Multi-task fine-tuning for generating keyphrases in a scientific domain

Abstract

Automatic selection of keyphrases (keywords) is a major challenge to finding and systematizing scholarly documents. This paper investigates the efficiency of using titles of scientific papers as additional information for keyphrase generation. We propose an approach to multi-task fine-tuning the BART model using control codes <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> . It is shown that the suggested approach can improve the performance of BART for the task of keyphrase generation. In some cases, the presented model outperforms state-of-the-art models for keyphrase extraction. Moreover, the results have demonstrated that multitask fine-tuning also increases the performance of title generation.