Everyone is welcome!

Who: Nikita Soni

When: Monday, November 11 at 12:00 PM

Where: NCS Room 120

Title: Human-Centered Large Language Modeling

Abstract: This thesis investigates human language understanding by integrating the human context of the language generator, i.e., who is speaking, where and in what situation, when they are speaking, and to whom it is addressed. For instance, a person feeling exhilarated on a hike might complete the statement I am feeling.. quite differently than they would when they are feeling dejected during a break-up. Factors such as demographics, personality, modes of communication, and emotional states have also been shown to play a crucial role in NLP models pre-LLMs (large language modeling) era. Advances in language modeling yielded in Transformer-based LLMs as the base of most current NLP systems.

However, traditional language modeling views words or documents devoid of the aforementioned human context. To address this, we have taken the first steps of mathematically defining the inclusion of human context in language modeling, and empirically comparing the effects of including different types of human contexts in language modeling on downstream tasks. So far, integrating human context into LMs has shown promise, and to realize its full potential we propose well-established empirical foundations for designing and benchmarking these human context-aware language models. This work will serve as a crucial resource in furthering the sub-field of human-centered large language modeling, providing a repertoire of datasets, human language modeling techniques, and a benchmark to evaluate against.