Seminar: Natural Language Generation

Natural language generation (NLG) deals with the computational synthesis of natural language texts. The goal is either to create new texts based on information from knowledge bases or to rewrite a given text into another text with specified properties. In the last years, NLG has made tremendous advances due to the development of powerful neural language models. Common tasks in NLG include automatic response generation, text style transfer, summarization, and debiasing. The aim of this seminar is to learn about state-of-the-art research in NLG as well as to discuss benefits, limitations, and ethical concerns of technologies that generated text.

Based on a few introductory talks, each participant will choose a sophisticated topic from recent related research. For this topic, knowledge from different literature has to be acquired and presented in a scientific talk. The talks are given in weekly sessions during the lecture time. In addition, the topic has to be summarized and discussed in detail in a paper-like article to be submitted in the middle of the lecture-free time.

• Basics of statistics
• Knowledge of programming, ideally Python
• Any course on machine learning or artificial intelligence
• Master's course: Statistical Natural Language Processing (preferred)
• Bachelor's course: Introduction to Natural Language Processing (alternatively)

• Daniel Jurafsky and James H. Martin. 2009. Speech and Language Processing: An Introduction to Natural Language Processing, Speech Recognition, and Computational Linguistics. Prentice-Hall, 2nd edition. Free draft of third edition: Speech and Language Processing

• Organizational Information (slides)
• Part I – Introduction to Natural Language Generation (slides)
• Part II – Overview of Seminar Topics (slides)
• Part III – Topic Assignment and Next Tasks (slides)
• Part IV – Basics of Scientific Presentation (slides)

The maximum number of participants is 20. The seminar grade depends on the talk (~50%), the article (~40%), and participation (~10%). All three aspects need to be passed individually.