Rhetoric: nextgen platform against online hate speech and polarization
Managing comments and social media dialogue has become extremely difficult, especially for news outlets. The volume they have to deal with is overwhelming. The manual process is painstaking, time consuming and comes with an emotional toll as well. Current textual solutions for automated moderation are not sufficient anymore, because they cannot deal with implicit toxicity or visual memes.
Introduction
Refusing anonymous comments has become a tactic easily circumvented. That is why Flemish news outlets VRT NWS and Het Nieuwsblad were looking for nextgen tools for their news editors. Their goal: to enable more respectful dialogue, counter polarization, and support civil discourse by managing textual and visual comments cleverly and efficiently.
Challenge
Implicit toxicity is by definition invisible to word-based classification engines, so our detection needed to go beyond the identification of swear words and slurs. It should involve a more detailed analysis of rhetorical structure and participant identification.
The key is stance and role detection, and classification of disagreement types. It is not sufficient anymore to identify word-based individual terms. To remain precise and minimize the number of false negatives, we implemented machine learning classifiers.
Toxic meme detection is quite the challenge as well. Memes have more interaction and spread the message faster. They often contain both text and a background image.
Many dashboards enable discussion, but do not empower and support people in their online debate. Neither do they stimulate people to have a civil and thoughtful discussion. Current tools do not include Dutch as target region.
Solution
We started the development of new, state-of-the-art tools to detect polarization and multimodal stance, and extract argument structure. We make use of our own Dutch pipelines in this case. Our new role detection engine can identify a participant’s role. We also develop a classifier measuring the type of disagreement based on the Paul Graham taxonomy.
Memory network models and Convolutional Neural Networks (CNNs) are proven to make textual stance detection happen: we will prove it is feasible for image-based sources as well.
To interpret images, we will map the feature space information from the images with the textual content.
While built for Dutch, our machine learning pipelines will be designed to be as language-independent as possible. It will have built-in features that try to explain and illustrate the result of the AI and why it chose the result it generated ((Explainable AI).
All this comes in tailor-made dashboards for news editors, journalists, moderators and conversation managers, offering metrics and visualization. They will be able to check the current state of stance and role detection, along with the classification of disagreement types in a specific conversation. The dashboards will also contain insights for news managers and moderators to complete.
Result
If this peaked your interest, have a look at the Rhetoric website