OPLOT

🧌 Internet trolls block civil debate, incite conflicts, spread misinformation and hatred.

💡 To cope with their invasion, we created the antitroll bot.

🔧 The bot is based on a troll recognition algorithm trained on 10,000+ comments from anti-war Telegram channels and the Botnadzor project. The developers marked those comments that were deleted by moderators, tried a dozen text processing models on them, chose the best one and added a neural network. Our bot demonstrates 74% accuracy on a test set.

🤖 To install the bot, email our team at [email protected]. We'll provide detailed setup instructions and walk you through step by step.

The current version of the Bot is available as an API, and we plan to add Telegram bot support in autumn 2024. To access the web interface, contact us ([email protected]).

Our goal is to provide tools to extract topics, tags and narratives from the news and social media texts.

We have launched:

• a Telegram bot (in test mode) that analyses news from different angles;
• a news digest;
• an interactive analytical dashboard for narrative detection;
• a demo dashboard of what information we can extract for your project: topics, interpreted with LLMs, named entities, narratives.

Analysis

• Themes - recurring & trending topics and ideas.
• Clusters - groups of similar articles.
• Narratives Identification - match to predefined propaganda frames.
• Named entities - persons, organizations, places mentioned.
• Sentiments associated with particular people and organisations. (coming soon)

All to be compared across different media outlets.

Target audience

• Researchers and journalists writing about Russian society and advocacy
• Teachers and students of media literacy and journalism courses
• Activists working in the field of counter-propaganda
• News readers interested in how different events are covered in the Russian media

Our goal is to provide tools to extract topics, tags and narratives from the news and social media texts.

Example:

Text

System assessment

“Many military and political experts assess the current situation in the world as extremely dangerous and acute, primarily because of the threat of the use of nuclear weapons. In recent decades, there has never been such a high risk of a war with the use of such weapons as there is now. All because of the aggressive actions of the US.” Source: RIA (translated)

Emotional coloring: The use of strong emotional words and expressions, such as "extremely dangerous and acute", "high risk of unleashing war", is aimed at evoking fear and anxiety in the reader. Use of generalizations and exaggerations: Phrases like "all because of aggressive actions by the US" can lead to a simplistic perception of the situation, ignoring the complexity of international relations and interactions.

You can use the system as an off-the-shelf solution or make it part of your product/workflow.

What can you use the service for?

• Automated rating of unbiased media/news Telegram feeds, scalable to a large number of texts.
• An extension of a news aggregator: the system can act as a safety net for your readers, allowing them to stay informed about the quality of information presentation even in passive consumption mode, and alerting them if they are about to share a manipulative text with friends or family.
• A research resource for journalists, academics and activists. Track in real time or retrospect the level of propaganda and types of manipulation in any text source in Russian. Analyse exactly how different sources manipulate their readers and compare them against each other.
• A component of educational initiatives, such as media literacy or critical thinking courses — also for high school and university students.
• Self-control — checking your own publications to ensure they are free of manipulation, similar to what Grammarly or Glavred services do for grammar correction.
• An easy illustration for relatives and friends to warn them about the extent of manipulation in their information space.

How can you access the service?

A continuously updated and automatically analysed news feed from Telegram channels is available here. API to access our predictive models is available on request, due to security reasons.

"Propaganda Thermometer", a live overview of known propagandistic Telegram channels, is available here.

A Telegram bot that can analyse a text you send it is available here.

We are now preparing our own Telegram bot with a personalised aggregated news digest, complete with a manipulation level indicator.

How does it work under the hood?

The core of our approach is natural language processing, or NLP. We use ChatGPT for advanced analytics, to provide detailed explanation. Our toolkit also includes faster and more cost and resource-efficient models that we have trained ourselves by fine-tuning the BERT (Transformer) neural network. Those models can determine with high accuracy (~ 85%) whether a sentence in Russian language contains manipulations or not.

To train the models, we used open data from scientific conferences (SemEval 2023), as well as data that our team collected and annotated ourselves. Our corpus contains 3000+ sentences labelled with 6 classes of manipulation (logical fallacies and lies, enemy demonisation, emotionally charged language, appeal to traditions & history, justification of violence, black-and-white thinking). The annotator agreement is comparable to scientific publications in this field.

To obtain the data, please fill in this Google form.

Get in touch!

If you would like to try out the model on your own text collection, or are interested in working directly with our corpus, please send us an e-mail. Our current partners include NeNorma, TrueStory (ex-Yandex) and AskRobot. We look forward to working with more projects!

We are always looking for NLP data scientists and engineers interested in improving the model (e.g., extending it to multi-class). Our next step is integrate AI into media literacy projects for school and university students in a gamified, interactive format.