Fluency and completeness in instance-based natural language generation
Sebastian Varges · 2002
A fundamental assumption underlying candidate ranking in corpus-based approaches to Natural Language Generation is the idea that in order to be fluent the output should be as similar to a (human-authored) corpus as possible. However, the goal of maximizing fluency can conflict with other goals, like conveying the maximal amount of input and being faithful. We employ an instance-based sentence generation system to investigate how the right balance between the different goals can be struck and show empirical results supporting our proposals.