The repo contains the source code for the paper: link here
To cluster the word embeddings to discover the latent topics, run the code/score.py file. Here are the arguments that can be passed in:
--entities : The type of pre-trained word embedding you are clustering with
choices= word2vec, fasttext, glove, KG
KG stands for your own set of embeddings
--entities_file: The file name contain the embeddings
--clustering_algo: The clustering algorithm to use
choices= KMeans, SPKMeans, GMM, KMedoids, Agglo, DBSCAN , Spectral, VMFM
--vocab: List of vocab files to use for tokenization
--dataset: Dataset to test clusters against against
default = 20NG
choices= 20NG, reuters
--preprocess: Cuttoff threshold for words to keep in the vocab based on frequency
--use_dims: Dimensions to scale with PCA (much be less than orginal dims)
--num_topics: List of number of topics to try
default: 20
--doc_info: How to add document information
choices= DUP, WGT
--rerank: Value used for reranking the words in a cluster
choices=tf, tfidf, tfdf
Example call:
python3 code/score.py --entities KG --entities_file {dest_to_entities_file} --clustering_algo GMM --dataset reuters --vocab {dest_to_vocab_file} --num_topics 20 50 --doc_info WGT--rerank tf