Topic Modeling
Topic modeling is an unsupervised natural language processing method that infers latent themes from a collection of documents, representing each document as a mixture of topics and each topic as a weighted distribution over words. Latent Dirichlet Allocation (LDA) is the most common algorithm; results depend on text preprocessing (tokenization, stop words, stemming), are judged with coherence measures and expert reading of top words, and can vary between runs because of random initialization.
Ensemble LDA for a scoping review of TAM-targeting nanoparticles (2024)
Brown and coworkers applied Gensim’s ensemble LDA, which clusters topics from many LDA runs with a DBSCAN variant to keep stable ones, to 854 screened abstracts on nanoparticles targeting tumor associated macrophages, after plain LDA gave noisy topics insensitive to hyperparameter tuning (Brown et al. 2024). Preprocessing used NLTK tokenization, English plus search-string-derived domain stop words, and Porter stemming. The model reached a Cv coherence of 0.589 and gave six expert-interpretable topics (Immune and TAMs, Nanoparticles, Imaging, Biologics Delivery, Vaccines, Multi-modal Therapies), though only the two largest reappeared consistently across reruns. Applied without retraining to 95 newer abstracts it recovered five of the six topics, a proof of concept for a living review. Summary: brown2024.