The paper proposes an advanced approach for identifying disinformation on Telegram channels related to the Russo-Ukrainian conflict, utilizing state-of-the-art (SOTA) deep learning techniques and transfer learning. Traditional methods of disinformation detection, often relying on manual verification or rulebased systems, are increasingly inadequate in the face of rapidly evolving propaganda tactics and the massive volume of data generated daily. To address these challenges, the proposed system employs deep learning algorithms, including LLM models, which are fine-tuned on a custom dataset encompassing verified disinformation and legitimate content. The paper's findings indicate that this approach significantly outperforms traditional machine learning techniques, offering enhanced contextual understanding and adaptability to emerging disinformation strategies.
The Russo-Ukrainian war has been fought on the ground and in the digital realm, where disinformation and propaganda have become critical tools of influence. Telegram's anonymity and widespread use in military conflict zones have turned it into a go-to platform for spreading disinformation, especially via channels that offer real-time updates from the frontlines. The spread of manipulated narratives and misleading information poses serious threats, including undermining public trust, destabilizing societies, and manipulating international perceptions.
Traditional methods of disinformation detection usually rely on manual verification and rulebased systems and have proven inadequate in addressing the evolving tactics employed by propagandists. The dynamic nature of content, coupled with the sheer volume of data generated daily, calls for more scalable solutions. Recent advancements in Deep Learning offer promising avenues to enhance the detection capabilities in this domain.
This paper introduces an automated detection system for identifying disinformation within suspicious Telegram channels about the Russo-Ukrainian conflict. Leveraging the power of DL algorithms and transfer learning techniques, the system is designed to analyze textual content, identify patterns indicative of misinformation, and adapt to emerging disinformation strategies. Furthermore, the integration of more deep models enables a nuanced understanding of context, semantics, and linguistic subtleties that are often exploited in deceptive narratives.
A critical component of our approach involves processing a custom dataset encompassing verified disinformation instances and legitimate content from Telegram channels. This dataset serves as the training ground for our models, ensuring they capture the multifaceted nature of disinformation prevalent in the Russo-Ukrainian war context. Preliminary results demonstrate the efficacy of our system in identifying disinformation, boasting precision, and recall rates. These findings underscore the potential of integrating SOTA AI technologies to fortify the digital information landscape against malicious incursions.
Detecting propaganda and disinformation in media is an evolving field, especially critical during
periods of conflict. Propaganda is strategically designed to manipulate public opinion by promoting
specific agendas, often through emotionally charged or misleading content. The paper "Fine-Grained
Analysis of Propaganda in News Articles" [
The relevance of this work [
The rise of disinformation surrounding the Russo-Ukrainian war is increasingly fueled by
advancements in generative technologies like Large Language Models (LLMs). A recent study [
Effective disinformation detection heavily relies on the quality and diversity of datasets used for
model training. The research [
There are also qualitative studies, e.g., in paper [
The findings from this paper suggest that frame analysis can be an effective tool for identifying the underlying themes and narratives in Telegram posts, which can then be used to enhance automated detection systems. By focusing on specific frames, such as those promoting regional separation or language persecution, it is possible to develop targeted algorithms that detect these narratives with higher accuracy. Moreover, the study's emphasis on the regional context of Odesa highlights the importance of considering local dynamics when developing detection systems, as disinformation strategies can vary significantly across different regions.
Though progress has been made in detecting disinformation and propaganda on various media platforms. There are still important gaps that need immediate attention, especially the spread of disinformation on Telegram channels. While current research provides valuable insights into identifying propaganda and disinformation, the lack of fully automated systems to recognize and flag harmful content in real-time is a pressing issue. Current methods often require significant manual intervention or are limited to post-hoc analysis, which is inadequate in the fast-paced environment of Telegram, where disinformation can spread rapidly.
Secondly, even current SOTA detection systems cannot evaluate the potential risk level of disinformation. Disinformation can have varying levels of impact, with particular messages posing a more immediate threat to public safety, social stability, or national security. An adequate system should not only identify disinformation but also provide a measurable assessment of its level of danger, allowing for timely interventions for the most harmful content.
These gaps highlight the necessity of developing an automated detection system that not only identifies disinformation in real-time but also assesses its danger level and functions at the granular message level, tailored specifically for the dynamic environment of Telegram channels in the RussoUkrainian war.
The original dataset was provided by the Disinformation Detection Сhallenge Kaggle competition
by AI HOUSE x Mantis Analytics [
Possible suspicious levels:
2. doubtful, potentially needed to be checked 3. surely suspicious disinformation In the manual pipeline, the annotator has a list of tips that helps them make decisions: • From war correspondents and the general situation at the front – 1 • Photos/videos of the attack - 2 (sometimes there are fake videos, but most of them cannot be verified) • Photos/videos of the attack, where posts mention leopards or other Western weapons - 3; there were fakes about leopards already, Russia is actively working to discredit Western equipment • Calls to support money raising – 1 • US statements about the new aid package – 1 • Video of mobilization in Ukraine – 3 • Interviews from Volyna about Azovstal and American soldiers – 3 • •
For our research, we additionally collected all posts from the telegram channels of the main dataset between July and October 2023, a total of 250k records.
