This paper introduces Key-Value Memory Networks (KV-MemNN), a new neural network architecture that improves the performance of question answering (QA) systems by directly reading documents. The paper addresses the challenge of answering questions by directly reading documents, which is difficult due to the unstructured and ambiguous nature of text. Traditional QA systems often rely on Knowledge Bases (KBs), which are structured but limited in scope and may not contain all necessary information. To bridge this gap, the authors propose KV-MemNN, which uses a key-value memory structure to store and retrieve information, allowing the model to learn from both structured and unstructured data. The paper presents a new QA benchmark called WIKIMOVIES, which contains 100,000 questions in the movie domain. It includes raw text, a preprocessed KB, and an imperfect KB derived from information extraction. The benchmark allows for comparison between KB-based and document-based QA systems. The authors also compare their model with existing methods, including Memory Networks and supervised embeddings, and show that KV-MemNN outperforms these methods on the WIKIMOVIES dataset. KV-MemNN is designed to store facts in a key-value structured memory, which allows the model to learn to use keys to address relevant memories with respect to the question. The model then uses the corresponding values to predict an answer. The key-value memory structure enables the model to encode prior knowledge and leverage complex transforms between keys and values, while still being trained using standard backpropagation via stochastic gradient descent. The authors also evaluate the performance of KV-MemNN on the WIKIQA benchmark, which is another Wikipedia-based QA dataset. They show that KV-MemNN achieves state-of-the-art results on this benchmark, surpassing the most recent attention-based neural network models. The paper also discusses related work, including early QA systems based on information retrieval, and the development of new QA methods based on semantic parsing. It highlights the challenges of using KBs, such as their sparsity and fixed schemas, and the need for efficient information extraction methods to populate KBs automatically. The paper concludes that KV-MemNN is a versatile model for reading documents or KBs and answering questions about them. It allows for the encoding of prior knowledge in the key and value memories, and could be applied to other tasks as well. Future work should aim to further close the gap between direct document reading and KB-based QA systems.This paper introduces Key-Value Memory Networks (KV-MemNN), a new neural network architecture that improves the performance of question answering (QA) systems by directly reading documents. The paper addresses the challenge of answering questions by directly reading documents, which is difficult due to the unstructured and ambiguous nature of text. Traditional QA systems often rely on Knowledge Bases (KBs), which are structured but limited in scope and may not contain all necessary information. To bridge this gap, the authors propose KV-MemNN, which uses a key-value memory structure to store and retrieve information, allowing the model to learn from both structured and unstructured data. The paper presents a new QA benchmark called WIKIMOVIES, which contains 100,000 questions in the movie domain. It includes raw text, a preprocessed KB, and an imperfect KB derived from information extraction. The benchmark allows for comparison between KB-based and document-based QA systems. The authors also compare their model with existing methods, including Memory Networks and supervised embeddings, and show that KV-MemNN outperforms these methods on the WIKIMOVIES dataset. KV-MemNN is designed to store facts in a key-value structured memory, which allows the model to learn to use keys to address relevant memories with respect to the question. The model then uses the corresponding values to predict an answer. The key-value memory structure enables the model to encode prior knowledge and leverage complex transforms between keys and values, while still being trained using standard backpropagation via stochastic gradient descent. The authors also evaluate the performance of KV-MemNN on the WIKIQA benchmark, which is another Wikipedia-based QA dataset. They show that KV-MemNN achieves state-of-the-art results on this benchmark, surpassing the most recent attention-based neural network models. The paper also discusses related work, including early QA systems based on information retrieval, and the development of new QA methods based on semantic parsing. It highlights the challenges of using KBs, such as their sparsity and fixed schemas, and the need for efficient information extraction methods to populate KBs automatically. The paper concludes that KV-MemNN is a versatile model for reading documents or KBs and answering questions about them. It allows for the encoding of prior knowledge in the key and value memories, and could be applied to other tasks as well. Future work should aim to further close the gap between direct document reading and KB-based QA systems.