BERTopic is a topic modeling technique that leverages 🤗 transformers and c-TF-IDF to create dense clusters allowing for easily interpretable topics whilst keeping important words in the topic descriptions. It even supports visualizations similar to LDAvis!

Corresponding medium post can be found here and here.

Installation can be done using pypi:

pip install bertopic

To use the visualization options, install BERTopic as follows:

pip install bertopic[visualization]

To use Flair embeddings, install BERTopic as follows:

pip install bertopic[flair]

For an in-depth overview of the features of BERTopic you can check the full documentation here or you can follow along with the Google Colab notebook here.

We start by extracting topics from the well-known 20 newsgroups dataset which is comprised of english documents:

from bertopic import BERTopic
from sklearn.datasets import fetch_20newsgroups
 
docs = fetch_20newsgroups(subset='all',  remove=('headers', 'footers', 'quotes'))['data']

topic_model = BERTopic()
topics, _ = topic_model.fit_transform(docs)

After generating topics and their probabilities, we can access the frequent topics that were generated:

>>> topic_model.get_topic_freq().head()
Topic	Count
-1	7288
49	3992
30	701
27	684
11	568

-1 refers to all outliers and should typically be ignored. Next, let's take a look at the most frequent topic that was generated, topic 49:

>>> topic_model.get_topic(49)
[('windows', 0.006152228076250982),
 ('drive', 0.004982897610645755),
 ('dos', 0.004845038866360651),
 ('file', 0.004140142872194834),
 ('disk', 0.004131678774810884),
 ('mac', 0.003624848635985097),
 ('memory', 0.0034840976976789903),
 ('software', 0.0034415334250699077),
 ('email', 0.0034239554442333257),
 ('pc', 0.003047105930670237)]

NOTE: Use BERTopic(language="multilingual") to select a model that supports 50+ languages.

After having trained our BERTopic model, we can iteratively go through perhaps a hundred topic to get a good understanding of the topics that were extract. However, that takes quite some time and lacks a global representation. Instead, we can visualize the topics that were generated in a way very similar to LDAvis:

topic_model.visualize_topics()

The parameter embedding_model takes in a string pointing to a sentence-transformers model, a SentenceTransformer, or a Flair DocumentEmbedding model.

Sentence-Transformers
You can select any model from sentence-transformers here and pass it through BERTopic with embedding_model:

from bertopic import BERTopic
topic_model = BERTopic(embedding_model="xlm-r-bert-base-nli-stsb-mean-tokens")

Or select a SentenceTransformer model with your own parameters:

from bertopic import BERTopic
from sentence_transformers import SentenceTransformer

sentence_model = SentenceTransformer("distilbert-base-nli-mean-tokens", device="cpu")
topic_model = BERTopic(embedding_model=sentence_model)

Flair
Flair allows you to choose almost any embedding model that is publicly available. Flair can be used as follows:

from bertopic import BERTopic
from flair.embeddings import TransformerDocumentEmbeddings

roberta = TransformerDocumentEmbeddings('roberta-base')
topic_model = BERTopic(embedding_model=roberta)

You can select any 🤗 transformers model here.

Custom Embeddings
You can also use previously generated embeddings by passing it through fit_transform():

topic_model = BERTopic()
topics, _ = topic_model.fit_transform(docs, embeddings)

To cite BERTopic in your work, please use the following bibtex reference:

@misc{grootendorst2020bertopic,
  author       = {Maarten Grootendorst},
  title        = {BERTopic: Leveraging BERT and c-TF-IDF to create easily interpretable topics.},
  year         = 2020,
  publisher    = {Zenodo},
  version      = {v0.5.0},
  doi          = {10.5281/zenodo.4430182},
  url          = {https://doi.org/10.5281/zenodo.4430182}
}