BBC Topic Modeling

Unsupervised BBC News Topic Modeling using BERTopic. Although BBCNews is typically a benchmark dataset for supervised document classification, in this repo we will approach it as an unsupervised learning problem.

How to use the repo

To use the repo:

1. Clone the repo git clone https://github.com/MauroCE/BBCTopicModeling.git
2. Create a virtual environment and install dependencies python3 -m venv bbctopic, source bbctopic/bin/activate, pip install -r requirements.txt.
3. Download data, unzip it and change the name of the resulting folder to data.
4. (Optional) create .env file and set TOKENIZERS_PARALLELISM=false

Model

I settled for BERTopic due to its ease of use, modularity (choices of clustering, dimensionality reduction, embeddings, etc) and good documentation. Additionally, it supports many machine learning modalities off-the-shelf, including unsupervised topic modeling, guided topic modeling and supervised document classification, among many others. Another reason for choosing this model/package is that it can seamlessly be used for dealing with a stream of incoming documents by using different clustering/dimensionality-reduction algorithms, as explained here.

Evaluation

Evaluating an unsupervised topic model is not straightforward and requires a number of choices. Due to time constraints, here I have focused on evaluation metrics as provided by gensim and, when BERTopic is instantiated with HDBSCAN, I also compute perplexity. I have left a notebook where I run BERTopic with k-means (with 5 centroids) so that this can be compared directly with the original categories the articles were in.

Dataset

To work with this repo, download the raw data, unzip it and call it data. The code in this repo expects the data folder to have sub-folders for each category tech, politics, business, entertainment, sport. This is exactly the structure obtained upon unzipping the data from the url above.

Testing

Testing is performed with pytest as it requires less boilerplate code and more plug-in functionalities. Due to time constraint, I have focused on testing the data loading and data formatting. In a production environment one would need to list all the features, inputs and expected outputs in the pipeline and test them exhaustively.

Notes

The file .env is not being tracked. If you want to suppress the tokenizer warning simply create a .env file and set TOKENIZERS_PARALLELISM=false.