Abstract

Recent years have witnessed increasing interest in developing interpretable models in Natural Language Processing (NLP). Most existing models aim at identifying input features such as words or phrases important for model predictions. Neural models developed in NLP , however, often compose word semantics in a disentangled manner. As such, interpretation by words or phrases only cannot faithfully explain model decisions. In our recent work, we propose a series of disentangle representation learning methods for interpretable language models, including the interpretation in text classification with hierarchical explanation, the uncertainty estimation in prediction, and controllable language generation with disentanglement. Experimental results on real world datasets show that our proposed approaches are able to generate interpretations more faithful to model predictions and better understood by humans.

Large language models (LLMs) can be used to generate human-like responses, offering a promising avenue for creating immersive and interactive environments. These environments have the potential to emulate the dynamic storylines readers might encounter in books, similar to those portrayed in the television series “Westworld.” However, the capability of LLMs to truly grasp an author’s intent remains a challenge. Narrative understanding seeks to capture the cognitive processes of authors, shedding light on their knowledge, intentions, beliefs, and desires.

We will introduce our recent work, NarrativePlay, a system that allows users to role-play a fictional character and interact with other characters in narratives such as novels in an immersive environment, guided by personality traits extracted from narratives. We also incorporate automatically generated visual displays of narrative settings, character portraits, and character speech, greatly enhancing the overall user experience.