Skip to main content


Evaluating machine-generated text

Research Achievements

Evaluating machine-generated text

IGERT Trainee Emily Pitler and her advisor Ani Nenkova (Computer Science) have been developing automatic methods for evaluating the quality of machine-generated text. Humans effortlessly produce novel utterances which are grammatical, factual, and relevant. In contrast, machine-generated texts often are ungrammatical, incoherent, or redundant. The ability to generate more natural text would greatly improve several applications of natural language processing, such as language translation, summarization, and question answering. An important first step in producing more fluent text is the ability to measure how fluent a machine generated text is. Using human evaluations of human-generated and machine-generated text, they have been able to develop a system that can evaluate both the syntactic and referential coherence of new, previously unseen text. This method will allow for significant breakthroughs in developing natural language computer systems.