Optimizing Neural Information Retrieval Techniques

An information retrieval (IR) system is software which facilitates the organization, retrieval, and ranking or evaluation of relevant documents from a digital repository based on input from a user. People interact with IR systems constantly--to search for a song from a streaming service, find a restaurant in a new city, or find the receipt for an order in their email inbox.

In recent years, there has been an explosion of development in using deep neural networks and deep learning methods for IR tasks. Instead of relying on heavily pre-defined, handcrafted features for query-document matching and ranking, neural network models can learn relationships between words and phrases, patterns, and hierarchical structure. However, using deep learning comes with a high computational cost, and require large amounts of training data before they can be deployed.

Out project looks at two neural information retrieval methods—ColBERT and SPLADE—and attempts to improve upon them using various techniques. Our most successful model uses quantization techniques to make the ColBERT model faster.