/DTV

Source code for the Dynamic Topic Vector model.

Primary LanguagePython

DTV

Source code for the Dynamic Topic Vector model. Code to reproduce the results from experiments on the Financial-News dataset from https://github.com/philipperemy/financial-news-dataset.

Requirements

Code requires a gensim-compatible version of GloVe embeddings which can be obtained by calling:

./models/download_data.sh

Pre-processed document embeddings are included in /data/FinancialNews/ and the gensim-usable GloVe model is stored at /models/glove2word2vec300d.txt.

Experiments

To run one experiment, specify a SAVEDIR and optionally a list of number of topics X in

./run_experiments_FN.sh,

then run from a shell.

After successful training, for every year results are stored in the specified SAVEDIR. This includes a pickled results dictionary res_n_X.p and a topics_n_X.txt file. The latter contains the top-10 words descriptive for one generated topic.