Information seeking over semi-structured tabular data

Information seeking systems aim to satisfy a user's information needs by finding relevant textual information and responding to user queries. In this thesis, we study how the information seeking task can be extended to semi-structured tables. Semi-structured tables are fact-heavy and pose significant challenges to language models that aim to effectively meet a user's information needs. To understand these challenges, we first study and compare language models' performance over either text or tabular context. Further, as real-world questions are often over multiple tabular contexts, we develop datasets and models to combine information across multiple input tables. Next, we study how to adapt multi-tabular information-seeking tasks over conversational agents in a single-turn setup. As the expected output modality of conversational agents is text, the multi-tabular information-seeking task enforces the agent to generate personalized text-based responses adapted to the user's query. To do so, we introduce a query-focused table summarization task and develop datasets and models. We further introduce question answering over tables in a low-resourced setting to study the challenges introduced to mathematical table reasoning in non-Latin scripts in a resource-scarce setup. Finally, as retrieval of relevant information is essential to information-seeking systems, we study sparse retrieval models for adaptation to new domains and focus on the trade-off between efficiency and effectiveness of text-based sparse retrieval models. In conclusion, this thesis aims to advance the development of information-seeking systems for semi-structured tables by studying various aspects of query-focused table processing and by designing large-scale methodologies and resources, such as datasets and models.