On Leveraging Encoder-only Pre-trained Language Models for Effective Keyphrase Generation

TL;DR

This paper explores the use of encoder-only pre-trained language models for keyphrase generation, demonstrating their effectiveness and proposing optimal architectural strategies, with findings showing their competitive performance and data efficiency.

Contribution

It investigates encoder-only PLMs for KPG, compares architectures, and highlights prefix-LM fine-tuning as a superior, data-efficient approach over traditional seq2seq models.

Findings

Encoder-only PLMs achieve broader keyphrase predictions.

Prefix-LM fine-tuning outperforms seq2seq PLMs.

Deeper models are more effective than wider ones.

Abstract

This study addresses the application of encoder-only Pre-trained Language Models (PLMs) in keyphrase generation (KPG) amidst the broader availability of domain-tailored encoder-only models compared to encoder-decoder models. We investigate three core inquiries: (1) the efficacy of encoder-only PLMs in KPG, (2) optimal architectural decisions for employing encoder-only PLMs in KPG, and (3) a performance comparison between in-domain encoder-only and encoder-decoder PLMs across varied resource settings. Our findings, derived from extensive experimentation in two domains reveal that with encoder-only PLMs, although KPE with Conditional Random Fields slightly excels in identifying present keyphrases, the KPG formulation renders a broader spectrum of keyphrase predictions. Additionally, prefix-LM fine-tuning of encoder-only PLMs emerges as a strong and data-efficient strategy for KPG, outperforming general-domain seq2seq PLMs. We also identify a favorable parameter allocation towards model depth rather than width when employing encoder-decoder architectures initialized with encoder-only PLMs. The study sheds light on the potential of utilizing encoder-only PLMs for advancing KPG systems and provides a groundwork for future KPG methods. Our code and pre-trained checkpoints are released at https://github.com/uclanlp/DeepKPG.