Conspiracy theories, prevalent in contemporary society, often propagate misinformation and distrust, impacting public opinion and decision-making processes. In this paper, we present an automated approach to detect and classify conspiracy theories in Italian to address the Automatic Conspiracy Theory Identification (ACTI) task. Our methodology leverages a transformer-based architecture trained on a multi-task problem to tackle the challenging task of conspiracy theory identification. Through this multi-task learning framework, we aim to build a single model capable of addressing both the detection and the classification tasks simultaneously. We show that tackling both problems in a multi-task setting results in improved performance w.r.t. simple transformer-based solutions.
Identification (ACTI) [
The prevalence and impact of conspiracy theories ne- cross-lingual detection of adjacent topics, such as hate
cessitate efective methods for their detection and classi- speech [10] and fake news [11], or sentiment analysis
ifcation. Manual identification and analysis of conspir- [12, 13], thus ofering useful building blocks for future
acy theories are time-consuming, resource-intensive, and possible applications.
subject to bias. Mainstream platforms (e.g. Reddit, Face- In this paper we propose leveraging Natural Language
book) need to apply moderation policies at the commu- Processing (NLP) techniques to address the ACTI task.
nity level. Given the limitations of the currently adopted We employ a pre-trained transformer-based model
fineapproaches [
the results achieved.
The ACTI task is focused on the detection and
classification of conspiracy theories based on short text posts. The
data is collected from Telegram Channels and is entirely
in Italian, with the exception of some short citations in
English. In particular, two subtasks are proposed:
estimates the probabilities for a post to be conspiratorial
in nature and Subtask B as a model : → [
Finally, we note that there are some overlaps in the posts contained in the two subtasks. While this is not, in general, a problem when the two training sets overlap (although it needs to be kept into account for a multi-task solution, to avoid data leakage during validation), we note that the test set of Subtask A has 164 posts in common with the training set and 66 with the test set of Subtask B, for a total of 230 posts that can be easily assumed to be conspiratorial in nature. In the spirit of a fair competition this information has obviously not been used in any way during the competition.
Let be the set of all possible inputs (i.e. Telegram posts)
and be the set of possible conspiracies. We can
formalize Subtask A as building a model : → [
one another (e.g. QAnon and Covid), so the identification of a single class may at times lead to ambiguous results.
Where (· ) is the sigmoid function. By contrast, the loss function for the conspiracy theory classification is a multi-class classification problem. As such, we aim to minimize the cross-entropy between the predicted probability distribution and the ground truth value (), defined as follows: ℒ()(, ()) = ∑︁ () ( ())
(2)
We note that the datasets available for the two tasks are not the same, although some overlaps occur. Because of this, the two losses ℒ() and ℒ() cannot be computed for all points. In particular, points that are not conspiratorial in nature (() = 0) are not associated with any conspiracy theory, thus making the term ℒ() meaningless. Similarly, points that are conspiratorial in nature but only appear as a part of the dataset for Subtask A are not annotated with a ground truth label regarding the conspiracy theory to which they conform. Because of this, all points that belong exclusively to the dataset for Subtask A can only be evaluated in terms of ℒ().
Instead, all points exclusively belonging to Subtask B are guaranteed to be conspiratorial in nature. Because of this, [CLS] <token 1> <token 2> … [SEP] (") ($) r e d o c n E
[CLS] <token 1> <token 2>
… [SEP] … we infer for those points that () = 1. This consider- unbalanced classes, as is for example the case with Subation makes it reasonable to assume that the multi-task task B. approach proposed should be particularly beneficial in We report some of the main results (i.e. encoders terms of improvements on Subtask A. choice and multi-task vs single tasks comparison) in
The overall loss is obtained as a weighted sum of the terms of performance on the final test set, whereas the above terms: other results on definition of in terms of performance on a validation set that has been obtained as a 20% hold ℒ = ℒ() + 1( ∈ )ℒ() (3) out from the available dataset.
