1 of 15

Multilingual Disinformation Detection for Digital Advertising

Maryline CHEN

Machine Learning Engineer

@ Criteo AI Lab

Žofia Trsťanová, Nadir El Manouzi, Maryline Chen, Andre L. V. da Cunha, Sergei Ivanov

2 of 15

Problem statement

Digital advertisement is funding the independent journalism.
3 main actors in this industry: the Advertiser, the Publisher and the internet users.

Advertisers: entities that want to show their ads.
Publishers – Internet Media: websites that accept to display ads.

Publishers are funded mostly via digital advertising (39% Share of media revenue from digital advertising in 2020 [Pew Research]).

Huge increase in disinformation on the internet.
Risk of supporting and financing disinformation publishers.
Negative impact for the advertiser, the user and the ad industry.

Our inventory: ~hundred thousands publisher domains, ~billion pages.
Q: How to remove such Publishers from our inventory?

Publishers

Advertisers

Internet Users

3 of 15

Related work

Work on social media content.

Propagation-based: analyse the way the content circulates. (Zhou et al., 2020a; Zhou & Zafarani, 2018; Wu et al., 2015; Castillo et al., 2011)

Content-based: analyse the textual and image contents. Two types: combinatorial (Pérez-Rosas et al., 2017; Shi & Weninger, 2016; Ciampaglia et al., 2015) or neural (Zhou et al., 2020b; Wang et al., 2018).

Work on news articles.

Using BERT-based and LSTM-based models for sentence representation (Vorakitphan et al. 2022; Blaschke et al. 2020).

Work on online propaganda.

BERT model for multi-class sentence classification (Chang et al. 2021).

4 of 15

Our solution : Repurposed Embeddings for Disinformation Detection (REDD)

Content-based approach using multi-lingual text embeddings as input.
Quantity: Machine learning analyses billions of webpages per week.

Constraints on scalability, latency and efficiency.

Quality: Human expert review in the loop.

Review a shortlist of top-ranked Publisher domains.

Empowers internal teams to proactively blocklist unsafe content.

5 of 15

The REDD pipeline

Leverage the online system.

Currently in production: a fine-tuned XLM-RoBERTa.
Repurpose the multilingual embeddings for input of our model.

From the embeddings, filter the topic to reduce and target relevant pages.

Embedding-based similarity search with a topic specific threshold.

A non-linear classifier predicting probability of being disinformation.

Human review of the top ranked domains.

Publisher text

Text embeddings

Topic projection

Disinformation Classifier

Human review

~0.5B

~12K

~4K

#pages

6 of 15

Dataset

Built from Publishers webpages on one disinformation topic.

Positive label: from 60 domains already blocked for spreading disinformation;
Negative label: from renowned media with a good reputation in fact-checking;

Manually annotated: quality over quantity.
Size after topic filtering: train dataset (17K URLs) vs test set (2K URLs)

53% of positive (disinformation) examples;

Multilingual but with some languages being more represented than others

Language bias: risk that the model might simply become a language detector;

7 of 15

Results: Topic filtering

	AUC-ROC
REDD	0.955
No topic filtering	0.65

8 of 15

Results: text vs embeddings as the input

REDD, the disinformation classifier: a model with three non-linear (SELU) layers.

Trained on 100 epochs, with AdamW optimizer and BCE Loss.
Multilingual embeddings as input.

3 main setups: REDD vs text XLM-R vs text XLM-R frozen embeddings.

AUROC at 1 for the full-text models (XLM-R)
BUT they only learned to distinguish one language from others: skewed dataset.

9 of 15

Results: Human review

Review only the top-ranked Publishers.
Page-level score from the REDD classifier are aggregated at the domain level.
Output from human inspection: 300 domains reviewed.

178 flagged as suspected of spreading disinformation.
26 directly block-listed from the publisher inventory.

more likely to be

ranked higher by REDD

10 of 15

Conclusion

The REDD pipeline: a tool that red-flags domains that potentially spread disinformation.

Topic filtering from text embeddings;
A non-linear classifier predicting probability of being disinformation;
Human review at the end;

Industrial application: A scalable solution that supports different languages and topics of interest.

Potential future work:

Leverage orthogonal sources of information such as images;
Use adversarial training;
Use semi-supervised learning;
Combine with a user-based disinformation graph;

11 of 15

Thank you!

Contact: @ Criteo AI Lab

ma.chen@criteo.com

12 of 15

Q&A

Contact: @ Criteo AI Lab

13 of 15

Appendix

The topic classification schema

14 of 15