The primary object of this session would be to give an overview of QA over knowledge graphs. This will include common source KGs, the datasets used to evaluate QA systems, an overview of the prominent types of approaches, and familiarizing the audience with the sub-tasks (like entity linking, predicate linking) and the terminology used in the community. We will then introduce the basic concepts of NLP needed for semantic parsing based QA systems (for the next session.).
In this hands-on session, we will build our first QA system using QANARY - a methodoloy for choreographing QA pipelines distributed over the Web. We will also demonstrate how to build components for QANARY. In the process, we will introduce different concepts of Semantic Parsing relevant for QA.
In this session, we will have a more specific discussion of different approaches for QA, based on a selection of recent works in the field. This will also include some less traditional approaches like rule learning, and question paraphrasing, using textual evidence. Towards the end, we will introduce the basics of Deep Learning, and the prerequisites for the next hands-on session.
We will give a practical introduction for the development and training of a simple question answering system using neural networks, focusing on the implementation of a learning-to-rank approach for QA in PyTorch. First, we will show how to write a basic neural model for ranking questions, how to prepare training data and how to train and evaluate the model. Then we will discuss possible improvements to the model, such as handling out-of-vocabulary words and better predicate representation.