A Roadmap to Realization Approaches in Natural Language Generation

Abstract

Text realization is the most significant step involved in natural language generation. It involves the approaches used to generate syntactically and semantically valid text, given an abstract linguistic representation. Based on the data and nature of data, a typical task generation includes text-to-text generation, database-to-text generation, concept-to-text generation, and speech-to-text generation. There are many approaches of natural language generation to generate texts, usually from non-linguistic structured data, which varies from a canned text approach to methods of learning from a text corpus and generating text based on the characters, content, keywords, size of the text, context, etc. Much work has also been done in learning and generating text mimicking a writing style. For applications like tutoring systems, wherein the text has to be manipulated and validated, we have to rely more on a template-based approach. Machine learning and other probabilistic-based statistical approaches can generate text for applications like report generation, summarization. This paper presents a roadmap and a comparative analysis of various text realization approaches.