topmost 0.0.2

Towards the Topmost: A Topic Modeling System Tookit

TopMost provides complete lifecycles of topic modeling, including datasets, preprocessing, models, training, and evaluations. It covers the most popular topic modeling scenarios, like basic, dynamic, hierarchical, and cross-lingual topic modeling.

This is our demo paper Towards the TopMost: A Topic Modeling System Toolkit.

This is our survey paper on neural topic models: A Survey on Neural Topic Models: Methods, Applications, and Challenges.

---

Table of Contents

• Overview

• Quick Start

  • Install TopMost

  • Discover topics from your own datasets

• Usage

  • Download a preprocessed dataset

  • Train a model

  • Evaluate

  • Test new documents

• Installation

  • Stable release

  • From sources

• Tutorials

• Notice

• Disclaimer

• Authors

• Contributors

• How to cite our work

• Acknowledgments

Overview

TopMost offers the following topic modeling scenarios with models, evaluation metrics, and datasets:

Scenario	Model	Evaluation Metric	Datasets
Basic Topic Modeling	LDA NMF ProdLDA DecTM ETM NSTM TSCTM ECRTM	TC TD Clustering Classification	20NG IMDB NeurIPS ACL NYT Wikitext-103
Hierarchical Topic Modeling	HDP SawETM HyperMiner ProGBN TraCo	TC over levels TD over levels Clustering over levels Classification over levels	20NG IMDB NeurIPS ACL NYT Wikitext-103
Dynamic Topic Modeling	DTM DETM	TC over time slices TD over time slices Clustering Classification	NeurIPS ACL NYT
Cross-lingual Topic Modeling	NMTM InfoCTM	TC (CNPMI) TD over languages Classification (Intra and Cross-lingual)	ECNews Amazon Review Rakuten

Quick Start

Install TopMost

Install topmost with pip as

$ pip install topmost

Discover topics from your own datasets

We can get the top words of discovered topics, topic_top_words` and the topic distributions of documents, doc_topic_dist. The preprocessing steps are configurable. See our documentations.

import topmost
from topmost.preprocessing import Preprocessing

# Your own documents
docs = [
    "This is a document about space, including words like space, satellite, launch, orbit.",
    "This is a document about Microsoft Windows, including words like windows, files, dos.",
    # more documents...
]

device = 'cuda' # or 'cpu'
preprocessing = Preprocessing()
dataset = topmost.data.RawDatasetHandler(docs, preprocessing, device=device, as_tensor=True)

model = topmost.models.ProdLDA(dataset.vocab_size, num_topics=2)
model = model.to(device)

trainer = topmost.trainers.BasicTrainer(model)

topic_top_words, doc_topic_dist = trainer.fit_transform(dataset, num_top_words=15, verbose=False)

Usage

Download a preprocessed dataset

import topmost
from topmost.data import download_dataset

download_dataset('20NG', cache_path='./datasets')

Train a model

device = "cuda" # or "cpu"

# load a preprocessed dataset
dataset = topmost.data.BasicDatasetHandler("./datasets/20NG", device=device, read_labels=True, as_tensor=True)
# create a model
model = topmost.models.ProdLDA(dataset.vocab_size)
model = model.to(device)

# create a trainer
trainer = topmost.trainers.BasicTrainer(model)

# train the model
trainer.train(dataset)

Evaluate

# get theta (doc-topic distributions)
train_theta, test_theta = trainer.export_theta(dataset)
# get top words of topics
topic_top_words = trainer.export_top_words(dataset.vocab)

# evaluate topic diversity
TD = topmost.evaluations.compute_topic_diversity(top_words)

# evaluate clustering
clustering_results = topmost.evaluations.evaluate_clustering(test_theta, dataset.test_labels)

# evaluate classification
classification_results = topmost.evaluations.evaluate_classification(train_theta, test_theta, dataset.train_labels, dataset.test_labels)

Test new documents

import torch
from topmost.preprocessing import Preprocessing

new_docs = [
    "This is a new document about space, including words like space, satellite, launch, orbit.",
    "This is a new document about Microsoft Windows, including words like windows, files, dos."
]

parsed_new_docs, new_bow = preprocessing.parse(new_docs, vocab=dataset.vocab)
new_doc_topic_dist = trainer.test(torch.as_tensor(new_bow, device=device).float())

Installation

Stable release

To install TopMost, run this command in your terminal:

$ pip install topmost

This is the preferred method to install TopMost, as it will always install the most recent stable release.

From sources

The sources for TopMost can be downloaded from the Github repository. You can clone the public repository by

$ git clone https://github.com/BobXWu/TopMost.git

Then install the TopMost by

$ python setup.py install

Tutorials

We provide tutorials for different usages:

Name	Link
Quickstart
How to preprocess datasets
How to train and evaluate a basic topic model
How to train and evaluate a hierarchical topic model
How to train and evaluate a dynamic topic model
How to train and evaluate a cross-lingual topic model

Notice

Differences from original implementations

1. Oringal implementations may use different optimizer settings. For simplicity and brevity, our package by default uses the same setting for different models.

Disclaimer

This library includes some datasets for demonstration. If you are a dataset owner who wants to exclude your dataset from this library, please contact Xiaobao Wu.

Authors

Xiaobao Wu

Fengjun Pan

Contributors

How to cite our work

If you want to use our toolkit, please cite as

@article{wu2023topmost,
title={Towards the TopMost: A Topic Modeling System Toolkit},
author={Wu, Xiaobao and Pan, Fengjun and Luu, Anh Tuan},
journal={arXiv preprint arXiv:2309.06908},
year={2023}
}

@article{wu2023survey,
    title={A Survey on Neural Topic Models: Methods, Applications, and Challenges},
    author={Wu, Xiaobao and Nguyen, Thong and Luu, Anh Tuan},
    journal={Artificial Intelligence Review},
    url={https://doi.org/10.1007/s10462-023-10661-7},
    year={2024},
    publisher={Springer}
}

Acknowledgments

• If you want to add any models to this package, we welcome your pull requests.

• If you encounter any problem, please either directly contact Xiaobao Wu or leave an issue in the GitHub repo.

• Icon by Flat-icons-com.