关键词: 研究方法, 实体抽取, 学科驱动特征识别, 特征词密集簇

Abstract:

[Purpose/Significance] The ability to solve complex scientific problems is increasingly dependent on knowledge of methods from multiple disciplines. Research methods, theories, software tools, data resources, and analytical models are essential elements of knowledge production, but they are often scattered across academic texts and expressed in diverse forms. Accurately identifying such knowledge can reduce the burden of literature analysis, clarify how methods are selected and combined, and provide structured evidence for knowledge discovery. Existing studies have mainly focused on method-sentence detection, single-type entity recognition, or frequency analysis, with limited integration between fine-grained extraction and the identification of disciplinary research characteristics. This study therefore develops a unified framework incorporating key technologies, system support, and core applications. We extend method entity research from textual recognition to the analysis of disciplinary knowledge production by connecting method entities with documents, authors, and journals. This supports the organization of academic texts, the recommendation of methods, and the identification of methodological preferences in economics and management. [Method/Process] Guided by clustering, this study proposed a fine-grained entity extraction strategy for Chinese academic literature by integrating a domain dictionary, linguistic rules, statistical representation, similarity calculation, and manual verification. The procedure consists of four stages. First, academic texts were pre-processed, stop words were removed, and method-related feature terms were combined with an initial dictionary to construct a method feature-term bag. Synonymous and semantically equivalent expressions were then merged. Second, a dense-cluster procedure locates text segments in which method-related terms are concentrated, reducing the amount of text requiring examination. Third, regular-expression rules, vector-space representation, and cosine similarity were used to identify and recommend candidate entities. Candidates satisfying the specified threshold were submitted for manual verification. Fourth, confirmed entities were classified into four categories: method and theory, software tool, data, and model. Each entity was assigned a unique code and linked to its source document and metadata. Two annotators verified the candidates according to a unified specification, while disagreements were resolved by a third annotator. The empirical dataset comprises 9 858 articles published from 2011 to 2020 in six journals in economics and management. Frequency statistics, co-occurrence analysis, and social network analysis were applied to identify disciplinary characteristics associated with methodological knowledge. [Results/Conclusions] A total of 8 153 method entities were identified from the titles, abstracts, and keywords of the sampled articles. The proposed strategy assists annotators in locating, screening, normalizing, and classifying candidate entities, improving the efficiency and consistency of large-scale annotation. Linking these entities with bibliographic and author metadata further reveals quantitative distributions, co-occurrence relationships, and scientific collaboration networks. Economics and management research relies more heavily on method-and-theory knowledge than on software tools, data resources, or models. Empirical research, questionnaire surveys, and the difference-in-differences method are widely used, reflecting the field's emphasis on empirical evidence, data analysis, policy evaluation, and causal identification. The results also show that research methods and data resources are frequently used in complementary combinations. The method-driven author collaboration network exhibits a complex structure and a tendency toward continuous development, with some stable groups maintained through shared methods and data-collection practices. Overall, fine-grained method entity extraction can support academic knowledge organization, method recommendation, disciplinary preference identification, and research collaboration analysis. Nevertheless, the approach still depends on expert-defined rules, domain dictionaries, manually selected thresholds, and Chinese lexical information, while its empirical scope is limited to six journals and one publication period. Future research should incorporate semi-supervised learning and contextual semantic representations, extend extraction to full texts, broaden the coverage of languages and disciplines, and evaluate cross-domain transferability. These improvements will enhance the generalizability of methods of knowledge extraction and support deeper analysis of disciplinary evolution and interdisciplinary knowledge production.

Key words: research methods, entity extraction, research feature recognition, intensive clusters of methodological feature words