The source code used for paper "Effective Seed-Guided Topic Discovery by Integrating Multiple Types of Contexts", published in WSDM 2023.
We use two benchmmark datasets, NYT and Yelp, in our paper, adapted from here. We use 60% as training corpus and the remaining 40% for evaluation.
Use the following command to generate PLM embeddings for the training corpus (gpu required)
python plm_emb.py
Before the first run, compile CatE by
cd cate
make cate
cd ..
Then run the following command for SeedTopicMine
python main.py --dataset nyt --topic locations
4 baselines are compared in our paper: SeededLDA, Anchored CorEx, KeyETM, and CatE.
To reproduce the results of SeededLDA and Anchored CorEx, please refer to ./baselines/SeededLDA.py and ./baselines/AnchoredCorEx.py, respectively.
To reproduce the results of KeyETM and CatE, please refer to their GitHub repositories (i.e., KeyETM and CatE).
To compute P@k and NDCG@k scores of SeedTopicMine and the baselines, we invite five annotators to independently judge if each discovered term is discriminatively relevant to a seed. We release the annotation results in ./annotations/. For example, ./annotations/yelp_sentiment_annotation.txt is as follows:
Term Annotator1 Annotator2 Annotator3 Annotator4 Annotator5
also none none none none none
amazing good none good good good
anger bad bad bad bad bad
apathetic bad bad bad bad bad
appalling bad bad bad bad bad
There are 6 columns. The first column is the term. The other 5 columns are the relevant category of the term according to the 5 annotators, respectively. If a term is relevant to more than one category or is irrelevant to any category, the category will be marked as "none".
If you find the implementation useful, please cite the following paper:
@inproceedings{zhang2023effective,
title={Effective Seed-Guided Topic Discovery by Integrating Multiple Types of Contexts},
author={Zhang, Yu and Zhang, Yunyi and Michalski, Martin and Jiang, Yucheng and Meng, Yu and Han, Jiawei},
booktitle={WSDM'23},
pages={429--437},
year={2023}
}