The primary metric for the evaluation is F1 with macro average [7]. It is particularly beneficial for multiclass tasks with class imbalance because it provides a balanced evaluation of model performance across all classes, regardless of their individual sizes. Even if a model performs well on the majority class, poor performance on a minority class will significantly lower the F1 macro score, highlighting the model’s weaknesses in those classes. We also use 5-fold cross-validation with shuffle for local validation purposes.
In this section, we will review all of our various experiments conducted for this study. Table 1 and Table 2 present all aggregated final metric values.
Our baseline ML solution is TF-IDF method for vectorization stacked with logistic regression [8]. This technique relies on statistical text representation combined with linear classification to deliver satisfactory results. The key idea is to convert text data into numerical features that a statistical model can process. TF-IDF vectorization accomplishes this by creating a matrix of numerical features, reflecting both the frequency of words in individual documents and their significance within the broader collection of text. After transforming the text into vectorized representation features, logistic regression is then used to classify the data. We used 10k max features in TF-IDF vectorizer and C=1.0 for logistic regression l2 regularization. This method did not work well for this task. It seems that it requires a much deeper contextual understanding of the input text.
Our classical DL solution is xlm-roberta-large (355M parameters) [9, 10] pretrained multilingual model with linear head finetuning on top of it. This method generally performed better than traditional ML because of its superior contextual representation of the text. However, in some experiments, it appeared that the model still struggles to understand the context of events in the ongoing Russo-Ukrainian war, leading to weird results. Additionally, it is almost impossible to achieve good results when fine-tuning the entire model with the backbone, rather than just the classifier head, because of almost immediate overfitting.
The idea of this approach is very similar to the previous one, with the key difference being that we performed the entire pre-training of xlm-roberta-large (355M) on the MLM task like in the original paper [10] ourselves using additional posts from Telegram, as mentioned earlier in the "Additional dataset" chapter. The results were the best among all the experiments that we tried. This can be qualitatively explained by the fact that the domain of warfare is very specific. Existing pre-trained models often have little knowledge in this domain area, so to fully leverage transfer learning, it is necessary to perform pre-training from scratch on the relevant data. The pre-training step here is very resource-intensive. For our additional dataset of 250k samples and 5 epochs, it took around 8 hours on the 4xV100 GPU cluster. For fine-tuning, we still tune only the final classification layer to avoid overfitting. This ensures the model has an excellent contextual understanding of the input text, and has strong generalization ability.
Although proprietary models and models larger than 1B parameters were prohibited in the original Kaggle competition, it makes sense from a research perspective to explore different techniques for working with LLMs. The simplest option is to use SOTA proprietary models with prompt engineering. In this experiment, we used GPT-4-1106 as a starting point. After experimenting with different prompts, we found the optimal prompt to be: "You are an expert in news analysis and disinformation. Your task is to classify each message into one of three possible categories (Suspicious_Level): <list of categories>. During the manual process of annotation, the annotation tool contains a list of tips that help you make decisions: <list of tips>". Where <list of categories> and <list of tips > were used with respect to the Main dataset chapter. The final results were quite contradictory. This can be qualitatively explained by the fact that the original dataset is very imbalanced. With prompt engineering, it is difficult to alter the class distribution in the model's output as generated text, which negatively impacts the macro F1 score.
In this approach, we aim to fully leverage both the power of state-of-the-art LLMs and fine-tuning to achieve the best possible results. Therefore, we use the following configuration: Therefore, we use the following configuration: • Gemma2-9b-it [11, 12] LLM model • LoRA [13] tuning with o_proj, v_proj, q_proj, k_proj target modules; rank=32; alpha=16 • 12 first layers freeze to reduce overfitting • batch_size=4; learning_rate=2e-4; scheduler=cosine This approach worked quite well and was relatively fast (10 minutes of training on 1xV100) because the current pre-trained model is already quite intelligent, and LoRA fine-tuning only trains 0.3% of all the parameters of the original model. However, the results were worse than the option with full pre-training using XLM-RoBERTa, likely because the pre-training data for Gemma2 either did not include Telegram data on the war or was partially censored, leading to worse out-of-the-box performance. Conducting a full pre-training of Gemma2 on Telegram data is not possible for us due to the enormous computational resources required.
