Topic modelling and NLP on scientific literature
Unsupervised extraction of latent topics from a scientific corpus using topic modelling and sentiment analysis
Exploring the thematic structure of a field without starting from predefined categories
Mapping the dominant topics of a scientific corpus by manual review is slow and depends on the judgement of whoever does the classifying. The challenge was to identify the thematic structure from the text itself, reproducibly and without predefined categories.
Classic narrative reviews rely on the reader's judgement to decide which topics group a set of articles together. That judgement introduces a bias that is hard to audit, and it does not scale as the corpus grows. In occupational health and home care services —a field that is expanding with population ageing and hospital-at-home care— a method was needed that would identify the focus areas of the literature systematically and replicably.
The task was to apply unsupervised natural language processing to the corpus of bibliographic records, so that the model would identify latent patterns without predefined thematic categories, for the researchers to interpret and label afterwards. This was combined with a lexical characterisation of the sentiments and emotions in the field's scientific language, making sure the interpretation did not go beyond what a descriptive analysis can support.
70 articles
clinical trials selected from an initial screening of 1,889 references
4 latent topics
identified in the corpus through unsupervised modelling
8 emotions
from Plutchik's model, quantified with the NRC lexicon
Topic modelling and sentiment analysis of the corpus, reproducible from start to finish
Building and cleaning the document corpus
Systematic retrieval of bibliographic records from MEDLINE (via PubMed) using search strategies built on the field's MeSH and DeCS descriptors, with a clinical trial filter to narrow down the evidence. Data integrity checked with a double extraction table, and the corpus stored as normalised plain text before it entered the analysis.
Unsupervised topic modelling with Latent Dirichlet Allocation
Application of the LDA algorithm, which treats each document as a mixture of topics and each topic as a mixture of words, so that documents can overlap in content instead of being forced into discrete groups. The model derives the thematic groupings from recurring patterns in the corpus, without predefined categories, and the highest-weighted terms are identified for each topic.
Sentiment analysis and lexical characterisation of sentiments and emotions
Lexical characterisation of the sentiments and emotions in the corpus with the NRC lexicon, which links each term to the eight emotions of Plutchik's wheel and to the two general sentiments. Comparison of the distribution of sentiments and emotions using non-parametric tests (Mann-Whitney U, Kruskal-Wallis), presented through word clouds, a radar chart and trends over time. The entire analysis carried out in R with a reproducible pipeline (`topicmodels`, `tidytext`).
A reproducible thematic snapshot of the field's scientific literature
Four latent topics identified reproducibly
Without supervision, the model identified the four focus areas of the corpus: home care and caregiver satisfaction, the breastfeeding period, rehabilitation programmes, and physical activity for pain relief. The thematic structure was obtained through a reproducible computational procedure, with no predefined initial categories, although the final interpretation and naming of the topics were the researchers' work.
Lexical characterisation of sentiments and emotions in scientific language
The sentiment analysis quantified a low frequency of emotion-related terms and a predominance of positive over negative sentiment, with the terms the lexicon assigns to the trust category the most frequent. The pattern remained stable over the period studied, in keeping with the neutral register of scientific writing, and the differences between sentiments were statistically significant.
A method that transfers to any scientific corpus
The procedure —building the corpus, modelling topics and interpreting them— is independent of the specific domain and can be applied to any field where the literature needs to be mapped systematically: monitoring research trends, exploratory work ahead of a systematic review, or research output analysis.
What this case shows about the way we work
Topics emerge from the text, not from predefined categories
The unsupervised approach lets the thematic structure emerge from the corpus, reducing the imposition of a priori categories and making the classification process auditable.
Reproducible and auditable
The whole pipeline runs from R code, so the analysis can be reproduced with the same data, parameters, versions and random seed, and every decision is traceable, in contrast to the opacity of a synthesis done by hand.
Two layers on the same corpus
Topic modelling and sentiment analysis are applied to the same text, combining which topics dominate the field with how the language that expresses them reads.
Backed by peer-reviewed publication
The methodology is part of a PhD thesis by publication and was published in a peer-reviewed scientific journal, which supports the rigour of the procedure beyond an internal demonstration.
Resulting scientific output
Capabilities deployed in this project
Shall we talk?
Tell us what you are working on and we will give you a straight answer on how we can help.