turftopic 0.2.4

Topic modeling with contextual representations from sentence transformers.

Topic modeling is your turf too.
Contextual topic models with representations from transformers.

Intentions

• Provide simple, robust and fast implementations of existing approaches (BERTopic, Top2Vec, CTM) with minimal dependencies.
• Implement state-of-the-art approaches from my papers. (papers work-in-progress)
• Put all approaches in a broader conceptual framework.
• Provide clear and extensive documentation about the best use-cases for each model.
• Make the models' API streamlined and compatible with topicwizard and scikit-learn.
• Develop smarter, transformer-based evaluation metrics.

!!!This package is still a prototype, and no papers are published about the models. Until these are out, and most features are implemented I DO NOT recommend using this package for production and academic use!!!

Roadmap

• Model Implementation
• Pretty Printing
• Publish papers :hourglass_flowing_sand: (in progress..)
• Thorough documentation and good tutorials ⏳
• Implement visualization utilites for these models in topicwizard ⏳
• High-level topic descriptions with LLMs.
• Contextualized evaluation metrics.

Implemented Models

Mixture of Gaussians (GMM)

Topic models where topics are assumed to be Multivariate Normal components, and term importances are estimated with Soft-c-TF-IDF.

from turftopic import GMM

model = GMM(10).fit(texts)
model.print_topics()

KeyNMF

Nonnegative Matrix Factorization over keyword importances based on transformer representations.

from turftopic import KeyNMF

model = KeyNMF(10).fit(texts)
model.print_topics()

Semantic Signal Separation (S³)

Interprets topics as dimensions of semantics. Obtains these dimensions with ICA or PCA.

from turftopic import SemanticSignalSeparation

model = SemanticSignalSeparation(10).fit(texts)
model.print_topics()

Clustering Topic Models

Topics are clusters in embedding space and term importances are either estimated with c-TF-IDF (BERTopic) or proximity to cluster centroid (Top2Vec).

from turftopic import ClusteringTopicModel

model = ClusteringTopicModel().fit(texts)
model.print_topics()

Variational Autoencoders (CTM)

Contextual representations are used as ProdLDA encoder network inputs, either alone (ZeroShotTM) or concatenated to BoW (CombinedTM).

pip install turftopic[pyro-ppl]

from turftopic import AutoencodingTopicModel

model = AutoencodingTopicModel(10).fit(texts)
model.print_topics()