Team IUCL used a fine-tuned DeBERTa [ 1] with a sequence classification head as its basis. DeBERTa was ifnetuned on an augmented dataset comprising the PAN24 [ 2] training set and the LOCO corpus [3]. LOCO is a corpus consisting of conspiracy and mainstream texts for a range of conspiracy topics, including COVID related narratives. By adding all LOCO texts, the training set increased by five times to around 20,000 texts. The team utilized a balanced subset of 400 samples from the original dataset as the development set. Data augmentation led to a significant improvement in model convergence, resulting in DeBERTa's performance equaling that of a Llama-2-7B model on the original dataset. The augmented DeBERTA model was evaluated on the PAN24 test set and achieved an MCC score of 0.8388, achieving the best result out of 83 teams.
The surge of conspiracy theories spread over social media and elsewhere has resulted in a renewed
interest among scholars to understand how conspiracy theories spread [
Ethics professor fired for having ethics . ‚Äú My school implores me to be an authority on ethics and I ’m here to tell you it ’s ethically wrong to coerce someone to take a vaccine . ‚Äù Fight back . AFLDS . org / legal
In terms of computational approaches to conspiracy theory detection, Peskine et al. [
3.1. Data Since we are participating in subtask 1, the binary classification, and on the English data, we describe only that dataset here. The datasets provided by the shared task organizers were collected from the Telegram platform, specifically addressing discussions surrounding the COVID-19 pandemic. The training set comprises approximately 4 000 texts, each annotated as either CRITICAL or CONSPIRACY. The training set is imbalanced, with 66% of the texts being critical and 33% conspiracy texts.
To assess the performance of our models, we partitioned the original training set and created a development set, randomly choosing 200 texts per category. This balanced distribution ensured a robust evaluation across both the categories. Furthermore, a test set consisting of 1 000 texts was made available by the shared task organizers.
To provide more data to our language model (LM) approach, we augmented the data and used the all
the texts (mainstream and conspiracy) from the LOCO [
The reason for choosing conspiracy texts across all conspiracy theories present in LOCO was to develop a more generalized LM based system capable of capturing a wide range of conspiracy texts, rather than focusing specifically on COVID-related texts.
After a brief skimming of the training data, we found that the URLs were tokenized, with spaces separating the URL path within the domain. Some of the URLs were shortened, and certain texts contained a combination of both standard URLs and shortened URLs.
For example, consider the text in Figure 1. This text shows a split URL and incorrectly rendered quotes. Cleaning such a dataset proved to be a challenging task in itself. We used regular expressions to cover as many of the URLs as possible, but there are some that our method did not identify.
For the statistical models we initiated data preprocessing by converting all text to lowercase. Subsequently, we replaced URLs with the token ’URL’ to mitigate their influence and enhance the dataset quality, using regular expressions. A frequency investigation of the words in the classes shows a higher prevalence of URLs in the conspiracy texts. However, upon examination, we found that both categories contained URLs, as depicted in Figure 2.
Utilizing the ’en_core_web_sm’ pipeline from spaCy2, we then removed stopwords to evaluate their impact on traditional models. Additionally, we extracted metadata such as sentence lengths and word counts. We performed part of speech tagging using the POS tagger provided in spaCy. Each word was then replaced with its corresponding POS tag.
We also investigated content word bleaching, in the assumption that words that are very specific for
a given conspiracy theory will be detrimental to detecting other conspiracy theories. Following Fort et
al. [
These preprocessing steps resulted in the creation of diferent types of texts, namely: 1. text: raw text 2. POS tagged: words replaced by their corresponding POS tags 3. preprocessed: lower cased text with URLs replaced 4. no stopwords: stop words removed 5. bleached: top 25 words replaced in the preprocessed text.
Due to time limitations and the size of the LOCO corpus, we only performed the experiments using preprocessed data on the original training set for statistical models. 2https://spacy.io 3https://radimrehurek.com/gensim/
We used the enriched texts and word embeddings to train Multinomial Naïve Bayes, Random Forest, XGBoost, Logistic Regression, and SVC models.
We chose the Naïve Bayes classifier for its simplicity and ability to handle missing data values. The Support Vector Classifier (SVC) excels at handling high-dimensional spaces and is robust against overfitting. Random Forest, an ensemble learning method, is also robust to overfitting and provides feature importance ranking, which helps identify the most influential features. Logistic Regression and Multinomial Naïve Bayes classifiers are both easy to interpret and computationally eficient. XGBoost ofers a highly eficient and scalable approach to handle data. All models, except for XGBoost, were implemented using scikit-learn. For XGBoost, we used the Python XGBoost module to train our systems. We optimized the classifiers using grid search and 5-fold cross-validation.
We also trained an ensemble of the top 15 models that performed best on our development set. This process resulted in the creation of seven diferent ensembles, trained on the top 3, 5, 7, 9, 11, 13, and 15 models. Among these, the ensemble of the top 11 models achieved the highest MCC score.
To create the best performing ensemble, we used the experiments reported in section 5.2. We show the models selected for the best performing ensemble of 11 models in Table 1.
