Knowledge Augmentation

Knowledge augmentation refers to methods for enhancing neural models—particularly language models—with access to external knowledge sources during both training and inference. Rather than encoding all world knowledge implicitly in model parameters, these approaches explicitly retrieve or condition on information from structured knowledge bases (e.g., knowledge graphs), textual corpora (e.g., Wikipedia), or other external sources. This strategy improves interpretability (retrieval decisions can be inspected), modularity (knowledge can be updated without retraining the full model), and performance on knowledge-intensive tasks.

Key papers

• [[2020-guu-realm]] — jointly pre-trains a retriever with a language model using masked language modeling signal

Related topics

• Information Retrieval (retrieval is a core mechanism)
• Language Models (the models being augmented)
• Question Answering Systems (a key application domain)