are sentences like “As the woman read, the magazine entertained the editors” that lead readers to incorrect interpretations. 

Human understanding goes beyond word-level prediction and involves grasping deeper structures such as syntax, semantics and pragmatics. 

To bridge this gap, the team used sheaf theory in conjunction with GPT-2, enabling them to access and use hidden structural information and improve its performance.  

Their experiments showed that the model not only processed difficult sentences better but also predicted sentence complexity, bringing its capabilities closer to human comprehension. 

Professor Mehrnoosh Sadrzadeh said: “Sheaf theory is a novel 20th-century branch of mathematics, developed to combine statistics and structure. In this capacity, it has been used in a variety of other fields from signal processing to quantum mechanics. It was lovely to see that it can also help us develop more humane AI.” 

The paper’s findings could pave the way for LLMs that are not only more precise but also more intuitive in understanding language at a human level, transforming how we interact with AI technologies.