Hate Speech Detection for Tweets with Surface-level Features

In this project, we detect whether tweets contain hate speech (offensive words, implications, slang, etc.). The problem can prove challenging due to several reasons such as:

• Tweets contain many out-of-dictionary words (e.g. gr8), slang abbreviations (e.g. smh) and hashtags (e.g. #CorruptPolitician).
• Tweets contain emojis, which strongly relate to the sentiment.
• Tweets are unstructured. They can contain one or more sentences.

We are required to develop a system that uses surface-level features (i.e. not a neural net) to outperform a given baseline classifier.

In our approach, we combine several features used in the literature to form the feature vector and use SVM as a classifier.

Our approach reaches accuracy as high as 72%, blind on a given testset.

File Description

The project contains the following files:

1. Code:
   • main.py: main code. Runs feature extraction and train the model on trainset.
   • feature_extraction.py: extracts features from tweets. Called in main.py.
   • download_data.py: Downloads the necessary data files from our git repo.
   • emoji_split.py: a small script that splits multiple emojis in the same tweet. We used this before feeding the data to TwitterNLP tool.
2. Data
   • nrc_unigrams: unigram sentiment scores from NRC dataset.
   • nrc_bigrams: bigram sentiment scores from NRC dataset.
   • sentiment140_unigrams: unigram sentiment scores from Sentiment140 dataset.
   • sentiment140_bigrams: bigram sentiment scores from Sentiment140 dataset.
   • word_clusters: mapping from words to 1000 clusters provided by CMU TweetNLP tool.
   • train_final_out: trainset tokenized by the CMU TweetNLP tool.
   • train_final_out_pos: trainset pos-tagged by the CMU TweetNLP tool.
   • dev_final_out: devset tokenized by the CMU TweetNLP tool.
   • dev_final_out_pos: devset pos-tagged by the CMU TweetNLP tool.
   • test_final_out: testset tokenized by the CMU TweetNLP tool.
   • test_final_out_pos: testset pos-tagged by the CMU TweetNLP tool.
3. Output
   • predictions.test: predictions of the test set evaluated on the improved model.

How To Run

1. Download the necessary data for feature extraction:

   python download_data.py

2. Run the code to train on train, and give train, dev accuracy and test predictions as test_predictions.tsv:

   python main.py

Libraries Required:

• Python 3
• Scikit Learn
• Pandas
• Numpy