Summary

Modern search engines need to be able to deal with user queries that are common, unpopular or have never been seen before, as well as queries that are error-free or contain errors. Furthermore, they must adapt to new domains and operate effectively in resource-constrained environments to ensure accessibility and scalability. To fulfill the user's expectations and address the increasing complexity of their queries, modern search engines need to (i) provide access to unstructured knowledge sources, as well as structured knowledge sources, (ii) further provide question-answering functionalities to enhance the search engine result page with direct answers to the given queries, and finally (iii) support multiple means via which the users can express their queries e.g., text, voice, and image. In this thesis, we focus on developing robust neural retrievers to support various functionalities of modern search engines. In particular, we investigate how to improve the robustness and effectiveness of neural retrievers in noisy and low-resource settings. We first explore the impact that errors in a query have on the retrieval performance of neural retrievers for ad-hoc retrieval and propose ways to robustify them. Then, we explore the use of neural methods for multi-modal retrieval over spoken queries and textual documents. Next, we examine the challenges of training an effective neural retriever with limited computational resources to tackle complex user queries requiring multi-hop retrieval. Lastly, we investigate how neural retrieval can be used to increase relation prediction performance in knowledge graph question-answering over previously unseen domains.