Diagnosis Then Aggregation: An Adaptive Ensemble Strategy for Keyphrase Extraction

Xin Jin, Qi Liu, Linan Yue, Ye Liu, Lili Zhao, Weibo Gao, Zheng Gong, Kai Zhang, Haoyang Bi

Abstract: Keyphrase extraction (KE) is a fundamental task in the information extraction, which has recently gained increasing attention. However, when facing text with complex structure or high noise, current individual keyphrase extraction methods fail to handle capturing multiple features and limit the performance of the keyphrase extraction. To solve that, ensemble learning methods are employed to achieve better performance. Unfortunately, traditional ensemble strategies rely only on the extraction performance (e.g., Accuracy) of each algorithm on the whole dataset for keyphrase extraction, and the aggregated weights are commonly fixed, lacking fine-grained considerations and adaptiveness to the data. To this end, in this paper, we propose an Adaptive Ensemble strategy for Keyphrase Extraction (AEKE) that can aggregate individual KE models adaptively. Specifically, we first obtain the multi-dimensional abilities of individual KE models by employing cognitive diagnosis methods. Then, based on the diagnostic abilities, we introduce an adaptive ensemble strategy to yield an accurate and reliable weight distribution for model aggregation when facing new data, and further apply it to improve keyphrase extraction in the model aggregation. Extensive experimental results on real-world datasets clearly validate the effectiveness of AEKE. Code is released at https://github.com/kingiv4/AEKE.