Topic Discovery via Latent Space Clustering of Pretrained Language Model Representations

Last update: Dec 18, 2022

Overview

TopClus

The source code used for Topic Discovery via Latent Space Clustering of Pretrained Language Model Representations, published in WWW 2022.

Requirements

At least one GPU is required to run the code.

Before running, you need to first install the required packages by typing following commands (Using a virtual environment is recommended):

pip3 install -r requirements.txt

You need to also download the following resources in NLTK:

import nltk
nltk.download('stopwords')
nltk.download('averaged_perceptron_tagger')
nltk.download('universal_tagset')

Overview

TopClus is an unsupervised topic discovery method that jointly models words, documents and topics in a latent spherical space derived from pretrained language model representations.

Running Topic Discovery

The entry script is src/trainer.py and the meanings of the command line arguments will be displayed upon typing

python src/trainer.py -h

The topic discovery results will be written to results_${dataset}.

We provide two example scripts nyt.sh and yelp.sh for running topic discovery on the New York Times and the Yelp Review corpora used in the paper, respectively. You need to first extract the text files from the .tar.gz tarball files under datasets/nyt and datasets/yelp.

You could expect to obtain results like the following (the Topic IDs are random):

On New York Times:
Topic 20: months,weeks,days,decades,years,hours,decade,seconds,moments,minutes
Topic 28: weapons,missiles,missile,nuclear,grenades,explosions,explosives,launcher,bombs,bombing
Topic 30: healthcare,medical,medicine,physicians,patients,health,hospitals,bandages,medication,physician
Topic 41: economic,commercially,economy,business,industrial,industry,market,consumer,trade,commerce
Topic 46: senate,senator,congressional,legislators,legislatures,ministry,legislature,minister,ministerial,parliament
Topic 72: government,administration,governments,administrations,mayor,gubernatorial,mayoral,mayors,public,governor
Topic 77: aircraft,airline,airplane,airlines,voyage,airplanes,aviation,planes,spacecraft,flights
Topic 88: baseman,outfielder,baseball,innings,pitchers,softball,inning,basketball,shortstop,pitcher

On Yelp Review:
Topic 1: steamed,roasted,fried,shredded,seasoned,sliced,frozen,baked,canned,glazed
Topic 15: nice,cozy,elegant,polite,charming,relaxing,enjoyable,pleasant,helpful,luxurious
Topic 16: spicy,fresh,creamy,stale,bland,salty,fluffy,greasy,moist,cold
Topic 17: flavor,texture,flavors,taste,quality,smells,tastes,flavour,scent,ingredients
Topic 20: japanese,german,australian,moroccan,russian,greece,italian,greek,asian,
Topic 40: drinks,beers,beer,wine,beverages,alcohol,beverage,vodka,champagne,wines
Topic 55: horrible,terrible,shitty,awful,dreadful,worst,worse,disgusting,filthy,rotten
Topic 75: strawberry,berry,onion,peppers,tomato,onions,potatoes,vegetable,mustard,garlic

Running Document Clustering

The latent document embeddings will be saved to results_${dataset}/latent_doc_emb.pt which can be used as features to clustering algorithms (e.g., K-Means).

If you have ground truth document labels, you could obtain the document clustering evaluation results by passing the document label file and the saved latent document embedding file to the cluster_eval function in src/utils.py. For example:

from src.utils import TopClusUtils
utils = TopClusUtils()
utils.cluster_eval(label_path="datasets/nyt/label_topic.txt", emb_path="results_nyt/latent_doc_emb.pt")

Running on New Datasets

To execute the code on a new dataset, you need to

Create a directory named your_dataset under datasets.
Prepare a text corpus texts.txt (one document per line) under your_dataset as the target corpus for topic discovery.
Run src/trainer.py with appropriate command line arguments (the default values are usually good start points).

Citations

Please cite the following paper if you find the code helpful for your research.

@inproceedings{meng2022topic,
  title={Topic Discovery via Latent Space Clustering of Pretrained Language Model Representations},
  author={Meng, Yu and Zhang, Yunyi and Huang, Jiaxin and Zhang, Yu and Han, Jiawei},
  booktitle={The Web Conference},
  year={2022},
}

Topic Discovery via Latent Space Clustering of Pretrained Language Model Representations

Related tags

Overview

TopClus

Requirements

Overview

Running Topic Discovery

Running Document Clustering

Running on New Datasets

Citations

Owner

Yu Meng

PyTorch implementation of Tacotron speech synthesis model.

Code for Universal Semi-Supervised Semantic Segmentation models paper accepted in ICCV 2019

Time Series Cross-Validation -- an extension for scikit-learn

PyTorch implementation of Convolutional Neural Fabrics http://arxiv.org/abs/1606.02492

This is the codebase for the ICLR 2021 paper Trajectory Prediction using Equivariant Continuous Convolution

Course content and resources for the AIAIART course.

Code for our NeurIPS 2021 paper: Sparsely Changing Latent States for Prediction and Planning in Partially Observable Domains

Use Python, OpenCV, and MediaPipe to control a keyboard with facial gestures

High-Resolution Image Synthesis with Latent Diffusion Models

This is the official PyTorch implementation of the CVPR 2020 paper "TransMoMo: Invariance-Driven Unsupervised Video Motion Retargeting".

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)

Joint project of the duo Hacker Ninjas

The official PyTorch code implementation of "Personalized Trajectory Prediction via Distribution Discrimination" in ICCV 2021.

Use CLIP to represent video for Retrieval Task

PyTorch implementation for Graph Contrastive Learning with Augmentations

A lightweight python AUTOmatic-arRAY library.

Repository for publicly available deep learning models developed in Rosetta community

Telegram chatbot created with deep learning model (LSTM) and telebot library.

Everything you need to know about NumPy( Creating Arrays, Indexing, Math,Statistics,Reshaping).

PyTorch wrapper for Taichi data-oriented class