Tags, when properly assigned to limited access papers, help users to estimate their relevance. This paper introduces a new approach for the selection of relevant tags as well as a recommendation for scientific papers tagging. The approach defines the relatedness between the tags attributed by users and the concepts extracted from the available sections of scientific papers based on statistical, structural and semantic aspects. Two different term-based graphs ([Formula: see text]-graph and [Formula: see text]-graph) were generated whose vertices indicate the terms and the edges represent the relatedness score between these terms. In addition, two algorithms were implemented to select and recommend the relevant tags: the neighbor-algorithm and the best-path-algorithm. The results of the experiments performed on a CiteULike collection of tagged papers show significant improvements only for the tagging of abstracted scientific articles. The approach was evaluated by referring to the full text of the papers with expert evaluation and comparing the tags generated by CiteULike users. Using the neighbor-algorithm, 80% of the top 10 recommended tags based on [Formula: see text]-graph and 76% of the top 10 recommended tags based on the [Formula: see text]-graph were relevant. While only 62% of those recommended by CiteULike users were relevant. The best-path-algorithm gave the best results in the top 20 and top 30 recommended tags and this in comparison with the tags recommended by the neighbor-algorithm and the tags assigned by CiteULike users.

中文翻译：

---

用于刚刚抽象的科学文章标记的基于图形的标记建议

标签，当正确分配给有限访问文件时，可以帮助用户估计它们的相关性。本文介绍了一种选择相关标签的新方法以及对科学论文标签的建议。该方法基于统计、结构和语义方面定义了用户归因的标签与从科学论文的可用部分中提取的概念之间的相关性。生成了两个不同的基于术语的图（[公式：参见文本]-图和 [公式：参见文本]-图），其顶点表示术语，边缘表示这些术语之间的相关性分数。此外，实施了两种算法来选择和推荐相关标签：邻居算法和最佳路径算法。在 CiteULike 标记论文集合上进行的实验结果表明，仅对抽象科学文章的标记有显着改进。该方法是通过参考论文全文和专家评估并比较CiteULike用户生成的标签来评估的。使用邻域算法，基于 [公式：参见文本]-graph 的前 10 个推荐标签中有 80% 和基于 [公式：参见文本]-graph 的前 10 个推荐标签中有 76% 是相关的。而 CiteULike 用户推荐的只有 62% 是相关的。best-path-algorithm 在前 20 个和前 30 个推荐标签中给出了最好的结果，这与邻居算法推荐的标签和 CiteULike 用户分配的标签相比。