The Future of Natural Language Generation in the Age of Large Language Models

Abstract

Natural language generation (NLG) research has focused in the past on mainly two aspects: (i) the process of language generation in itself (e.g., to give theoretical insights into models of human language) and (ii) the use of such approaches to solve applied problems (e.g., generation of medical reports, weather forecasts, etc.). Since the advent of large language models (LLMs), given their "generation capabilities", there has been a tendency among researchers to use LLMs for an increasing number of novel tasks. The use of LLMs raises questions for NLG researchers about which tasks, both theoretical and practical, are still considered NLG tasks per se, and which are still relevant, given that LLMs have solved many former problems in language generation. Through a structured literature search, we investigate domain shift in NLG before and after LLMs and discuss the positive and negative aspects that arise from this, with the aim of providing researchers with greater clarity on where we stand as a field.