The paper introduces Key-Value Memory Networks (KV-MemNNs) to address the challenge of directly reading documents for question answering (QA). Traditional QA systems often rely on Knowledge Bases (KBs), which can be restrictive and sparse. KV-MemNNs utilize a key-value structured memory to improve the performance of reading documents. The model stores facts in a key-value memory and reasons over them to predict answers. The authors construct the WIKIMOVIES dataset, which contains raw text and a preprocessed KB in the domain of movies, to evaluate the performance of different knowledge sources. Experiments show that KV-MemNNs outperform the original Memory Network and achieve state-of-the-art results on the WIKIQA benchmark, demonstrating the effectiveness of their approach in bridging the gap between KB-based and document-based QA.The paper introduces Key-Value Memory Networks (KV-MemNNs) to address the challenge of directly reading documents for question answering (QA). Traditional QA systems often rely on Knowledge Bases (KBs), which can be restrictive and sparse. KV-MemNNs utilize a key-value structured memory to improve the performance of reading documents. The model stores facts in a key-value memory and reasons over them to predict answers. The authors construct the WIKIMOVIES dataset, which contains raw text and a preprocessed KB in the domain of movies, to evaluate the performance of different knowledge sources. Experiments show that KV-MemNNs outperform the original Memory Network and achieve state-of-the-art results on the WIKIQA benchmark, demonstrating the effectiveness of their approach in bridging the gap between KB-based and document-based QA.