Authors : Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, Richard Socher

• Salesforce Research

Paper : https://arxiv.org/pdf/1909.05858.pdf Code : https://github.com/salesforce/ctrl

Summary

Control Code 를 이용하여 Language model 에 Condition을 주어 도메인, 스타일, 토픽 등과 같이 생성 되는 문장의 내용 및 스타일을 컨트롤 하여 생성 함

• Generation의 Control 가능성
• Conditional Transformer Language Model 사용
• Unsupervised learning (co-occurs)

Abstract

Large-scale language models show promising text generation capabilities, but users cannot easily control particular aspects of the generated text. We release CTRL, a 1.63 billion-parameter conditional transformer language model, trained to condition on control codes that govern style, content, and task-specific behavior. Control codes were derived from structure that naturally co-occurs with raw text, preserving the advantages of unsupervised learning while providing more explicit control over text generation. These codes also allow CTRL to predict which parts of the training data are most likely given a sequence. This provides a potential method for analyzing large amounts of data via model-based source attribution. We have released multiple full-sized, pretrained versions of CTRL at https://github.com/salesforce/ctrl.

1. Introduction

• Computer Vision 분야에서는 Generative adversarial networks가 image generation을 퍼포먼스를 상승 시켰으며, control 하는 연구도 진행 되었음
• Natural Language Processing 분야에서도 마찬가지로 conditional language model을 사용하여 text generation에 대한 연구 및 transfer learning 등 다양한 연구가 되었음.