Table 1 presents the results of all experiments, with an added row labeled "Naïve Random" which refers to predicting the most frequent class. This allows us to see how much better our ML solutions perform compared to the baseline. The "5-Fold CV" column shows the metric values on the validation splits of the public training dataset, while the "Kaggle Public" and "Kaggle Private" columns display the metric values on the public and private subsets of the Kaggle benchmark, respectively. These columns enable comparison of our solutions with those of other participants in the competition.
It is quite noticeable that the metric results in the three columns differ significantly from each other. This can be explained by the obvious temporal factors in the data, as different data subsets contain posts from different dates, which affects the topics and vocabulary of the posts and introduces a domain shift in the data.
Although our results are only ranked 2nd out of 38 for this benchmark on the private leaderboard, they are stable and consistent across different test data splits, unlike the top private leaderboard entry, which does not appear in the top 5 on the public leaderboard and vice versa. This consistency is beneficial for potential deployment in production, as it suggests that the model's performance on new, previously unseen data is likely to be more stable. We also present the top Public and Private Kaggle leaderboards in Figure 3 and Figure 4, respectively.
Both from a human labeling perspective and from a model performance perspective, it can be challenging to distinguish between different suspicious levels in texts among those that are potentially suspicious. According to the Figure 2, both suspicious classes 2 and 3 could be considered as minority classes, where class 3 consists of rather more suspicious posts than class 2. We can consider a binary version of this problem by simply dividing all posts into suspicious and nonsuspicious, where non-suspicious includes class 1 and suspicious includes 2 and 3. For a binary version of the problem we use ROC-AUC, Precision-Recall AUC (AUC-PR), binary F1 score and 5folds CV similar to the previous section.
In terms of accuracy by metrics in both binary and multiclass cases, models based on fine-tuned transformers work significantly better than classical methods and than third-party APIs of pretrained LLMs. From the hardware optimization side, the current version of the best solution for metrics (XLM-RoBERTa-LARGE with full pre-training) could work in CPU environments approximately 1 request per second on 4xCPU 16 GB RAM machine with 8-bit quantization. If we want to use this on a large scale while being gpu-poor, then the most optimal solution would be to use a binary version TF-IDF + LogisticReg model with a moderate threshold as an input filter and then perform more accurate classification using the XLM-RoBERTa model in order to reduce number of posts classified by the compute-intensive XLM-RoBERTa model, but this approach could lead to evaluation metrics degradation.
We have obtained very promising results from our experiments with employing an automatic detection system for identifying disinformation in Telegram channels discussing the RussoUkrainian conflict. As we performed pre-training on a custom dataset as part of our transfer learning, we found that other approaches such as traditional machine learning techniques and deep learning models were outperformed on different evaluation metrics. Despite these findings, there are still challenges to overcome before this system can be used in real-world scenarios. This stage of pretraining is highly dependent on GPU and requires substantial computational power; thus, scalability in this approach is limited. Moreover, tests conducted on different data subsets showed significant variations that could arise from changes over time in the dataset. Hence, this may undermine the system’s utility in practical situations. These discrepancies must be addressed to adapt to changing conditions and enhance usability across diverse contexts.
First and foremost, I want to thank my students from the Kaggle team Adam’s sons: Andrii Shevtsov, Andrii Zhuravlov and Yevhen Herasimov, for their part in the Kaggle experiments. I am also grateful to Volodymyr Sydorskyi, Oleg Yaroshevskiy, Maxim Tereschenko and the Mantis Analytics team for preparing the dataset and organizing the Kaggle competition. [6] AI House. "Disinformation Detection Challenge." , 2023. URL: .https://aihouse.org.ua/en/event/disinformation-detection-challenge/. [7] Scikit-learn. "F1 Score — Scikit-learn 0.24.2 Documentation.", 2024. URL: https://scikitlearn.org/stable/modules/generated/sklearn.metrics.f1_score.html. [8] Manning, Christopher D., Prabhakar Raghavan, and Hinrich Schütze. Introduction to Information
Retrieval. Cambridge University Press, 2008. [9] Hugging Face. "XLM-RoBERTa Large pre-trained model weights: FacebookAI/xlm-robertalarge.", 2023. URL: https://huggingface.co/FacebookAI/xlm-roberta-large. [10] Conneau, Alexis, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. "Unsupervised Cross-lingual Representation Learning at Scale." In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 8440-8451. Association for Computational Linguistics. [11] Hugging Face. "Gemma 2-9b-it pre-trained model weights: Google/gemma-2-9b-it.", 2023. URL: https://huggingface.co/google/gemma-2-9b-it. [12] Team Gemma, Morgane Riviere, Shreya Pathak, Pier Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Lan, Sammy Jerome, and Alek Andreev. "Gemma 2: Improving Open Language Models at a Practical Size," 2024. [13] Hu, Edward, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. "LoRA: Low-Rank Adaptation of Large Language Models," 2021.