Source-linked AI summary
COVID-19 Multidimensional Kaggle Literature Organization
Maksim E. Eren, Nick Solovyev, Chris Hamer, Renee McDonald, Boian S. Alexandrov, Charles Nicholas
TL;DR
The rapid expansion of COVID-19 publications creates a need for methods that organize relevant information in large research corpora. This paper applies tensor factorization to a multidimensional representation of CORD-19, showing that latent components group related articles, journals, authors, and topic keywords. The results are presented through an interactive visualization.
Problem
The growing COVID-19 literature requires improved document organization methods to help identify relevant information.
Method
The paper applies tensor decomposition to a four-dimensional CORD-19 tensor representing first authors, documents, journals, and words.
Results
Tensor-factorization components group similar articles, journals, and researchers while capturing topic keywords across the CORD-19 corpus.
Takeaways & Limitations
Higher-order representation supports simultaneous organization of CORD-19 papers, journals, authors, and keywords, with results available in an interactive visualization.
Abstract
from arXiv · showhide
The unprecedented outbreak of Severe Acute Respiratory Syndrome Coronavirus-2 (SARS-CoV-2), or COVID-19, continues to be a significant worldwide problem. As a result, a surge of new COVID-19 related research has followed suit. The growing number of publications requires document organization methods to identify relevant information. In this paper, we expand upon our previous work with clustering the CORD-19 dataset by applying multi-dimensional analysis methods. Tensor factorization is a powerful unsupervised learning method capable of discovering hidden patterns in a document corpus. We show that a higher-order representation of the corpus allows for the simultaneous grouping of similar articles, relevant journals, authors with similar research interests, and topic keywords. These groupings are identified within and among the latent components extracted via tensor decomposition. We further demonstrate the application of this method with a publicly available interactive visualization of the dataset.
1 INTRODUCTION
The rapid growth of COVID-19 literature makes improved document organization increasingly important. This paper applies tensor-based multidimensional analysis to organize CORD-19 by grouping related papers, journals, authors, and keywords.
- 117,000 new COVID-19 publications appeared in Google Scholar during the seven months before July 2021.
- Over 534,000 COVID-19 publications have appeared since 2020, intensifying the need for better document organization.
- The paper organizes the CORD-19 collection of over 400,000 scholarly articles using a multidimensional analysis method.
- Tensor decomposition represents CORD-19 documents across author, title, journal, and keyword dimensions to extract multifaceted latent patterns.
- Tensor factorization groups similar authors, articles, and journals within and across latent components while identifying topic keywords.
2 RELEVANT WORK
Prior research has widely applied matrix and tensor decomposition to text, medical data, and document organization. The paper situates its four-dimensional CORD-19 analysis among these approaches and related CORD-19 methods.
- Tensor and matrix decomposition are established approaches for text analysis.
- Selecting the number of latent topics remains a difficult problem in both matrix and tensor decomposition.
- Random sampling methods have been used to accelerate large-tensor analysis, with latent factors grouping similar Reddit subreddits.
- Prior biomedical-text work used a Term x Keyword x Document tensor with TF-IDF entries and k-means clustering.
- This paper instead analyzes a four-dimensional tensor to group documents, authors, and journals and extract topic keywords.
- Earlier CORD-19 work applied hierarchical non-negative matrix factorization to a Word x Document representation.
3 METHODS
The method represents the CORD-19 corpus as a four-dimensional tensor and applies CP-based factorization to identify distinct topics and groupings across documents, journals, authors, and words.
- Preprocessing: Preprocessing removes non-English or textless documents, normalizes selected metadata, and drops duplicate titles or abstracts before tensor construction.The corpus is cleaned before tensor analysis, following improvements over the authors’ prior data-cleaning procedure.
- Tensor factorization: CP decomposition represents the tensor with a low-rank approximation formed from latent factor matrices and component weights.The paper uses CP-ALS to compute decompositions at ranks 20, 40, 60, 80, 100, 120, and 200.
- Tensor construction: The corpus is modeled as an order-four tensor spanning first author, document title, journal, and words, with log-transformed word counts as entries.Each entry records how often a first author used a word in a document published in a journal.
- Component selection: Cosine similarity between word-dimension latent factors filters redundant topics before component interpretation.The similarity threshold is 0.35 or higher, yielding 73 selected components from 620.
- Component interpretation: The selected components are interpreted by examining high-valued latent factors for documents, journals, and authors, while word factors form topic word-clouds.The analysis plots entries with the highest values for three latent-factor vectors and uses the fourth dimension for topic keywords.
4 RESULTS
The tensor model makes the corpus interpretable across dimensions, revealing coherent groupings of keywords, papers, journals, and authors within latent components. These groupings capture topics including public health, education, aviation security, vaccination, respiratory research, and niche specialties.
- Interpretability: The extracted latent factors expose meaningful relationships by jointly representing paper, journal, author, and word groupings.This interpretability comes from analyzing information from each tensor dimension simultaneously and manually inspecting the resulting groupings.
- Topic groupings: Topic keywords form coherent groups, including public health, pandemic education, aviation security, vaccination, and respiratory studies.The keyword groupings semantically parallel the documents, journals, and author publications associated with their components.
- Component 56: Component 56 links respiratory-COVID-19 research with asthma, air pollution, and allergy across papers, journals, and author publications.It clusters the World Allergy Organization Journal with other relevant allergy and asthma journals while also showing distinct groupings from other research fields.
- Author groupings: Several components isolate niche authors and same-author textbook chapters while still yielding relevant keywords for those groups.Examples include aviation-law author Ruwantissa Abeyratne and public-health-policy keywords in Theodore Tulchinsky’s components.
5 CONCLUSION
The paper uses a four-dimensional CP-decomposed tensor to organize CORD-19 literature, capturing topic keywords and groupings of related articles, journals, and researchers. The authors identify non-negative factorization and alternative tensor dimensions as future directions.
- Conclusion: A four-dimensional CORD-19 tensor and CP decomposition capture topic keywords and group similar articles, journals, and researchers within and across components.The conclusion describes this as an expansion of prior COVID-19 literature organization using tensor analysis.
- Future work: Future work could replace CP-ALS with non-negative tensor factorization and explore different combinations of tensor dimensions.