Introduction

LDAShiny is a web application built with Shiny (Chang et al., 2017) and the golem framework (Fay et al., 2022). It lets you run an exploratory review of the scientific literature with the Latent Dirichlet Allocation model (LDA; Blei, Ng and Jordan, 2003) without writing code: from loading bibliographic data to interpreting topics and tracking how they evolve over time.

This guide walks through every tab of the application in the recommended order:

Installation and launch

The package requires R >= 4.1.0. Once installed, the application starts with a single function:

library(LDAShiny)
run_LDAShiny()

Bibliographic files (for example, Scopus exports) can be large. Shiny limits file uploads to 5 MB by default; if your files are larger, raise the limit before launching the application:

options(shiny.maxRequestSize = 1024 * 1024^2)  # 1 GB
run_LDAShiny()

Step 0. About

The About tab summarizes what the application does, its technical specifications (version, framework and main packages), the recommended way to cite LDAShiny, and information about the research group and the author.

Step 1. Importing data (Import Data)

First choose an action in the Select an action menu:

The fields used by the rest of the application are title, abstract, year, Journal, doi and database; the PRISMA cleaning also uses affiliations. For Excel files, column names are matched case-insensitively and the usual aliases are accepted (for example TI, AB, PY, SO, DI, DB, C1). The Excel file should include a database column, because the training module uses it to compute topic intensity by source.

Step 1b. Geographic and journal analysis (Geographic & Journal Analysis)

This module is available as soon as data have been imported. It provides three complementary views of the corpus and does not affect the modeling workflow.

Step 2. Preprocessing the text (Preprocess)

This module turns text into a document-term matrix. Everything is configured in the Settings tab and run with Run Preprocessing.

The text is always converted to lowercase and English stopwords are always removed, together with those from the uploaded file and any added interactively. For this reason the application is designed for corpora in English.

The other tabs of the module show and let you refine the result:

• DTM Summary: dimensions of the matrix before and after the sparsity filter. If the DTM ends up empty, or the filter removes every term, the application reports it; in the second case, try a less strict sparsity level.
• Term Frequency (tf_mat): table of terms with their total frequency and document frequency.
• Refine Stopwords: table sorted by descending document frequency. Terms that appear in many documents are often candidates for stopwords. Select the rows, click Add Selected & Reprocess, and the DTM is rebuilt automatically. The cumulative list of words added during the session can be downloaded as Excel or cleared with Clear Custom Stopwords.
• Downloads: the full frequency table (tf_mat.xlsx) and the final DTM (dtm_final_sparse.rds).

Step 3. Choosing the number of topics (LDA Inference)

The module fits an LDA model for each value of K in the chosen range, in parallel, and computes the coherence of every topic (Mimno et al., 2011). The summary measure for each model is the mean coherence of its topics.

Click Run LDA Inference. When it finishes, the status reports the K with the highest mean coherence and the elapsed time. Results appear in two tabs:

• Coherence Table: coherence value for each K, downloadable as Excel (coherence_mat.xlsx).
• Coherence Plot: coherence curve against K, with customization options (labels, theme, line color and width, points) and export as PNG, TIFF, JPEG or PDF.

The K with maximum coherence is a starting point, not a final decision. Coherence curves often have several regions with similar values, which is why the next step evaluates how robust the choice is.

Step 4. Stability and sensitivity (Stability & Sensitivity)

This module checks whether the chosen K is robust, using three ideas:

1. Stability against initialization. For each candidate K several models are trained, and the variability of their coherence and the Adjusted Rand Index (ARI; Hubert and Arabie, 1985) between the dominant-topic assignments of each document are measured. Before comparing runs, the topics of each run are aligned with those of the first run using the cosine similarity of their term distributions.
2. Composite score. Each K receives a score that combines four criteria.
3. Sensitivity to the weights. The module checks whether the winning K changes under alternative weighting schemes.

Step 5. Training the final model (Final LDA Model)

model <- readRDS("lda_final_model.rds")

names(model)            # model components
dim(model$phi)          # K topics x terms
dim(model$theta)        # documents x K topics
round(model$prevalence, 1)

# Top ten terms of each topic
textmineR::GetTopTerms(model$phi, M = 10)

Step 6. Temporal trends (Topic Trends)

This module starts from the intensity of each topic by year, computed as the average gamma of the documents in that year, and evaluates its evolution with two complementary tests:

• Ordinary least squares (OLS) regression of intensity on year: slope (reported multiplied by 1000), R-squared and p-value.
• Mann-Kendall test (Mann, 1945), non-parametric: coefficient τ and p-value.

Files generated by the application

Tips and troubleshooting

• Document your decisions. Keep the PRISMA table, the list of added stopwords, the chosen K and the model parameters (Model Evaluation Metrics tab), and store the .rds file of the final model.
• Read the topics, not just the metrics. Coherence and the composite score guide the choice of K, but the final decision should rest on reading the terms and the representative documents of each topic.
• Refine stopwords iteratively. After a first fit, use Refine Stopwords to remove generic domain terms and run the workflow again.
• The DTM is empty or has no terms. Check the chosen text column and relax the sparsity filter (Sparse filter).
• A large file cannot be uploaded. Raise shiny.maxRequestSize before launching the application (see “Installation and launch”).
• LDAvis does not open. Allow pop-ups in the browser for the application address.
• Topic Trends shows no results. Train the final model first: this module uses the topic-by-year intensity computed in Final LDA Model, which requires the imported data to have year and database columns.
• Computations take a long time. LDA Inference and Stability & Sensitivity train many models. Narrow the range of K, use a step larger than 1 for a first exploration, and assign more CPUs.

How to cite

If you use LDAShiny in your research, please cite:

De la Hoz-M., J., Fernandez-Gomez, M. J. and Mendes, S. (2021). LDAShiny: An R Package for Exploratory Review of Scientific Literature Based on a Bayesian Probabilistic Model and Machine Learning Tools. Mathematics, 9(14), 1671. https://doi.org/10.3390/math9141671

References

Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B, 57(1), 289-300.

Blei, D. M., Ng, A. Y. and Jordan, M. I. (2003). Latent Dirichlet Allocation. Journal of Machine Learning Research, 3, 993-1022.

Chang, W., Cheng, J., Allaire, J. J., Sievert, C., Schloerke, B., Xie, Y., et al. shiny: Web Application Framework for R. R package.

Fay, C., Rochette, S., Guyader, V. and Girard, C. (2022). Engineering Production-Grade Shiny Apps. Chapman & Hall/CRC.

Hubert, L. and Arabie, P. (1985). Comparing partitions. Journal of Classification, 2, 193-218.

Mann, H. B. (1945). Nonparametric tests against trend. Econometrica, 13(3), 245-259.

Mimno, D., Wallach, H., Talley, E., Leenders, M. and McCallum, A. (2011). Optimizing semantic coherence in topic models. Proceedings of EMNLP, 262-272.

Page, M. J. et al. (2021). The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ, 372, n71.

Sievert, C. and Shirley, K. (2014). LDAvis: A method for visualizing and interpreting topics. Proceedings of the Workshop on Interactive Language Learning, Visualization, and Interfaces, 63-70.