The test set made available for the competition is split Where is used to balance the important that the two intro a public and a private subsets (used for diferent loss terms play in the overall predictions, and 1(· ) is a parts of the competition itself). Both scores have been selector that applies the second loss term only for terms made available, for each submission, at the end of the where that portion can be applied meaningfully. challenge. For Subtask A, the private/public test is a 70/30
Figure 1 summarizes the architecture for the proposed split, whereas for Subtask B the private/public test split methodology from inputs to the computation of the over- is approximately 50/50. all loss. Since we make use of transformer-based en- For conciseness, we report an aggregated result that coders, we note that we adopt the final hidden state cor- covers both private and public scores simultaneously responding to the [CLS] token as representative of the with the aforementioned weights: more specifically, if semantic contents of the entire message [14].
The experimental section aims to evaluate the efectiveness of our proposed approach for detecting and classifying messages containing conspiracy theories. We present the main choices made in terms of encoder adopted, as well as results in terms of choice of hyperparameters.
As already discussed, the models are evaluated on both subtasks by means of the macro 1 score metric. This metric is particularly useful when evaluating models on 1(,) is the score obtained on the public set for subtask A and 1(,) is the score obtained on the private set, we report 1() = 0.3 1(,)+0.7 1(,). Similarly, we report 1() = 0.5 1(,) + 0.7 1(,). We note that, despite assigning the same weights to the same partitions, the reported metrics are not the same as computing the 1 scores on the entire test set – an operation that cannot be performed with the information at hand.
The overall metric used to evaluate the performance in the competition is a weighted average of the performance obtained on the two tasks, with weights 0.6 and 0.4 for Subtasks A and B respectively. Thus, we additionally report the overall score 1 = 0.6 1() + 0.4 1(). 0.85 0.80 e r o c s10.75 F
The main hyperparameters to be configured for the proposed pipeline is the coeficient. Other parameters that are generally important, but not specific to this precise context (e.g., number of training epochs, learning rate, layer sizes, optimizer) will not be covered in detail and can be found as a part of the provided source code.
The choice of a valid value for is quite specific to the conspiracy theory detection problem, as it represents the trade-of coeficient between the capability of detecting conspiratorial contents and being able to correctly assign them. Figure 2 shows how the performance of the model (based on BERT-Italian-XXL-uncased – as explained in the next subsection) varies as increases. Although the best value in terms of overall 1 score is observed for = 0.2, we identify = 1 as being a more robust choice, considering the overall optimal behavior in the interval centered around that value. BERT-Italian-XXL-uncased BERT-Italian-XXL-cased BERT-Italian-base-uncased BERT-Italian-base-cased
BART-IT-WITS BART-IT-IlPost BART-IT-FanPage
BART-IT ELECTRA-XXL-discriminator
ELECTRA-XXL-generator 0.8748 0.8556 0.8605 0.8488 0.8469 0.8452 0.8393 0.8437 0.8635 0.8251
ically tailored to the Italian language. BART-IT has been shown to outperform other state-ofthe-art architectures on various tasks. We note that, on top of the original BART-IT model, three additional versions have been fine-tuned on abstractive summarization tasks on various datasets: FanPage, IlPost [19] and WITS (Wikipedia for Italian Text Summarization) [20]. Since these three data sources have rather diferent scopes and styles, we assess the quality of the various ifne-tuned versions. • Italian BERT [21], a BERT-based model [14] trained on a recent Wikipedia dump as well as data from the OPUS corpora3 collection. We use both a cased and uncased version as well as a base and an XXL ones.
In recent years a wide variety of transformer-based models have been introduced for various languages, including Italian. An a priori choice regarding the most suitable model is not trivial to make. Therefore, we ran a benchmark study to assess how well various models behave on both tasks. In particular, we compare results obtained for the following models: by training the same model on only one of the tasks at a time. As expected, we observe a significant improvement in performance for Subtask A, whereas the performance of Subtask B does not benefit from the introduction of a multi-task approach. This can be explained in terms of benefits that are introduced by the multi-task loss: while points belonging to could all be labelled as being conspiratorial (thus enhancing the dataset available for subtask A), points in are not beneficial to the improvement of Subtask B (as they are either nonconspiratorial, or the conspiracy theory to which they conform is unknown).
The pipeline presented in this work is similar to the one used to take part to the competition, with some minor changes mainly regarding the adoption of BERT-ItalianXXL-uncased instead of BART-IT, as well as not adopting a diferent convolutional model in parallel. For the sake of completeness, the private scores obtained during the challenge are reported in Table 3, along with the best ones obtained with the proposed method.
