PUB - Publikationen an der Universität Bielefeld

The task of answering natural language questions over structured data has received wide interest in recent years. Structured data in the form of knowledge bases has been available for public usage with coverage on multiple domains. DBpedia and Freebase are such knowledge bases that include encyclopedic data about multiple domains. However, querying such knowledge bases requires an understanding of a query language and the underlying ontology, which requires domain expertise. Querying structured data via question answering systems that understand natural language has gained popularity to bridge the gap between the data and the end user. In order to understand a natural language question, a question answering system needs to map the question into query representation that can be evaluated given a knowledge base. An important aspect that we focus in this thesis is the multilinguality. While most research focused on building monolingual solutions, mainly English, this thesis focuses on building multilingual question answering systems. The main challenge for processing language input is interpreting the meaning of questions in multiple languages. In this thesis, we present three different semantic parsing approaches that learn models to map questions into meaning representations, into a query in particular, in a supervised fashion. Each approach differs in the way the model is learned, the features of the model, the way of representing the meaning and how the meaning of questions is composed. The first approach learns a joint probabilistic model for syntax and semantics simultaneously from the labeled data. The second method learns a factorized probabilistic graphical model that builds on a dependency parse of the input question and predicts the meaning representation that is converted into a query. The last approach presents a number of different neural architectures that tackle the task of question answering in end-to-end fashion. We evaluate each approach using publicly available datasets and compare them with state-of-the-art QA systems.