Using NLP to identify online harassment
Like this project? Contact me on LinkedIn
- Link to data: Jigsaw Toxic Comment Classification Challange
- Slideshow
- Collab Notebook
- BERT Collab Notebook
Online harassment is a pervasive issue that affects a significant portion of the population, with serious consequences for individuals and society. According to a 2021 Pew research poll, a majority of respondents believe that social media companies are not doing enough to address online harassment. Harassment affects everyone, but has a more significant impact on BIPOC and LGBTQIA communities, as well as women.
Social media platforms play a crucial role in disseminating information and shaping public sentiment. Therefore, use of social media sites through creation of a hostile environment by an extreme few can not, and should not be condoned. Detecting and addressing hate speech and toxic comments are essential tasks in the information age.
Our client, a small social media site that specializes in neighborhood forums, is looking to combat an uptick in trolling on their site. The goal of this capstone project is to build a virtual moderator (vMOD) using Natural Language Processing (NLP) techniques. The vMOD will accurately predict whether a comment is harassment or not. This model will not only help in reducing the impact of toxic commentary but also alleviate the financial onus of needing to employ a large group of human moderators, in addition to lessening the emotional toll of the remaining human moderators.
Our dataset is derived from the Kaggle competition titled Identifying Toxic Comments: Jigsaw Toxic Comment Classification Challenge. It contains 159,571 comments from Wikipedia, with six target variables: toxic, severe_toxic, obscene, threat, insult, and identity_hate.
We began by cleaning and preprocessing the data. We used two tecniques for preprocessing:
- A regex function for our MNB model:
- remove stopwords
- removing non-ASCII characters, websites, and non-words
- replacing contractions
- removing punctuation and blank space
- removing stopwords
- For our model with embeddings we:
- tokenized data
- padded sequences
We also considered the length of comments to prepare for tokenization.
We primarily explored two different models, a Multinomial Naive Bayes model with binarized target data and oversampling to account for our class imbalance, and a neural network with embeddings.
We encountered challenges in reducing false negatives and false positives- Below, we can see the confusion matrix results of out MNB, as well as our model with embeddings, demonstrating the difficulty our models had with minimizing both false negatives and false positives. While working with global models, we used F1 scores to evaluate our results for the aforementioned reason, and had our best results with the MNB model (.94 for nontoxic, and .63 for toxic). However, in order to compare our MNB with our model with embeddings, we turn to accuracy as well as evaluating our confusion matrices for a more nuanced understanding of our models performance. Our model with embeddings had it's best F1 score in it's fourth epoch, with a F1 score of 56%, and accuracy of 94% on the validation data.
MULTINOMIAL NAIVE BAYES
MODEL WITH EMBEDDINGS
Our capstone project aims to address the critical issue of online harassment through NLP models. While we've made progress in classifying toxic comments, there's room for improvement, especially in reducing false negatives. Our initial results suggest that oversampling the minority class may be necessary. That being said, we can utilize either our binary MNB classifier to begin the work of filtering comments, or our model with embeddings, depending on if we'd like to minimize our false positives (toxic speech classified as non-toxic, emb model) or our false negatives (normal speech classified as toxic, MNB model)
-
Further Model Enhancements: Continue exploring advanced MNB patameter and hyperparameter tuning.
-
BERT Model: Complete the BERT (Bidirectional Encoder Representations from Transformers) model and assess its precision in detecting toxic comments.
-
Deployment: Consider how to deploy the virtual moderator to actively moderate online platforms and contribute to safer online spaces.