inverse prompting

Controllable Generation from Pre-trained Language Model via Inverse Prompt

Abstract

our results show that our proposed method substantially outperforms the baselines and that our gengeration quality is close to human performance on some of the tasks.

• Demo ``` poem gengeration https://pretrain.aminer.cn/apps/poetry.html

QA https://models.aminer.cn/os/qa ```

Introduction

Inverse prompting can be decoupled into three steps

• first:given a piece of generated text , an inverse prompt is constructed using the generated text.
• sencond: the conditional likelihood of the original prompt given the inverse prompt is computed based on the pre-trained language model
• Third ,the conditional likelihood is used as a score in beam search for selecting the best generation candidates.

Related Work

inverse prompting does not require any gradient update to the original model and is free of any additional attribute models.

• Open-Domain Long-Form Question-Answering
  • human evaluation
• Traditional Chinese Poem Generation
  • jiuge(no contemporary notion)

Methodology

Baseline: Prompting and Beam Search

Inverse Prompting

Implementation

Base Language Model

we train our base Chinese language model using Megatron-LM with Transformer-XL.

Open-Domain Long-Form Question-Answering

Open-Domain Poem Generation

Self Training for Poem Generation

Experiments