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.

BERTopic supports guided, supervised, semi-supervised, manual, long-document, hierarchical, class-based, dynamic, and online topic modeling. It even supports visualizations similar to LDAvis!

Corresponding medium posts can be found here, here and here. For a more detailed overview, you can read the paper or see a brief overview.

Installation, with sentence-transformers, can be done using pypi:

pip install bertopic

If you want to install BERTopic with other embedding models, you can choose one of the following:

pip install bertopic[flair]
pip install bertopic[gensim]
pip install bertopic[spacy]
pip install bertopic[use]

For an in-depth overview of the features of BERTopic you can check the full documentation or you can follow along with one of the examples below:

Name	Link
Topic Modeling with BERTopic
(Custom) Embedding Models in BERTopic
Advanced Customization in BERTopic
(semi-)Supervised Topic Modeling with BERTopic
Dynamic Topic Modeling with Trump's Tweets
Topic Modeling arXiv Abstracts

We start by extracting topics from the well-known 20 newsgroups dataset containing 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, probs = topic_model.fit_transform(docs)

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

>>> topic_model.get_topic_info()

Topic	Count	Name
-1	4630	-1_can_your_will_any
0	693	49_windows_drive_dos_file
1	466	32_jesus_bible_christian_faith
2	441	2_space_launch_orbit_lunar
3	381	22_key_encryption_keys_encrypted

The -1 topic refers to all outlier documents and are typically ignored. Next, let's take a look at the most frequent topic that was generated:

>>> topic_model.get_topic(0)

[('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)]

Using .get_document_info, we can also extract information on a document level, such as their corresponding topics, probabilities, whether they are representative documents for a topic, etc.:

>>> topic_model.get_document_info(docs)

Document                               Topic	Name	                        Top_n_words                     Probability    ...
I am sure some bashers of Pens...	0	0_game_team_games_season	game - team - games...	        0.200010       ...
My brother is in the market for...      -1     -1_can_your_will_any	        can - your - will...	        0.420668       ...
Finally you said what you dream...	-1     -1_can_your_will_any	        can - your - will...            0.807259       ...
Think! It's the SCSI card doing...	49     49_windows_drive_dos_file	windows - drive - docs...	0.071746       ...
1) I have an old Jasmine drive...	49     49_windows_drive_dos_file	windows - drive - docs...	0.038983       ...

Note

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

After having trained our BERTopic model, we can iteratively go through hundreds of topics to get a good understanding of the topics that were extracted. 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()

Find all possible visualizations with interactive examples in the documentation here.

By default, the main steps for topic modeling with BERTopic are sentence-transformers, UMAP, HDBSCAN, and c-TF-IDF run in sequence. However, it assumes some independence between these steps which makes BERTopic quite modular. In other words, BERTopic not only allows you to build your own topic model but to explore several topic modeling techniques on top of your customized topic model:

You can swap out any of these models or even remove them entirely. Starting with the embedding step, you can find out how to do this here and more about the underlying algorithm and assumptions here.

BERTopic has many functions that quickly can become overwhelming. To alleviate this issue, you will find an overview of all methods and a short description of its purpose.

Below, you will find an overview of common functions in BERTopic.

After having trained your BERTopic model, several attributes are saved within your model. These attributes, in part, refer to how model information is stored on an estimator during fitting. The attributes that you see below all end in _ and are public attributes that can be used to access model information.

There are many different use cases in which topic modeling can be used. As such, several variations of BERTopic have been developed such that one package can be used across many use cases.

Evaluating topic models can be rather difficult due to the somewhat subjective nature of evaluation. Visualizing different aspects of the topic model helps in understanding the model and makes it easier to tweak the model to your liking.

To cite the BERTopic paper, please use the following bibtex reference:

@article{grootendorst2022bertopic,
  title={BERTopic: Neural topic modeling with a class-based TF-IDF procedure},
  author={Grootendorst, Maarten},
  journal={arXiv preprint arXiv:2203.05794},
  year={2022}
}