Repository contains demo code for MTAnchor, an interactive, multilingual topic modeling system. The code accompanies the paper Multilingual Anchoring: Interactive Topic Modeling and Alignment Across Languages (Yuan et al., 2018).
- Python 3
- Numpy
- Scipy
- Scikit-learn
- Flask
- Flask_sqlalchemy
- Flask_migrate
- Anchor-topic
All above packages can be installed with pip install.
- Clone repository:
git clone https://github.com/forest-snow/mtanchor_demo.git - Move to folder:
cd mtanchor_demo - Install dependencies:
pip install -r requirements.txt - Run bash script:
./mtanchor.sh - Go to
http://127.0.0.1:5000/in your browser
- Look at the most likely words for each topic in the blue boxes and the anchor words in the purple boxes.
- Move words from blue boxes to purple boxes to add them as anchors (note that interface will not let you mix words from languages). You may also remove words from the purple boxes to delete them as anchors.
- Hover over words to see their translation (if available).
- Click on words to see occurrences of words and their translations highlighted across the interface.
- Search for words and see what words exist in the corpora. Once you have selected a word from the dropdown menu, then you have the option of adding the word and/or its translation as an anchor to the topic.
- Add or remove topics from the interface.
- Update topics with your newly chosen anchors (note that interface will not let you update topics if no anchor word is chosen).
- Restart your progress if you want to start over.
- End app with
Ctrl+C.
The demo uses a SQLite database (stored in your local files) to save data. Since this is only a demo, only top 1000 words from each corpus is included in the vocabulary. The demo doesn't let user submit their results to prevent database from getting too large. If you find that the database is taking up too much space, run delete_db.sh to delete all data from the database.
If you want the Wikipedia data used in the experiments, you may download it here. The dataset contains Chinese and English Wikipedia articles that are labeled in one of six categories: film, music, animals, politics, religion, and food. Please cite this paper if you use the data.
- Interface was initially based upon Jeff Lund's tbuie interface.
- Topic model is built using code from anchor-topic package.
@inproceedings{yuan2018mtanchor,
title={Multilingual Anchoring: Interactive Topic Modeling and Alignment Across Languages},
author={Yuan, Michelle and Van Durme, Benjamin and Boyd-Graber, Jordan},
booktitle={Advances in neural information processing systems},
year={2018}
}Copyright (C) 2018, Michelle Yuan
Licensed under the terms of the MIT License. A full copy of the license can be found in LICENSE.