Where Is the Automation Discourse Heading? Exploring Evolution and Trends of RPA Research Through Topic Modeling

This repository provides a comprehensive resource for bibliometric and topic modeling analyses of Robotic Process Automation (RPA) research. It includes the raw and processed data, as well as the analytical scripts used to extract and validate the 100 latent research topics. 1. Bibliometric and Topic Modeling Data The dataset consists of primary Excel/CSV files alongside text files for preprocessing. data.xlsx: Includes multiple sheets containing Scopus bibliographic data, topic distributions, and derived metrics, enabling detailed exploration of thematic structures and trends. Data from Scopus: Contains bibliographic metadata exported from Scopus, designed for use in Latent Dirichlet Allocation (LDA) analyses. Key fields include authors, abstracts, publication years, and other relevant publication information. Each paper is annotated with manually or algorithmically assigned categories and topics. Figures 8A and 9A: Provides a detailed table of all identified topics, including associated terms, total paper counts, and citation counts per topic. Calculation of Combined Score: Topics are ranked according to a calculated combined score, which reflects their relative importance or influence within the dataset (aggregating research interest and impact). RPA_LDA_RESULTS_(k=100).xlsx: Contains the results of the LDA analysis with 100 topics, organized into two sheets: top_100_terms_in_each_topic: Lists the top 100 most representative terms for each topic, facilitating interpretation of thematic content. probabilities_with_each_topic: Provides the topic probabilities for all 3,275 documents across the 100 topics, enabling exploration of topic distributions at the document level. Human Validation & Stratification Data (New): topic_stratified_20%_origin.xlsx (and its CSV equivalent filtered_abstracts_lda_sample20.csv): Contains a 20% stratified sample of the original abstracts. This dataset ensures proportional representation across all LDA topics to create a balanced dataset for human and LLM validation. finalTableLDA+HumanLabels.csv: A ground-truth validation dataset containing the sampled abstracts mapped against their algorithmically assigned LDA topics and manually annotated human labels. Preprocessing Files: Elimination terms (extended stopwords) for the LDA analysis are provided in stopwords.txt, and general words are listed in general_words.txt. 2. Analytical Scripts and Validation Datasets The analytical code is separated into dedicated repositories and scripts, each containing detailed execution structures for reproducibility. LDA Analysis R Scripts: This repository (Supplementary files R.zip) contains the R scripts used for the end-to-end Latent Dirichlet Allocation (LDA) topic modeling pipeline. It includes the code for the systematic data preprocessing, statistical determination of the optimal number of topics, model training (Gibbs sampling), and posterior inference for new documents. Human Validation & Stratification Scripts (New): sample_stratified.py: A Python script used to group the baseline data by LDA topic and extract a reproducible 20% stratified sample (using a fixed random seed), generating the subset used for manual review. cohens_kappa_analysis.py: A Python script designed to process finalTableLDA+HumanLabels.csv and calculate the Cohen's kappa (κ) agreement coefficient. It evaluates the inter-rater reliability between the automated LDA model topic assignments and the human labels, outputting both the score and a cross-tabulation matrix. LLM Validation Scripts: This repository (CohensKappaLLMValidation.zip) contains the scripts and the dataset used to validate the robustness of the identified topics via Large Language Models (LLMs). It includes the zero-shot classification prompts provided to the LLMs (GPT-4o, Mistral Large, and Claude Haiku), the resulting LLM labeling dataset, and the scripts used to calculate the Cohen’s kappa (κ) agreement coefficients between the probabilistic LDA allocations and the LLM assignments.