Original title: Drilling Down into the Discourse Structure with LLMs for Long Document Question Answering
Authors: Inderjeet Nair, Shwetha Somasundaram, Apoorv Saxena, Koustava Goswami
How Do Large Language Models Navigate Long Documents for Answers? The article delves into evidence retrieval for long document question answering. It explores leveraging the capabilities of large language models (LLMs) for zero-shot evidence retrieval, aiming to avoid high computational costs. Current models struggle with limited context lengths and may miss inter-segment relationships. To tackle this, the article proposes leveraging document discourse structures to condense representations. This technique enables a more comprehensive understanding of document relationships while significantly reducing the token count. The proposed method retains exceptional performance, achieving nearly the same results as the best zero-shot approach while processing only a fraction of the tokens. Moreover, when combined with a self-asking reasoning agent, it comes remarkably close to the zero-shot performance achieved using gold evidence, particularly in complex multi-hop question answering scenarios.
Original article: https://arxiv.org/abs/2311.13565