Repository for WOAH paper "StereoKG: Data-Driven Knowledge Graph Construction for Cultural Knowledge and Stereotypes" (paper)
Software:
PRAW
NLTK
PyTorch
SimpleTransformers
HuggingFace Transformers
- PRAW - Set up a Reddit Account
- Twitter - Access Twitter Developer API Credentials
In Credentials.py, include all metadata like the subreddits to query, Twitter developer credentials, absolute paths to respective directories, and a list of all subject entities in the KG.
The question templates are stored for each entity in the questions folder. For using additional entities, simply create a questions file for that entity and enter the questions in the format
subject, <question pattern>
- Scraping Reddit - scrape_reddit.py
- Scraping Twitter - scrape_twitter.py
The extracted questions are inherently converted to sentences and stored in the filepath specified in Credentials.py.
germans are obsessed with the wednesday frog meme
indians are inherently happy
christians are supporting donald trump
Muslim women wear burkha
fast_clustering.py can be run to perform clustering on the sentences. If the create_singleton_mode is true, separate files are created for singleton and non-singleton clusters.
['Christians believe everyone else should live and abide by their rules', 'Christians love to assume that everyone is supposed to live by the rules of their religion']
['indians are so loud', 'indians are always so loud in public places', 'indians are such noisy eaters']
['French people speak so fast', 'Serious question why do french people talk so quick', 'French people speak so quickly I am just a child']
triple_generation.py can be used to extract triples from sentences in clusters using a Python based OpenIE wrapper.
This process uses the DistilBERT-CoLA model for choosing the most grammatically appropriate triple from the cluster of triple samples.
<indians, don't sacrifice, freedom>
<christians, don't remember, sabbath day>
<atheists, are at, peace with death>
The resultant KG of triples is saved in kg/stereoKG.tsv.
Using the triple_to_text model trained on WebNLG data, this same KG is converted to a verbalized form and saved in kg/stereoKG_linearised.txt.
The code for language modeling integration experiments with intermediate pretraining can be found in lm_integration.
The best models in our research paper are Twitter-RoBERTa models with intermediate pretraining on the structured and unstructured KG triples. They can be found at the following links:
- StereoKG-DT-SK - Domain (Twitter) pre-trained RoBERTa with intermediate pretraining on verbalised triples.
- StereoKG-DT-UK - Domain (Twitter) pre-trained RoBERTa with intermediate pretraining on scraped sentences.