Improving Re-Ranking using Neural Models

Information retrieval (IR) has become increasingly important in our digital lives, with search and ranking playing a central role. To effectively address user information needs, neural language models are used to understand context and rank queries beyond simple term matching. However, neural rankers come with certain limitations. This thesis focuses on two key challenges: First, while these models excel at capturing context, they require large amounts of training data to fine-tune for specific ranking tasks, making it difficult to generalize well on smaller datasets without overfitting. Second, although neural rankers perform well on common (head) queries, their effectiveness on harder or tail queries remains limited, highlighting a need for more robust solutions. To address the challenge of training rankers on small datasets, we propose data augmentation techniques that aim to improve both effectiveness and robustness. Using supervised and unsupervised methods that generate training samples from parts of relevant documents, we improve sample efficiency. We further adapt a family of contrastive loss functions tailored for ranking, enabling the model to learn better from augmented data. Our results demonstrate that combining data augmentation with ranking-aware contrastive losses leads to models that generalize well, both on in-domain datasets (e.g., TREC-DL) and out-of-domain benchmarks. Hard queries are challenging due to their vague or uncommon nature. To improve ranking, we use LLM-based query enrichment and fine-tune a specialized ranker, combining its scores with a base ranker and a query performance estimate. Experiments on DL-Hard show significant gains. We also introduce an intent-based dataset, created using LLMs and crowd-sourcing, to support the evaluation and improvement of hard query handling. By integrating these methodologies, this thesis aims to develop a robust and versatile ranking model that excels across various query types and datasets. The proposed framework seeks to advance the state-of-the-art in information retrieval, providing practical solutions for more accurate and contextually relevant search results.