Abstract

Conversational assistants become increasingly popular, enabling users to interact with search systems through dialogue.
Although conversations offer a compelling and natural way of interacting with search engines, they pose significant challenges to system performance. This is mainly due to the verbosity and complexity of the multi-turn interactions with the search engine. Additionally, there is a lack of large scale high quality training data.
To this end, the thesis first explores how conversational verbosity can be tackled in a zero-shot setting. It introduces a retriever that resolves the conversational dependencies of the last user utterance by contextualizing it with relevant context from the conversation.
Next, it investigates how a system can resolve query ambiguity, by enabling mixed initiative interactions and asking clarifying questions. It does so by using a Retrieval Augmented Generation model that identifies potential search intents from user queries and the corpus.
Then, we explore how introducing clarifying questions in conversational search affects passage ranking systems. We conclude that conversations become more complex, as users add constraints or negations to queries and systems often fail to integrate those effectively.
In the light of that, we take a step back from conversational search to study passage ranking with ad-hoc complex compositional queries.
We investigate how to construct effective compositional embeddings for first stage ranking. We first propose a zero-shot framework to construct compositional embeddings. Next, we improve ranking performance on negations by allowing retrieval models to assign negative scores to unwanted aspects.