A Tutorial on Evaluation Metrics used in Natural Language Generation

There has been a massive surge of Natural Language Generation (NLG) models in the recent years, accelerated by deep learning and the availability of large-scale datasets. With such rapid progress, it is vital to assess the extent of scientific progress made and identify the areas/components that need improvement. To accomplish this in an automatic and reliable manner, the NLP community has actively pursued the development of automatic evaluation metrics. Especially in the last few years, there has been an increasing focus on evaluation metrics, with several criticisms of existing metrics and proposals for several new metrics. This tutorial presents the evolution of automatic evaluation metrics to their current state along with the emerging trends in this field by specifically addressing the following questions: (i) What makes NLG evaluation challenging? (ii) Why do we need automatic evaluation metrics? (iii) What are the existing automatic evaluation metrics and how can they be organised in a coherent taxonomy? (iv) What are the criticisms and shortcomings of existing metrics? (v) What are the possible future directions of research?