Summarization with Latent Structure, Context Factors and Quantitative Precision

Large language models (LLMs) have demonstrated remarkable results in NLP, achieving significant progress in many tasks, including document summarization. I will show through three avenues how we can support LLMs to provide summaries that better suit user needs. I will first show how latent structure can improve the summaries when we require synthesis of different parts of the input document to achieve a reasonable summary. I will then show how we can consider context factors when summarizing, as they are often latent but significantly affect the desired outcome. I will continue by showing that “verifier models” can mitigate the well-known problem of imprecision with generated text, focusing on the accuracy of specifying quantitative information. I will conclude with thoughts about where we are going with LLMs, and what we can do better.

Shay Cohen is a Reader at the University of Edinburgh (School of Informatics). Before this, he was a postdoctoral research scientist in the Department of Computer Science at Columbia University and held an NSF/CRA Computing Innovation Fellowship. He received his B.Sc. and M.Sc. from Tel Aviv University in 2000 and 2004 and his Ph.D. from Carnegie Mellon University in 2011. His research interests span a range of topics in NLP and machine learning, with a focus on structured prediction (for example, parsing) and text generation.