---

AU Interacting Minds Centre, NLP Workshop - November 7th

Stance Detection & Topic Modelling of Social Media Users' Content

Rebekah Baglini, Luca Nannini, and Arnault-Quentin Vermillet

Link

Program

1. Data Preprocessing

• Load Dataset
• Tokenization/Stopword Removal
• Clean Tweets Strings with Regular Expressions
• Lemmatization/Stemming

2. Topic modeling

• Create, Run, and Train the HDP model via Gensim
• Visualize topics through an interactive graphs - pyLDAvis
• Visualize cosine metrics of topics as a heatmap
• HDP and LDA via Gensim Models

3. Supervised text classification with BERT

Datasets

Vacc_tweets_raw_n5000.csv

• Random sample of 5000 (out of > 1 million) tweets from 2019 containing string 'vaccin'
• Collected using GetOldTweets3 scraper

Additional sets in Data folder

5-topic_stance_tweets_training_n2814.csv

• Training set from SemEval2016 Task 6 for stance detection task
• Labels: FAVOR, AGAINST, UNKNOWN
• Topics:
  • Atheism
  • Climate change is a concern
  • Feminist movement
  • Hillary Clinton
  • Legalization of abortion

Vacc_articles_w_stance_n3303.csv

• From Vaccine sentiment project, contains 3303 sentences extracted from online articles labelled pro (n=24), neg (n=22), and neu(tral) (n=7).

Vacc_tweets_w_stance_n1131.csv

• 1131 tweets containing string 'vaccin' labeled for stance
• Labels = pro, anti, unknown

Or upload your own dataset!

• Using the upload widget in the Colab file.

Code files

• During the tutorial, we will be working from a Google Colab notebook. This means you will not have to install or load anything locally.
• If you'd like to run locally, we've included list of dependencies in Requirements.txt