Retrieval-augmented language models (RALMs) represent a substantialadvancement in the capabilities of large language models, notably in reducingfactual hallucination by leveraging external knowledge sources. However, thereliability of the retrieved information is not always guaranteed. Theretrieval of irrelevant data can lead to misguided responses, and potentiallycausing the model to overlook its inherent knowledge, even when it possessesadequate information to address the query. Moreover, standard RALMs oftenstruggle to assess whether they possess adequate knowledge, both intrinsic andretrieved, to provide an accurate answer. In situations where knowledge islacking, these systems should ideally respond with "unknown" when the answer isunattainable. In response to these challenges, we introduces Chain-of-Noting(CoN), a novel approach aimed at improving the robustness of RALMs in facingnoisy, irrelevant documents and in handling unknown scenarios. The core idea ofCoN is to generate sequential reading notes for retrieved documents, enabling athorough evaluation of their relevance to the given question and integratingthis information to formulate the final answer. We employed ChatGPT to createtraining data for CoN, which was subsequently trained on an LLaMa-2 7B model.Our experiments across four open-domain QA benchmarks show that RALMs equippedwith CoN significantly outperform standard RALMs. Notably, CoN achieves anaverage improvement of +7.9 in EM score given entirely noisy retrieveddocuments and +10.5 in rejection rates for real-time questions that falloutside the pre-training knowledge scope.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)