基于语义感知的英文文献自动标引概念遴选方法*

孟旭阳 白海燕 梁 冰 王 莉

(中国科学技术信息研究所 北京 100038)

Automatic Indexing Concept Selection Method of English Documents Based on Semantic Perception

Meng Xuyang Bai Haiyan Liang Bing Wang Li

(Institute of Scientific and Technical of Information of China，Beijing 100038)

Abstract：[Purpose/Significance]In the era of resource digitalization, the literature service is changing to knowledge service. High-quality subject indexing is the foundation and key to improve the ability of literature knowledge services. Aiming at the low accuracy of automatic indexing of English scientific and technological literature, a concept selection optimization method based on semantic perception is proposed.[Method/Process]Based on the automatic subject indexing of knowledge organization system, word embedding in natural language processing is used to represent the semantic vector of concept and literature content, and then perform semantic perception and evaluation to achieve the selection of concept indexing results at the semantic level. This method adopts a technical method based on the combination of knowledge organization system and natural language processing, which makes up for the lack of semantic level, further reduces the impact of unrelated concepts, and improves the accuracy of concept indexing results.[Result/Conclusion]The experimental results show that the method in this paper has good semantic perception performance, effectively reduces irrelevant concepts, greatly improves the relevance of indexing results and literature, and provides valuable reference and support for the construction of scientific and technological literature resource knowledge service and related research.

Keywords：automatic indexing; concept selection; semantic perception; word embedding

0 引 言

近年来，随着信息资源的迅速增长以及科技文献服务需求的日益知识化，对科技文献主题自动标引的准确度提出了更高的要求。文献主题的自动标引是对文献数字资源进行组织的有效手段，标引的质量直接影响到数字资源的质量和利用[1]，也是知识化服务重点需要解决的问题之一，具有重要的研究意义和较高的实用价值。

基于知识组织系统的自动主题概念标引可以建立不同术语表达形式的一致性规范概念，为不同术语之间提供语义关联，能够提高概念的知识化关联，实现资源的组织化。然而，目前的文献主题概念标引结果并不理想，以基于英文超级科技词表(STKOS)[2]的英文文献主题概念自动标引系统[3]为例，经统计分析发现，在英文科技文献自动标引的概念遴选步骤中，一些与文献主题内容不相关的概念却被保留在最终标引结果中，忽略了概念结果与文献之间的语义相关性分析，是自动标引质量不佳的重要及主要原因。

为了进一步提高标引结果与文献内容的相关性，本文基于英文超级科技词表(STKOS)的英文文献主题标引，提出了基于语义感知的自动标引概念遴选优化方法。首先，通过运用自然语言处理技术，训练科技领域的词向量的语言模型，建立了词本身与文献内容之间的语义联系。其次，构建了基于多层注意力加权的文献语义表示模型对文献进行表示。……