Scientific articles can be annotated with short sentences, called highlights, providing readers with an at-a-glance overview of the main findings. Highlights are usually manually specified by the authors. This paper presents a supervised approach, based on regression techniques, with the twofold aim at automatically extracting highlights of past articles with missing annotations and simplifying the process of manually annotating new articles. To this end, regression models are trained on a variety of features extracted from previously annotated articles. The proposed approach extends existing extractive approaches by predicting a similarity score, based on n-gram co-occurrences, between article sentences and highlights. The experimental results, achieved on a benchmark collection of articles ranging over heterogeneous topics, show that the proposed regression models perform better than existing methods, both supervised and not.

科学文章可以用短句子（称为重点）进行注释，从而使读者可以快速浏览主要发现。重点通常由作者手动指定。本文提出了一种基于回归技术的监督方法，其双重目的是自动提取缺少注释的过去文章的重点，并简化手动注释新文章的过程。为此，在从先前注释的文章中提取的各种特征上训练了回归模型。所提出的方法通过基于文章句子和重点之间的n-gram共现来预测相似性分数来扩展现有的提取方法。实验结果是根据涵盖不同主题的文章的基准集合获得的，