Against sexism like a machine
“ Most women no longer have the desire or the knowledge to develop a high quality character, even if they wanted” (🧛♂️ misogynist tweet example )
Despite regulatory structures in place across the EU to prohibit discrimination and promote equality, women continue to be under-represented in decision-making roles and excluded from key economic sectors. Daily, sexism is also present right on the surface of the Internet, and it is an adequate place to shut down misogyny before it spreads.
At Recognai we are a culturally and sexually diverse team and we embrace equality and inclusion. In that way we are sensible and compromise into gender equality (and we still need to improve: 2 women vs 3 men are working full time here, and both founders on 3 are males). In any case, sexism, such as racism, has no place here.
Ignacio, our amazing Intern, run his Bachelor’s thesis on Misogyny Detection on Spanish texts using Deep Learning Models. The objective was to create models capable of detecting and categorizing sexist content, and several approaches were researched, using data from previous shared task that the NLP community carried out, pushing the state of the art forward.
So we embarked on this project about creating and optimizing a social media misogyny detector. Sexism in social media mostly affects women and it is an issue that has been addressed countlessly by many organizations and influential personas. There are moderation teams in almost all digital platforms which do their best to answer report and keep those environments as free from misogynistic content as possible. But, what if there was an automatic system to identify signs of sexism online and help fight it? The EXIST task at IberLEF 2021 was all about this and more. We are talking about a shared task that rewards the best sexism identifiers and classifiers for social media posts and interactions.
The technical name for this is Automatic Misogyny Identification (AMI) and it learns to distinguish misogynist contents from non-misogynous ones and to categorize their types.
The first task was based on binary classification, where the system had to decide whether the tweet it analyzed was sexist or not.
The second, more advanced task was about categorizing sexism into sub-categories such as stereotyping, dominance, or sexual violence. The model was also trained to spot certain misogyny types that were active or passive (addressed to individuals or the general public). The second task further allowed us to understand which types of sexism were more widely spread across social media.
The communion between biome.text and Rubrix, our two open-source tools, was fundamental for carrying out this task. NLP models were trained on biome.text, using Transformer models from HuggingFace Transformer, and both the data and the obtained predictions were explored and validated using Rubrix.
Later on Ignacio’s thesis, we even used Rubrix to include the data corpus of this shared task into a bigger corpus, to create a multipurpose misogyny detector. In that scenario, we needed to adapt the instances from EXIST into a different categorization. The whole team dedicated some hours to annotate tweets with the No-code manual labeling application.
The model demonstrated that AI can work together with humans to aid the fight for social justice and equality as it deals with crucial, time-bearing processes. Filtering through the dirt manually would become an endless task of identifying and tagging sexist content as thousands of tweets in this fashion flood the platform every day.
How we can guide AI to a future without discrimination
At Recognai we believe that discrimination and prejudice are human traits that affect everyone during their lifetime. Machines and AI can help monitor these occurrences to lower them to a minimum and help prevent extreme cases such as hate crimes or domestic abuse. In the long run, projects like this will add up to make the difference we want to see in the world and make it a more inclusive place regardless of gender or beliefs. We are very proud to take part in events of this type and scale especially considering their purpose and social benefit. At the end of the day AI is a direct reflection of what we teach it to be so implementing it in positive social projects like this one could go a long way in the future.
If you have an idea to make a positive change with NLP, contact us, we can join forces and support you with access to Rubrix our Open-source tool for tracking and iterating on data for AI projects.
Interested in natural language processing, open science, and open-source?
⭐ Star Rubrix and biome-text projects to stay tuned about our open-source projects.
🐦 Follow us on Twitter
📇 Follow us on LinkedIn