In this work we presented a transformer-based multi-task approach to addressing a joint detection and classification problem. We have shown the importance of choosing the most suitable encoding model, as well as the benefits of adopting a multi-task approach. We highlighted how the current framing of the multi-task problem is only beneficial to one of the subtasks. As a future work, we aim to reframe the problem through a semi-supervised approach that allows for the pseudo-labelling of all points, thus aiming to improve the performance uniformly across subtasks.
This study was carried out within the FAIR - Future Artificial Intelligence Research and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) – MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 – D.D. 1555 11/10/2022, PE00000013), with partial support from SmartData@PoliTO center on Big Data and Data Science. This manuscript reflects only the authors’ views and opinions, neither the European Union nor the European Commission can be considered responsible for them. [7] S. Shahsavari, P. Holur, T. Wang, T. R. Tangherlini, //www.mdpi.com/1999-5903/15/1/15. doi:10.339 V. Roychowdhury, Conspiracy in the time of corona: 0/fi15010015. automatic detection of emerging covid-19 conspir- [18] M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Moacy theories in social media and the news, Journal hamed, O. Levy, V. Stoyanov, L. Zettlemoyer, Bart: of computational social science 3 (2020) 279–317. Denoising sequence-to-sequence pre-training for [8] T. R. Tangherlini, S. Shahsavari, B. Shahbazi, natural language generation, translation, and comE. Ebrahimzadeh, V. Roychowdhury, An automated prehension, arXiv preprint arXiv:1910.13461 (2019). pipeline for the discovery of conspiracy and con- [19] N. Landro, I. Gallo, R. La Grassa, E. Federici, Two spiracy theory narrative frameworks: Bridgegate, new datasets for italian-language abstractive text pizzagate and storytelling on the web, PloS one 15 summarization, Information 13 (2022) 228. (2020) e0233879. [20] S. Casola, A. Lavelli, Wits: Wikipedia for italian [9] G. Russo, C. Gote, L. Brandenberger, S. Schlosser, text summarization., in: CLiC-it, 2021.
F. Schweitzer, Helping a friend or supporting a [21] S. Schweter, Italian bert and electra models, 2020. cause? disentangling active and passive cospon- URL: https://doi.org/10.5281/zenodo.4263142. sorship in the U.S. congress, in: Proceedings of doi:10.5281/zenodo.4263142. the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Toronto, Canada, 2023, pp. 2952–2969. URL: https: //aclanthology.org/2023.acl-long.166. [10] E. W. Pamungkas, V. Basile, V. Patti, A joint learning approach with knowledge injection for zero-shot cross-lingual hate speech detection, Information
Processing & Management 58 (2021) 102544. [11] J. Köhler, G. K. Shahi, J. M. Struß, M. Wiegand,
M. Siegel, T. Mandl, M. Schütz, Overview of the clef2022 checkthat! lab task 3 on fake news detection,
Working Notes of CLEF (2022). [12] F. Giobergia, L. Cagliero, P. Garza, E. Baralis,
Cross-lingual propagation of sentiment information based on bilingual vector space alignment., in:
EDBT/ICDT Workshops, 2020, pp. 8–10. [13] G. Russo, N. Hollenstein, C. C. Musat, C. Zhang,
Control, generate, augment: A scalable framework for multi-attribute text generation, in: Findings of the Association for Computational Linguistics: EMNLP 2020, Association for Computational Linguistics, Online, 2020, pp. 351–366. URL: https: //aclanthology.org/2020.findings- emnlp.33.
doi:10.18653/v1/2020.findings-emnlp.33. [14] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova,
Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018). [15] R. Guarasci, A. Minutolo, E. Damiano, G. De Pietro,
H. Fujita, M. Esposito, Electra for neural coreference resolution in italian, IEEE Access 9 (2021) 115643–115654. [16] K. Clark, M.-T. Luong, Q. V. Le, C. D. Manning,
Electra: Pre-training text encoders as discriminators rather than generators, arXiv preprint arXiv:2003.10555 (2020). [17] M. La Quatra, L. Cagliero, Bart-it: An eficient sequence-to-sequence model for italian text summarization, Future Internet 15 (2023). URL: https: