Topic modelling learning journal (part 3)

Created on March 03, 2026

2026   ·   research

In my previous post, topic modelling learning journal (part 2), I explored the overall structure of the ARC Discovery Projects dataset by projecting text data into a two-dimensional space.

In this final post of the series, I move one step further and use Latent Dirichlet Allocation (LDA) to uncover the underlying topics that connect these projects.

Discovering underlying topics

To extract explicit themes from the project summaries, I trained a topic model using a Bag-of-Words representation together with LDA.

Training an LDA model itself is relatively straightforward, but one hyperparameter plays a particularly important role: the number of topics.

Unfortunately, there is no universally correct answer. The “best” number depends not only on statistical measures but also on whether the resulting topics are interpretable to humans.

To guide this choice, I experimented with several different values. I first evaluated the models using perplexity and coherence scores, which are commonly used metrics for topic models. The results did not show a very clear optimum, but six topics appeared to be a reasonable compromise.

I then manually inspected the results by printing the top keywords for each topic and looking at representative project summaries associated with them. The themes looked reasonably coherent and interpretable, so I decided to proceed with six topics for now, acknowledging that this choice is somewhat subjective.

Interpreting the topics

To better understand what each topic represents, I calculated the lift of each keyword. Lift measures how uniquely a word is associated with a particular topic compared to the rest of the corpus.

In practice, this helps filter out very common words that appear across many documents (for example domain-specific stopwords such as project or aim), allowing the more distinctive topic-specific terms to stand out.

The visualization below shows the top 30 high-lift keywords for each topic using a Termite-style topic–term matrix, a design originally proposed by researchers at Stanford. In this plot, each column represents a topic, each row represents a keyword, and the size of the bubble indicates the lift value.

The Termite-style topic–term matrix reveals a very clear pattern: most high-lift words are strongly associated with only one topic. This suggests that the model has learned well-separated themes with relatively little overlap in vocabulary.

In other words, each topic appears to correspond to a fairly distinct research domain, such as biology, materials science, ecology, or artificial intelligence.

Based on the high-lift keywords, I asked my best friend ChatGPT to help summarise what each topic seems to represent. For each topic, I also linked the top-ranked Discovery Project (DP) summary associated with that topic.

Topics across research fields

The topic keywords already provide a good intuition about the disciplines involved. However, to see whether the model aligns with existing classifications, I compared the inferred topics with the Field of Research (FOR) codes associated with each project.

In the figure below, each circle represents a topic–discipline combination, and the size of the circle corresponds to the number of projects assigned to that topic.

Some interesting patterns emerge.

Topics 3 and 4 appear to be more closely related to engineering and chemical sciences, while topics 1 and 5 seem more strongly connected to biological sciences. Topic 0 appears to correspond mainly to social science research.

One topic stood out to me in particular: Topic 2, which is dominated by keywords related to artificial intelligence.

Unlike the other topics, it appears as a distinct theme that cuts across multiple research areas. While AI is often thought of primarily as a computer science discipline, the model suggests that AI-related projects appear across a much broader range of fields. In this dataset, they are especially common in engineering, followed by mathematics and even psychology.

The popularity of AI is not too surprising, but seeing this pattern emerge automatically from text alone still felt quite remarkable.

Another question that naturally came to mind is how these topics evolve over time. In particular, I was curious about how funding is distributed across topics between 2023 and 2025.

The figure below shows the proportion of projects associated with each topic across different funding years. The size of the bubbles represents the total funding allocated to each topic.

Interestingly, the AI-related topic does not show a dramatic increase in either proportion or funding share over this period.

Given the strong public narrative about rapidly increasing AI investment, I expected to see a more obvious shift. At least according to this dataset and this probabilistic topic model, the distribution appears relatively stable.

Of course, this result should be interpreted cautiously: topic models are approximate, and small datasets can easily hide subtle trends.

At the end of this three-part series, I’m quite happy with the results I have so far. They are certainly far from perfect, especially compared with modern NLP architectures based on transformers, but I am still amazed by how much structure can be extracted using relatively simple statistical models.

If I continue exploring more modern approaches in the future, I would love to compare them with these traditional methods.

So stay tuned!



Enjoy Reading This Article?

Here are some more articles you might like to read next:

  • Topic modelling learning journal (part 2)
  • Topic modelling learning journal (part 1)
  • Finding my employability after the PhD
  • DIY Perovskite crystal model - Origami, knitting and 3D printing
  • From a distant dream to reality - my AI4Science journey