An Efficient Deep Learning Framework for Revealing the Evolution of Characterization Methods in Nanoscience

Abstract

Text mining has emerged as a powerful strategy for extracting domain knowledge structure from large amounts of text data. To date, most text mining methods are restricted to specific literature information, resulting in incomplete knowledge graphs. Here, we report a method that combines citation analysis with topic modeling to describe the hidden development patterns in the history of science. Leveraging this method, we construct a knowledge graph in the field of Raman spectroscopy. The traditional Latent Dirichlet Allocation model is chosen as the baseline model for comparison to validate the performance of our model. Our method improves the topic coherence with a minimum growth rate of 100% compared to the traditional text mining method. It outperforms the traditional text mining method on the diversity, and its growth rate ranges from 0 to 126%. The results show the effectiveness of rule-based tokenizer we designed in solving the word tokenizer problem caused by entity naming rules in the field of chemistry. It is versatile in revealing the distribution of topics, establishing the similarity and inheritance relationships, and identifying the important moments in the history of Raman spectroscopy. Our work provides a comprehensive tool for the science of science research and promises to offer new insights into the historical survey and development forecast of a research field.

1. Introduction

Data-driven methods have attracted much interest in literature survey and fundamental research. They help researchers forecast the hotspots in the near future, and administrators facilitate the formulation of funding policies [1]. As a distributed repository of scientific knowledge, scientific literature represents the fundamental data unit for studying the structure and evolution of science [2]. Traditionally, researchers summarized the patterns and trends of scientific development by reading a large amount of literature one by one. Unfortunately, this paradigm is confronted with unprecedented challenges in the field of nanoscience and nanotechnology. As a field that has attracted much interests from scientists, it usually contains millions of scientific literature, leaving a big challenge to extract research trends and potential research hotspots in nanoscience and nanotechnology manually. To address this issue, in recent years researchers started to utilize quantitative research methods to analyze the evolution of scientific structure and research hotspots, such as literature metrology [3] and science mapping analysis [4]. Nevertheless, most methods rely on particular data formats and literature indicators, resulting in failures to comprehend substantive content and academic ideas.

Text mining methods provide an opportunity for automatically reading literature and extracting the viewpoints therein and are beneficial to reducing time costs and avoiding human errors [5]. The topic models, as generally divided into structural [6], dynamic [7], and neural [8] topic models, had been proven efficient in deducing potential topic distributions and obtaining a birds-eye view of topic evolution [9–11]. The reference section was widely recognized as a significant component of a piece of published literature because it is a complex combination of considerations and it informs the substantial knowledge transfer of important arguments, experimental methods, and discoveries [12]. To date, most of the works were concentrated on the textual information of the literature, paying less attention to considering the inter-reference information. The lack of reference information prevents researchers from delving into the potential connections among literature and results in an incomplete knowledge graph. However, there is a limited method that can incorporate citation information into the topic information in the state-of-art text mining.

Recently, the Bidirectional Encoder Representations from Transformers Topic (BERTopic) model was proposed to generate coherent topic representations [13]. It was a scalable framework that allowed researchers to integrate external information and to construct a complete domain knowledge graph. Herein, we developed a novel method that integrated the BERTopic model and citation analysis to demonstrate the entire evolution of domain knowledge. To get a corpus, the web crawling technique was applied to gather literature from the Web of Science.