基于词嵌入的科研主题排序研究

Automatic Scientific Topic Ranking based on Pre⁃trained Neural Embedding

为准确把握科研领域内文献主题的发展变化,常利用隐式语义特征提取科研主题分布。但由于主题挖掘技术本身的限制,并非所有主题都具有同等重要性或意义。有些主题可能包含太多背景词,信息空泛,或者主题词之间缺乏连贯性,导致主题缺乏实际意义。针对上述问题,在已有研究基础上,基于词嵌入,提出一种新的多维度评估主题质量算法;针对科研文档的特点,利用语料库的统计特征对无意义主题距离评估方法进行优化,并最终将二者融合到一个统一的主题排序框架中。实验结果表明,本文提出的方法可以有效提高主题排序整体效果,能够识别出非重要和质量差的主题,主题排序的整体效果优于现有方法。

To accurately explore the development and changes of topics in the field of scientific research,implicit semantic features are often used to extract the scientific topics.However,due to the limitation of the topic mining technology itself,not all the topics are of equal significant or meaningful.Some topics may contain background terms or lack coherence between topic terms,resulting in the lack of practical significance.According to the existing research,this paper proposes a new multi⁃dimensional topic quality evaluation algorithm based on word embedding,and uses the statistical features of the corpus to optimize the insignificant topic distance scoring method based on the characteristics of scientific documents,and finally integrates the two into a unified topic ranking framework.Experimental results show that our method can effectively improve the overall effectiveness of topic ranking,and can identify and distinguish the insignificant and poor⁃quality topics from the legitimate ones.The overall effect of topic ranking is better than existing methods.