English

Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning

Computation and Language 2021-07-06 v2 Artificial Intelligence Machine Learning

Abstract

Query reformulation aims to alter noisy or ambiguous text sequences into coherent ones closer to natural language questions. This is to prevent errors from propagating in a client-facing pipeline and promote better communication with users. Besides, it is crucial to maintain performance in downstream environments like question answering when rephrased queries are given as input. We show that under the previous framework (AQA), attempts to alter RL algorithms do not bring significant benefits to either reward acquisition or sequence fluency. Instead, we leverage a query-reformulating text-to-text transformer (QRT5) and apply policy-based RL algorithms to further nudge this reformulator and obtain better answers downstream by generating reward-acquiring query trajectories. QRT5 shows better sample efficiency in RL to achieve the same level of QA performance as the previous approach. It can generate reformulations with more readability based on query well-formedness evaluations and can generalize to out-of-sample data. Our framework is demonstrated to be flexible, allowing reward signals to be sourced from different downstream environments such as intent classification.

Keywords

Cite

@article{arxiv.2012.10033,
  title  = {Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning},
  author = {Jerry Zikun Chen and Shi Yu and Haoran Wang},
  journal= {arXiv preprint arXiv:2012.10033},
  year   = {2021}
}

Comments

Workshop on the 9th Dialog System Technology Challenge (DSTC-9), AAAI 2021

R2 v1 2026-06-23T21:04:03.944Z