We employed traditional features, using TF-IDF with the maximum feature parameter varied at 1000, 5000, and 10000, and bag of words with -gram ranges from 1 to 5, as well as modern embedding-based approaches to create features for the statistical models.
Initially, we employed BERT-based embeddings to train our statistical models. However, BERT
models can handle a maximum sequence length of 512 tokens, which was insuficient for our needs
[
While training the statistical models, we found that using the POS-tagged text led to significant information loss. Therefore, we discontinued the POS based models early on. The experiments included the four remaining text representations, three feature extraction techniques, and six types of models.
We used the transformers library by Hugging Face [14] to finetune our language models. The language
models were loaded in 4 bit precision to reduce memory usage and speed up training. We used the
Parameter Eficient Finetuning (PEFT) [ 15] library to load the models. Our experiments focused primarily
on the DeBERTa [
Both language models were trained on the first 512 tokens per text. DeBERTa also predicted on the first 512 tokens while Llama-2-7b performed best when predicting on the first 2048 tokens. Both language models were first trained only on the PAN24 corpus, followed, where applicable, by training on an augmented training set that included the LOCO corpus.
Although the LMs can handle longer sequences, computational constraints limited the sequence length to 512 tokens for training. However, we conducted one experiment using 2048 tokens for training, which did not yield better performance compared to 512 tokens. Even with this limitation of training on 512 tokens, we performed inference on 2048 tokens for the models based on Llama-2-7b and Llama-3-8b architectures to capture a much larger context. This led to a significant increase in the performance of the LMs.
The hyperparameters used for the submitted system using DeBERTa were as follows: learning rate of 5e-5, token length of 512, gradient accumulation steps of 8, and a batch size of 1. The Low-Rank Adaptation parameters used for the PEFT library were as follows: Rank (r) of 64, alpha of 16, a dropout of 0.1 and a task type of Sequence Classification. All linear layers were updated during training. No bias was used.
The organizers used the Matthews Correlation Coeficient (MCC) [ 17] as the evaluation metric to assess the systems. We adopted the same metric to perform internal evaluation of our statistical models, language models, and ensembles.
Here, we first provide an overview of the oficial results of our two systems then we will discuss results of our internal experiments with the statistical models, the LMs, and the ensembles, on our development data.
We were allowed to submit two systems. For the first system, we chose a DeBERTa model finetuned on the entire PAN and LOCO datasets. Fr the second system, we chose an ensemble that includes the augmented DeBERTa system plus Llama2 finetuned on the PAN dataset, plus the statistical models of the best ensemble of traditional models. The latter was chosen since it performed best in our internal evaluation (see next section).
SVC preprocessed no stopw. bleached text preprocessed no stopw. bleached
Features BOW TF-IDF embeddings BOW TF-IDF embeddings BOW TF-IDF embeddings BOW TF-IDF embeddings BOW TF-IDF embeddings BOW TF-IDF embeddings BOW TF-IDF embeddings BOW TF-IDF embeddings
The oficial results in Table 2 show that our system using DeBERTa with the augmented training set reached an MCC of 0.8388, and was ranked the best system in the competition, out of 83 teams. Our second submitted system reached an MCC of 0.7845. It was not oficially ranked but places between ranks 27 and 28 on the oficial ranking 4, thus not reaching the baseline BERT system.
The results for the two best performing models among the statistical models, SVC and Logistic Regression, are shown in Table 3. The best result is reached by using the SVC with TF-IDF and bleached topic words. The second highest result is reached by Logistic Regression when using a standard bag of words after bleaching topic words.
Overall, we find that the best performing settings depend on the model choice. For Logistic Regression, the bag of words outperforms TF-IDF features as well as word embeddings, and stopword removal works well. For SVC, TF-IDF features are best, followed by word embeddings. Here, bleaching topics improves results, while the other preprocessing strategies have a mixed influence, depending on the feature representation.
Our preliminary results showed that the removal of URLs has an overall negative impact on statistical classifiers. There are exceptions (e.g., when using the SVC with embeddings or using Logistic Regression with TF-IDF features), but there are no consistent trends.
The MCC results mostly depend on the model’s success on the conspiracy class. This makes sense as the conspiracy is the minority class, and thus harder to predict. However, the F-scores for this class 4https://pan.webis.de/clef24/pan24-web/oppositional-thinking-analysis.html tend to be around 2-3 points lower that the corresponding F-scores for the critical class. Thus, the data imbalance does not hamper the systems much.
The results of our experiments using LMs, evaluated on the development data, are shown in Table 4. They show that the best results are reached by using Llama 2. For this model, augmenting the training data results in a minor improvement from an MCC of 0.8601 to 0.8607. For DeBERTa, in contrast, augmenting the training data results in a sizable increase from an MCC of 0.7951 to 0.8358. Additionally, we found that augmenting the language models with the LOCO data accelerates convergence and improves results.
