# Automatic and Human-AI Interactive Text Generation

**Authors:** Yao Dou, Philippe Laban, Claire Gardent, Wei Xu  
**arXiv:** [2310.03878](https://arxiv.org/abs/2310.03878)  
**Date:** 2023-10-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.03878v1/

## Key points

- Text-to-text generation involves improving a piece of text while maintaining its original meaning and length based on specific criteria.
- Applications such as text simplification, paraphrase generation, and style transfer fall under this category.
- These tasks are more constrained in terms of semantic consistency and targeted language styles compared to open-ended text completion tasks.
- The tutorial focuses on two main areas: text simplification and revision.
- Significant advances discussed include non-retrogressive approaches, prompting with large language models instead of fine-tuning, new learnable metrics for evaluation, studies on non-English languages, and interdisciplinary research combining HCI+NLP+Accessibility.
- Insights from the InstructGPT paper reveal that "Rewrite" (text revision) accounts for 6.6% of use cases in OpenAI's API prompts.
- Various topics covered include Tasks and Datasets (e.g., Text Simplification), Neural and Language Models (e.g., Edit-based models), Automatic and Human Evaluation methods, Human-AI Collaborative Writing tools pre/post LLMs era with commercial tools showcased in live demos.
- Ethical considerations surrounding text generation are addressed along with conclusions and future directions in the field.

## Summary

This expanded tutorial delves into the realm of text-to-text generation. This subset of natural language generation tasks involves improving a piece of text while maintaining its original meaning and length based on specific criteria. Applications such as text simplification, paraphrase generation, and style transfer fall under this category. Unlike open-ended text completion tasks, these tasks are more constrained in terms of semantic consistency and targeted language styles. This level of control makes them ideal for studying models' ability to generate semantically adequate and stylistically appropriate text. The tutorial focuses on two main areas: text simplification and revision. It provides an overview of state-of-the-art research in natural language generation across four key aspects: Data, Models, Human-AI Collaboration, and Evaluation. Significant advances discussed include non-retrogressive approaches, prompting with large language models instead of fine-tuning, new learnable metrics for evaluation, studies on non-English languages, and interdisciplinary research combining HCI+NLP+Accessibility to create writing assistant systems. Insights from the InstructGPT paper reveal that "Rewrite" (text revision) accounts for 6.6% of use cases in OpenAI's API prompts. The tutorial outlines various topics including Tasks and Datasets (e.g., Text Simplification), Neural and Language Models (e.g., Edit-based models), Automatic and Human Evaluation methods (including reading comprehension questions for text simplification), Human-AI Collaborative Writing tools both pre-LLMs era and post-LLMs era with commercial tools showcased in live demos. Ethical considerations surrounding text generation are also addressed along with conclusions and future directions in the field. The tutorial aims to cater to a diverse audience ranging from researchers to practitioners in academia and industry with basic knowledge of natural language processing.