Unsupervised Text Generation by Learning from Search
Dec 1, 2020·
,,,,,·
1 min read
Jingjing Li
Zichao Li
Lili Mou
Xin Jiang
Michael R. Lyu
Irwin King
Abstract
We propose a novel unsupervised text generation approach that learns to generate text by searching through a large text corpus. The key idea is to treat text generation as a search problem, where we search for the best sequence of tokens that maximizes a predefined objective function. We introduce a search-based text generation framework that consists of two main components: (1) a search module that explores the token space efficiently, and (2) a learning module that updates the search policy based on the search results. The search module uses a beam search algorithm to explore the token space, while the learning module updates the search policy using reinforcement learning. We demonstrate the effectiveness of our approach on various text generation tasks, including story generation, dialogue generation, and code generation. Our results show that the proposed method can generate high-quality text without requiring any labeled data.
Type
Publication
In Advances in Neural Information Processing Systems 34 (NeurIPS 2020)
This paper presents a novel approach to unsupervised text generation that leverages search-based learning. The key contributions include:
- A search-based text generation framework that treats text generation as a search problem
- An efficient beam search algorithm for exploring the token space
- A reinforcement learning-based approach for updating the search policy
- Comprehensive experiments on various text generation tasks showing the effectiveness of the proposed method
The paper demonstrates that high-quality text can be generated without requiring labeled data, making it particularly useful for scenarios where labeled data is scarce or unavailable.