The increase for DeBERTa when using the augmented training data is surprising in that the LOCO data used for augmenting the training data difer from the training data in the text type (websites vs. Telegram), they cover a wider range of conspiracy theories, and the texts were sampled to highlight the distinction between conspiracy and mainstream content, rather than conspiracy and critical thinking. The fact that DeBERTa can successfully use the additional training data leads us to the assumption that there is a language of conspiracy theories, with a distinct style that this LM can learn to recognize.
The results of the ensembles, either combining LMs with statistical models or statistical models only, are shown in Table 5. These results show that the best ensemble of 11 classifiers gains about 2.6 points on the best performing single classifier. However, this combination does not reach the ensemble including LMs. Our second submission system, IUCL2, reaches an MCC of 0.90 on our development set. Note that this is about 4 points higher than the results of the augmented Llama 2 and about 6.5 points higher than the augmented DeBERTa model. This trend is in contrast to the results on the test data, where the single DeBERTa model performed significantly better.
Conspiracy:
We conducted a post-result analysis to interpret the model performance of our winning model and check the saliency of individual words. We utilize the Integrated Gradients method [18] to calculate the importance of each word in the text. We then visualize the importance of each word in the text using the Transformers-Interpret and Captum libraries.
Our analysis shows that the model pays attention to words in the text that are often interpretable. We show two examples from the development set visualizing the saliency in the text in Figure 3. The critical example shows that the tokens “New” and ”on_Private” are highly indicative of the critical class while ‘No” would point to conspiracy. In the conspiracy text, the term “Illuminati” is the best indicator of the conspiracy class.
Figure 4 shows an example from the test set where our system classified a text as critical. This is an example where there are no solid indicators as to whether a conspiracy belief is present, and the only token receiving attention is the period, showing that this is a dificult example to classify.
We conducted an extensive analysis of various statistical and language models, evaluating their individual performance, and the efectiveness of their ensembles, to classify texts into conspiracy and critical texts. We found that statistical models tend to underperform in comparison to LMs: the best statistical model is about 10 points below the best LM, and ensembling them increases results by about 2 points. However, this ensemble is only 3 points below the non-augmented DeBERTa.
On the oficial test set, our best performing model is a DeBERTa model with augmented training data. Our findings also indicate that ensembles combining traditional models with LMs outperformed individual LM results on the development set but not on the test set. This disparity suggests that ifnetuning sensitivity to data characteristics may have influenced performance, as gains observed in the development set did not fully generalize to the test set.
For future work, we plan to further explore and interpret the model results to gain insights into how the language of conspiracy theories difers from that of critical texts. Another avenue of research involves investigating the uniformity of labeling and annotations, as well as the impact of this on the classification results.
This work is based on research in part supported by US National Science Foundation (NSF) Grant #2123618.
This research was supported in part by Lilly Endowment, Inc., through its support for the Indiana University Pervasive Technology Institute. [13] I. Beltagy, M. E. Peters, A. Cohan, Longformer: The long-document transformer, arXiv:2004.05150, 2020. arXiv:2004.05150. [14] T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, A. Rush, Transformers: State-of-the-art natural language processing, in: Q. Liu, D. Schlangen (Eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for Computational Linguistics, Online, 2020, pp. 38–45. URL: https://aclanthology.org/2020.emnlp-demos.6. doi:10.18653/v1/2020.emnlp-demos.6. [15] S. Mangrulkar, S. Gugger, L. Debut, Y. Belkada, S. Paul, B. Bossan, Peft: State-of-the-art parametereficient fine-tuning methods, https://github.com/huggingface/peft, 2022. [16] H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, D. Bikel, L. Blecher, C. C. Ferrer, M. Chen, G. Cucurull, D. Esiobu, J. Fernandes, J. Fu, W. Fu, B. Fuller, C. Gao, V. Goswami, N. Goyal, A. Hartshorn, S. Hosseini, R. Hou, H. Inan, M. Kardas, V. Kerkez, M. Khabsa, I. Kloumann, A. Korenev, P. S. Koura, M.-A. Lachaux, T. Lavril, J. Lee, D. Liskovich, Y. Lu, Y. Mao, X. Martinet, T. Mihaylov, P. Mishra, I. Molybog, Y. Nie, A. Poulton, J. Reizenstein, R. Rungta, K. Saladi, A. Schelten, R. Silva, E. M. Smith, R. Subramanian, X. E. Tan, B. Tang, R. Taylor, A. Williams, J. X. Kuan, P. Xu, Z. Yan, I. Zarov, Y. Zhang, A. Fan, M. Kambadur, S. Narang, A. Rodriguez, R. Stojnic, S. Edunov, T. Scialom, Llama 2: Open foundation and fine-tuned chat models, 2023. arXiv:2307.09288. [17] D. Chicco, N. Tötsch, G. Jurman, The Matthews correlation coeficient (MCC) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation, BioData Mining 14 (2012). [18] M. Sundararajan, A. Taly, Q. Yan, Axiomatic attribution for deep networks, in: D. Precup, Y. W.
Teh (Eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, PMLR, 2017, pp. 3319–3328. URL: https://proceedings. mlr.press/v70/sundararajan17a.html.