Memory-augmented QA

Neural QA system focused on compositional questions using attention over key–value facts: attend, retrieve, answer.

This project focuses on developing a memory-augmented question answering (QA) system that leverages external knowledge bases to enhance its ability to answer questions accurately.

The system is designed to retrieve and utilize relevant information from a structured set of facts, improving its performance on various QA tasks. The key components and techniques used in this project include:

• Implemented a Key-Value Memory Network (KVMemNet) to store facts as (key, value) pairs and retrieve relevant values via attention.
• Added batched dot-product attention over keys with value aggregation; supported single-hop reasoning.
• Constructed a tractable dataset through templated questions, vocabulary reduction, and distractor sampling to control difficulty and size.
• Trained the model using CrossEntropy loss, monitored performance metrics, and implemented strategies to mitigate overfitting.
• Improved retrieval by natural-language key formatting, and balanced distractors to encourage fine-grained discrimination to ensure robust performance.