nbalepur / dynamite

Code for Findings of ACL 2023 Paper "DynaMiTE: Discovering Explosive Topic Evolutions with User Guidance"

DynaMiTE: Discovering Explosive Topic Evolutions with User Guidance

This repository is the official implementation of "DynaMiTE: Discovering Explosive Topic Evolutions with User Guidance", which was accepted to Findings of ACL 2023

Datasets

The datasets used in our experiments can be found on Huggingface here! There are three splits in the dataset: arxiv, un, and newspop, which correspond to the three datasets used in the paper. Each dataset also has two columns:

• text: The text of the document in the corpus
• time_discrete: The time stamp of the document in the corpus

Requirements

This code was run on Python 3.8.10. We recommend creating a virtual environment to run DynaMiTE.

To install requirements:

pip install -r requirements.txt

Preprocessing

First, create a folder with your dataset name in the data folder (e.g. data/arxiv). To load in your dataset, create a data.csv file where each row is a document. This CSV must contain at least two columns, text and time_discrete, which correspond to the text of the document as well as an ordinal integer representing the time step of the document.

The zip file containing AutoPhrase must also be downloaded from here. This zip file should be moved into the preprocessing folder.

To preprocess the dataset, navigate to /preprocess/, specify the parameters at the top of the preprocess.py file, and run the following command:

python preprocess.py

The folder with your CSV will become populated with more data. Expect ~15min to process each dataset.

We provide links to download the Arxiv, UN, and Newspop datasets.

Training

First, navigate to train_model/train.py and specify your parameters at the top of the file. Then, navigate back to parent directory.

You can train DynaMiTE by running the following command:

python train_model/train.py

You must specify an output folder, which defaults to the results folder. You must also add the same dataset folder to the output folder specified in the preprocessing step (e.g. results/arxiv/). After training, the specified output folder will be populated with the topic evolutions in a text file, along with the embeddings from the discriminative dynamic word embedding space.

Experiments

We include code for running quantitative experiments for NPMI and the category shift analysis. Both experiments require the outputs from training.

You can calculate NPMI by running the following command:

python eval.py

You can run the category experiment through the following command:

python shift_study.py

dynamite's Issues

Memory Issue for Arxiv [OOM error while preprocessing]

I am running the preprocess.py on a 16GB memory. After a while it gives OOM issue. Is there a memory efficient code you have? Or you have tried to make it run with a larger memory.

Results Not Matching with Table 2

Hi Thanks for organizing the datasets.

For each dataset given the seed words in the paper, I obtained the test.txt files with evolving topics. After running the eval I am getting the following. I was trying to match the results with Table 2 but I guess it does not really match.

data	pmi	npmi	lcp	diversity
arxiv	0.702312	0.077952	-0.283055	1.0
un-debates	0.244851	0.023682	-0.322926	1.0
newspop.	0.71346	0.089052	-0.488389	1.0

I wonder how to interpret this result with Table 2 in the paper.

For your convenience I have attached the results folder.
results.zip

This folder contains only test.txt for each.

Preprocessing Issue

While running it on debate data I got the following issues

Traceback (most recent call last):
File "/Users/akalia/Research_Papers/Real-Time-Event-Extraction/Dynamite/DynaMiTE-main/preprocessing/preprocess.py", line 210, in
emb, pmi_samples, cx, cxy, vocab = create_pmi(year_texts_phrase[year])
File "/Users/akalia/Research_Papers/Real-Time-Event-Extraction/Dynamite/DynaMiTE-main/preprocessing/preprocess.py", line 191, in create_pmi
PMI = csc_matrix((data, (rows, cols)))
File "/Users/akalia/miniforge3/lib/python3.9/site-packages/scipy/sparse/_compressed.py", line 53, in init
self._coo_container(arg1, shape=shape, dtype=dtype)
File "/Users/akalia/miniforge3/lib/python3.9/site-packages/scipy/sparse/_coo.py", line 148, in init
raise ValueError('cannot infer dimensions from zero '
ValueError: cannot infer dimensions from zero sized index arrays

So far I have generated the following files.

data.csv
phrase_text.txt
text.txt
wordIDHash.csv
wordPairPMI_2012.csv
wordPairPMI_2013.csv
wordPairPMI_2014.csv
wordPairPMI_2015.csv