Advancing neural machine continual learning and unlearning for language models in information retrieval systems

Information retrieval (IR) refers to obtaining relevant information from a repository, such as a database or the internet, based on a user's query. The rise of deep learning has led to the development of neural network-based information retrieval (NIR) technologies, which have become one of the most effective approaches in this domain. IR systems in the real world are continuously evolving, thereby necessitating updates to incorporate new information and modify or remove outdated data. However, most NIR models operate in a static mode--trained on fixed datasets and deployed without further updates--limiting their effectiveness in dynamic information contexts. This thesis aims to address this research gap.

The first key contribution of this thesis is enabling NIR models to continuously learn and incorporate new data without forgetting previously accumulated knowledge. To achieve this, this thesis defines the task of continual neural information retrieval and proposes a multiple-topic dataset to evaluate this task. A continual learning framework for neural information retrieval (CLNIR) is then proposed. By allowing for the pairing of different NIR models and continual learning strategies, CLNIR offers a flexible and customisable approach to continual neural information retrieval, which effectively addresses new data challenges in IR.

The second key contribution is addressing data removal requests in IR systems without compromising the integrity of the remaining data or the predictive capabilities of NIR models. Data removal in machine learning models (machine unlearning) is widely studied but mostly designed for classification tasks and is ineffective in retrieval contexts. This thesis introduces a novel task termed Neural Machine unRanking (NuMuR), tailored to IR contexts, and includes two sub-tasks: query removal and document removal. A novel method Contrastive and Consistent Losses (CoCoL) is then proposed to address this task. CoCoL achieved superior performance in both query and document removal.

The third key contribution is providing data correction capabilities to NIR models, effectively decreasing the influence of outdated data and enhancing the relevance of new data. This thesis extends NuMuR to introduce the task of Corrective Neural Machine Unranking. To perform this task, a novel approach termed Corrective unRanking Distillation (CuRD) is proposed. By dynamically adjusting the relevance scores produced by NIR models, CuRD allows new data to occupy the original position of old data and minimises the relevance of old data, thus achieving correction.