At the NeurIPS 2018 workshop on AI for Social Good we presented a piece of work we performed in collaboration with Amnesty International. We leveraged a mixture of crowdsourcing and deep learning to study the nature and quantity of abuse suffered by prominent women on twitter.
At this point quite a lot has been written about the project: if you’re interested in the technical side you can check out our paper and the more extended methodological note; Amnesty have a very nice interactive description of the findings ; and there was an extraordinary amount of news coverage, from Forbes to Wired.
From the NLP team, our major contribution was an abuse classifier than outperformed Google’s API when tested upon this dataset. It was based upon the hugely influential BERT model, and relied upon a superb open-source implementation of BERT from HuggingFace. This is ironic – there’s an overarching ‘anti-big-tech’ narrative to which we were arguably contributing, and yet our tooling relied upon pre-trained models from Google and a framework (Pytorch) from Facebook. This serves to underline the fact that these big companies, Twitter included, produce an enormous variety of impacts – some hugely positive, others overwhelmingly negative. Few companies are unambiguously bad (or good).
One of the most interesting – and unexpected- outcomes of the work was the (albeit temporary) hit that Twitter’s stock suffered. Following the report, an apparently quite influential analyst from Citron Research downrated the stock, causing a 12% plummet – something like $2.5 billion of value lost.
The violence of the market’s response to our work suggests that we hit a nerve. I think that we were probably confirming something that people already knew (Twitter is a toxic place for some women), but hadn’t had the quantitative basis to assert so forcefully.
The bounce-back in the following month might hint that Twitter’s response to the report was effective, reassuring investors’ of imminent action – or simply that markets are flighty and volatile places.