The taught component of the course will consist of four parts:
I. General Introduction. In the first part of the course you will learn what the different aims of practical and theoretical NLG can be, what are the main elements of the standard NLG pipeline, how NLG systems are built, and how they are evaluated. Template-based and end-to-end systems will be discussed briefly.
II. Practical systems. You will get acquainted with a range of practical applications of NLG; a few will be discussed in detail: candidates applications are medical decision support, knowledge editing, and robo-journalism. Strengths, weaknesses, and opportunities for the practical deployment of these systems will be discussed. If time allows, we will devote attention to multimodal systems, which produce documents in which pictures or diagrams complement a generated text.
III. Module in focus: Referring Expressions Generation. We will zoom in on one part of the standard NLG pipeline, which is responsible for the generation of referring expressions (e.g., as when an NLG system says “the city where you work”, or “the area north of the river Rhine”). We will discuss a range of rule-based algorithms, and some that are based on Machine Learning.
IV. Perspectives on NLG. We will discuss what linguists, philosophers, and other theoreticians have to say about human language production, and how this relates to NLG. We may start with a Gricean approach, and continue with the Bayesian-inspired Rational Speech Acts approach. We will ask how accurate and how explanatory existing NLG algorithms are as models of human language production (i.e., human speaking and writing), and what are the main open questions for research in this area.
The core of the course will be presented in lectures. Additionally, students will be asked to read, present, and discuss some key papers and systems which illustrate the issues listed above.