We modify the code to train on abstracts of papers from arxiv.org.
In main.py
:
We try to match the settings from Dieng's paper (https://arxiv.org/abs/1907.05545), fix run time errors, add comments. We also added a warning when the vocabulary is not prefitted enough (this is important).
In scripts/data_undebates
:
We modify so that it can process the meta data json file contains arxiv abstracts, with the option to select a category (default category: high energy physics phenomenonlogy hep-ph
).
In plot_word_evolution.py
:
We refactorized the code to make it clearer. This script is not plug and play, many lines must be hard-coded.
New file: requirements.txt
.
The plot below shows results for DETM on hep-ph
(high energy physics phenomenology) category. Six topics out of 50 are shown here. For each topics, probabilities of some selected words are plotted against time (2007-2020). In topics #33 and #34, peak probability of the word 750
coincides with the flurry of papers on a possible discovery of new physics around 2015-2016, which turned out to be just a statistical fluke. Topic 38 shows the increase in higgs
around the time of the discovery of Higgs boson in 2012.
Other steps involving, from getting data, preprocessing, embedding words and runtime, etc can be found here: https://github.com/quynhneo/detm-arxiv