Predicting Heursitscs Related to the Controversiality of a Social Media Post

Team: 10

School: Academy For Tech & Classics

Area of Science: Natural Language Processing

Interim:

Problem Definition: Social media platforms are currently confronted with a large amount of controversial content, which carries significant consequences for the emotional health of their users [1]. Without an effective way to classify this content, social media platforms face difficulty when trying to limit such content.

Plan for Solving Problem Computationally: In order to accurately categorize controversial content, some researchers have found success using manually annotated datasets. However, these datasets are limited in size and would require constant updating to stay up to date with what is currently controversial. [4] To avoid these issues, we plan on instead predicting various heuristics which could indicate controversy (rather than controversy itself). For example, on Twitter, we could use the following variables:

  1. Like to retweet ratio
  2. Sentiment of comments (through NLTK [5])
  3. Likes to number of comments
While these do not exactly indicate if a post is controversial, they can still be correlated with the controversy surrounding the post. A successful model which can predict these heuristcs could be expanded upon and used to categorize content as controversial. In order to predict these heuristics ahead of time, we plan on fine tuning a BERT model. With a BERT model, large organizations, such as Google [3], have already trained a representation of language features in a large neural network. We can add our own output layer to this neural network and train that on our specific task. By doing this, we benefit from the deep understanding of language only achievable through large, extremely expensive neural networks, without having to train one ourselves [2]. For the collection of the data which we will use to train our model, we plan to use Twitter, a site well-known for its blatantly controversial, or even offensive, posts. It is worth mentioning that since fluctuations in smaller posts are readily apparent - a post with 2 likes and 2 retweets, for example, could fluctuate to 5 likes and 2 retweets or vice versa - we plan on using tweets that have garnered more public attention. Additionally, we may use Facebook as well, though our attempts at creating a program to scrape data from the platform seem to indicate this to be less than certain. After allowing our model to train on a substantial number of epochs, we would then plan to evaluate the results at a confidence interval of approximately 95% in order to prevent false positives, an error of which many consumers of social media platforms are wary.

Current Progress: Currently, we have made significant progress towards collecting the necessary data which we can use to train our model. While we have not yet deployed our data collection code at scale, we plan to do so very soon, and we have made significant progress writing a program to collect the necessary data from social media platforms. We have a working program to collect data from the Twitter API and have made significant progress towards doing the same with Facebook. We hope to have these programs collecting and storing data very soon, giving us the necessary data to train our model. However, due to Facebook’s restrictions on scraping data, we were unable to get any from the aforementioned site. Additionally, we have written the code which will take our raw text output and represent it as a series of word embeddings, which can be input into our BERT model in order to make predictions.

Expected Results: Because of the extremely stochastic nature of social media, it is fairly unrealistic to expect extraordinary precise accuracy with our model’s predictions. However, we anticipate that our model will be able to give a generally accurate prediction of certain post heuristics, which would align with the post's controversy. While it doesn't make sense to censor all forms of controversial content, the ability to de-prioritize it represents a powerful tool for the reduction of the emotional and societal harm largely caused by social media.

References: [1] William J. Brady, et al. "How social learning amplifies moral outrage expression in online social networks". Science Advances 7. 33(2021): eabe5641. [2] Devlin, Jacob et al. "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding." (2018). [3] Devlin, Jacob, and Ming-Wei Chang. “Open Sourcing Bert: State-of-the-Art Pre-Training for Natural Language Processing.” Google AI Blog, Google, 2 Nov. 2018, https://ai.googleblog.com/2018/11/open-sourcing-bert-state-of-art-pre.html. [4] Mozafari M, Farahbakhsh R, Crespi N (2020) Hate speech detection and racial bias mitigation in social media based on BERT model. PLoS ONE 15(8): e0237861. https://doi.org/10.1371/journal.pone.0237861 [5] NLTK Project. (2023, January 2). nltk.sentiment.sentiment_analyzer module. NLTK. Retrieved January 9, 2023, from https://www.nltk.org/api/nltk.sentiment.sentiment_analyzer.html

Team Members:

  Gene Huntley
  Henry Tischler

Sponsoring Teacher: Jenifer Hooten

Mail the entire Team