Reproduce “From Statistical Methods to Deep Learning, Automatic Keyphrase Prediction: A Survey” on OpenResearch

About this paper

Binbin Xie, Jia Song, Liangying Shao, Suhang Wu, Xiangpeng Wei, Baosong Yang, Huan Lin, Jun Xie and 1 more

Researchers from Xiamen University and Alibaba Group published an extensive survey on Automatic Keyphrase Prediction (AKP), systematically categorizing models from statistical to deep learning and large language models. The work includes the first standardized empirical comparison of representative AKP models, offering insights into their performance across various datasets and highlighting key challenges like domain generalization.

Key insights

• Supervised deep learning models generally outperform unsupervised and statistical methods for keyphrase extraction, but performance can degrade significantly due to domain discrepancies in test data characteristics.
• The ONE2SET paradigm for keyphrase generation demonstrates greater training stability and superior performance in predicting the correct number of keyphrases compared to ONE2ONE and ONE2SEQ approaches.
• While large language models (LLMs) like ChatGPT show strong zero-shot and few-shot capabilities, they are generally surpassed by state-of-the-art supervised models fine-tuned on extensive domain-specific datasets.

Summary by alphaXiv.