This PhD project investigates the critical challenge of catastrophic forgetting in neural information retrieval (NIR) systems. Contemporary NIR architectures exhibit significant performance degradation when integrating new information while attempting to preserve existing knowledge - a fundamental limitation in our expanding digital landscape.

Background:

Neural information retrieval has transformed traditional search systems through innovative deep learning implementations. Modern approaches include embedding-based architectures (DRMM, KNRM, DUET) and pre-training based frameworks (BERTdot, ColBERT), which have demonstrated remarkable success in static environments. However, these models face considerable challenges in continuous learning scenarios.

The phenomenon of catastrophic forgetting emerges as a central challenge when models incorporate new information, leading to deterioration of previously acquired knowledge. While continual learning strategies have shown promising results across various domains, their application within NIR systems demands deeper exploration, particularly concerning topic distribution shifts and data volume dynamics. Moreover, adaptive learning strategies are necessary to ensure models can adjust effectively to evolving data and retrieval requirements without compromising performance.

Existing NIR systems require complete retraining to integrate new information-an approach that is computationally demanding and impractical for real-world deployment. Recent research suggests promising directions in applying continual learning to NIR, yet fundamental challenges remain in developing specialised strategies, understanding topic shifts, and implementing efficient memory management solutions.

Research Objectives:

1. Develop a framework for continual and adaptive learning in NIR systems, addressing catastrophic forgetting while enabling the model to dynamically adjust to new data and retrieval tasks.

2. Design and optimise advanced continual learning strategies, focusing on memory management, handling topic diversity, and adapting to variations in data volume, ensuring models can learn continuously and flexibly.

3. Integrate domain adaptation techniques and refine evaluation metrics, ensuring the scalability and practical efficiency of the proposed strategies across dynamic information retrieval environments.

4. Case Study: Apply the developed methods in a real-world NIR system to demonstrate their practical effectiveness in dynamic and ever-evolving environments.