Repository for "Automatic Detection of Sexist Statements Commonly Used at the Workplace" by Dylan Grosz & Patricia Conde-Cespedes. The associated paper was accepted to the PAKDD Workshop on Learning Data Representation for Clustering (LDRC 2020).
Associated Paper: https://hal.archives-ouvertes.fr/hal-02573576/
Detecting hate speech in the workplace is a unique classification task, as the underlying social context implies a subtler version of conventional hate speech. Applications regarding a state-of-the-art workplace sexism detection model include aids for Human Resources departments, AI chatbots and sentiment analysis. Most existing hate speech detection methods, although robust and accurate, focus on hate speech found on social media, specifically Twitter. The context of social media is much more anonymous than the workplace, therefore it tends to lend itself to more aggressive and “hostile” versions of sexism. Therefore, datasets with large amounts of “hostile” sexism have a slightly easier detection task since “hostile” sexist statements can hinge on a couple words that, regardless of context, tip the model off that a statement is sexist. In this paper we present a dataset of sexist statements that are more likely to be said in the workplace as well as a deep learning model that can achieve state-of-the art results. Previous research has created state-of-the-art models to distinguish “hostile” and “benevolent” sexism based simply on aggregated Twitter data. Our deep learning methods, initialized with GloVe or random word embeddings, use LSTMs with attention mechanisms to outperform those models on a more diverse, filtered dataset that is more targeted towards workplace sexism, leading to an F1 score of 0.88.
After cloning the repository, access the resulting folder and run pip install -r requirements.txt. It is recommended to install all of the packages in requirements.txt. However, some of these packages may be superfluous, as most of this paper's development occurred in PyCharm.
Download a GloVe dataset from https://nlp.stanford.edu/projects/glove/. For both memory considerations and replication of this research, download the 6 billion token, 50D word vectors http://nlp.stanford.edu/data/glove.6B.zip. 50 dimensions were chosen over larger embedding dimensions in order to prevent overfitting/memorization during training on such a small dataset.
A similarly trained gender balanced GloVe dataset (GN-GloVe) can be downloaded here: https://drive.google.com/file/d/1v82WF43w-lE-vpZd0JC1K8WYZQkTy_ii/view
After downloading the dataset, make sure to either name your embedding text file vectors.txt after placing it in the data folder, or edit embedding_fn in SexisteDetectionMain.py (line 33) so it points to the correct filepath.
The dataset we used for training is provided in data/SD_dataset_FINAL.csv. It is also present on Kaggle: https://www.kaggle.com/dgrosz/sexist-workplace-statements. If you would like to update the dataset and reupload, either place it in the data folder and rename it to SD_dataset_FINAL.csv or edit the sexist_dataset_fn variable on line 32 of SexisteDetectionMain.py.
After all the embeddings and sexist workplace statements datasets are properly loaded and referenced, run SexisteDetectionMain.py. There are 8 versions of the model that we have coded, each with their own architecture:
- ModelV1: word embeddings are concatenated per phrase, averaged and put through logistic regression
- ModelV2: word embeddings are sequentially pushed through a Vanilla Unidirectional LSTM architecture, specifically
Embedding->LSTM->Dropout->LSTM->FullyConnected->Softmax - ModelV3: word embeddings (you can toggle between random embeddings and loaded GloVe embeddings in the function call) are sequentially pushed through a Bidirectional LSTM architecture, specifically
Embedding->BiLSTM->Dropout->BiLSTM->FullyConnected->Softmax - ModelV4: embeddings (you can toggle between random embeddings and loaded GloVe embeddings in the function call) are pushed through a Bidirectional LSTM architecture with attention, specifically
Embedding->BiLSTM->Dropout->BiLSTM->Attention->FullyConnected->Softmax
If you don't change the initial code and run SexisteDetectionMain.py, all of these listed models will run sequentially, performing 10 iterations of training and testing (respective training/testing sets are chosen at random for each iteration). You can edit number of iterations on line 335, and you can edit the list of which models to run in the runModels boolean list (lines 342-346).
For any questions, please email Dylan Grosz at dgrosz 'at' stanford 'dot' edu.