BanMANI: A Dataset to Identify Manipulated Social Media News in Bangla

Misinformation is an increasingly pressing concern in the current social and political landscape where information frequently spreads through social media platforms with few constraints to reflect the information in reliable sources. This is further exacerbated by the presence of “bots” made by malicious actors that are designed to artificially spread ideas that distort reality Ferrara (2020); Lei et al. (2023). In order to mitigate this issue, considerable work has been done to identify fake articles Shu et al. (2020), verifying scientific and encyclopedia claims Wadden et al. (2020); Thorne et al. (2018), and identifying claims on social media that distort news from trusted sources Huang et al. (2023). However, most such work is limited to only English.

Bangla, with the fifth-most L1 speakers worldwide, at 233.7 million²²2https://en.wikipedia.org/wiki/List_of_languages_by_total_number_of_speakers, only has prior work in detecting fake articles Hossain et al. (2020). More work in this direction is needed for Bangla on social media platforms, as demonstrated by the 2012 Ramu incident. In the Ramu incident, a Facebook post from a fake account led to the destruction of a Buddhist temple and dozens of houses in Bangladesh by an angry mob of almost 25,000 people Ahmed and Manik (2012). In this vein, we construct a dataset of news-related social media content for identifying news manipulation in social media, BanMANI. This dataset is the comparable Bangla counterpart to the ManiTweet dataset Huang et al. (2023) in English. Figure 1 shows an example of a reference news article alongside both a manipulated and a non-manipulated social media post.

This paper’s contributions are the following.

• •

  We construct a publicly available Bangla dataset of 800 news-related social media items that are annotated as manipulated or not relative to 500 reference news articles.

• •

  We present a semi-automatic method for generating such a dataset, which allows scalable dataset collection using annotators efficiently for languages with few available NLP tools.

• •

  We demonstrate that current SOTA LLMs struggle on this task, both in zero-shot and fine-tuned settings.

This paper is most closely related to fact-checking and fake news detection tasks. While much work in this direction has been done in English, Hossain et al. (2020) only recently started work in this domain in Bangla by releasing a dataset for fake news detection.

In English, Huang et al. (2023) released a dataset for identifying news manipulation in Tweets. In order to supplement fully-human data, they used a semi-automatic approach of generating Tweets using ChatGPT and using human annotators to validate and label the results. They found that ChatGPT and Vicuna failed to solve this new task, even after fine-tuning. In their work, they used FakeNewsNet Shu et al. (2019) dataset to seed their reference articles.

Fact-checking tasks closely resemble our task in that claims must be compared against reference evidence, such as in the FEVER Thorne et al. (2018) and SCIFACT Wadden et al. (2020) datasets. Techniques for this kind of fact-checking work often use a retrieval module that pulls relevant data from the supplied candidate pool. The degree of consistency between a piece of evidence and the input claim is then evaluated using a reasoning component Pradeep et al. (2021).

While our task compares text against a reference article, models must be able to separate social media news related to the reference article from those that only convey opinions to ensure the successful completion of our task. This is the key difference between these (i.e., fact-checking and fake news detection) and our work.

Our goal is to identify whether a news-related social media item (a post or a comment) is manipulated. If the social media item is manipulated then furthermore to determine what particular information is being manipulated relative to a related reliable reference article. We divide this task into three parts.

We confined our test data collection to Facebook since this platform is more commonly used by Bangla speakers compared to typically studied media platforms for English speakers (e.g., Twitter, Instagram, etc.).³³3According to StatCounter.com (https://gs.statcounter.com/), Twitter held 22.01% of thie social media market share in the US in June 2023, but only 1.41% in Bangladesh. On the other hand, Facebook held 48.2% of the market share in the US but 78.84% in Bangladesh. We have created a dataset that contains 800 news-related social media items with 500 associated news articles. Our dataset contains 530 manipulated items and 270 non-manipulated items. The breakdown of our dataset is shown in Table 1.

From Table 4, we see most of the manipulated news is political. Some people spread manipulated news on social media to influence public opinion, promote a particular political party, etc., and these might be the reasons behind the manipulated political news. Also, we notice that national and international news are manipulated in a bigger amount. Sites and pages with low trustworthiness are most likely to spread manipulated news. The followers of those sites and pages are most likely unaware of the fact and accidentally post manipulated news.

The result of the zero-shot ChatGPT and our fine-tuned model is presented in Table 2 and Table 3 respectively. From Table 2 and Table 3, we can see that our fine-tuned model outperforms the zero-shot ChatGPT model for subtask 1, where the F1 score of zero-shot ChatGPT and fine-tuned model is 57.02% and 65.77% respectively. In terms of EM, we can see that our fine-tuned model performs better for both subtask 2 and subtask 3. For subtask 2, if we look at the RL value of our fine-tuned model, we can see that the precision of RL is 69.26%, which is 33.2% more than the zero-shot model. That is also the case for the F1 score of RL. In the same way, for sub-task 3, we can see that the precision and F1 score of RL outperforms the zero-shot model.

Due to our budget limitation, we were not able to collect a large set of human-written social media items. This means that there exists a gap between the quality of the training and test data; the training set was automatically created, unlike the test data. In the future, we will collect more human-written items from social media to create an entirely human-written training dataset. Our prompts are also purposefully simple, as this was the first step in creating such a dataset. We expect to get qualitative gains in the automatically generated data with more careful prompt engineering. Finally, our experiments were limited to only a single popular LLM for each setting. Expanding the experiments to cover other LLMs, especially open-source LLMs would lead to more robust experimental results and better replicability. We also leave the few-shot method as our future work.

In this paper, we presented the BanMANI dataset, the semi-automatically constructed dataset of news manipulation in social media. This dataset extends Huang et al.’s 2023 ManiTweet dataset to Bangla. Our semi-automatic collection process generates social media posts from seed articles using a multilingual LLM and a Bangla NER system. These results are filtered using human annotators for efficient use of annotator time. We find that both zero-shot and fine-tuned LLMs struggle on this dataset, pointing to important directions of future work. Surprisingly, we find that LLMs perform similarly effectively on this dataset when compared to the English variant. We hope that this new resource can help with combating information manipulation in Bangla-speaking social media communities. Furthermore, we believe that the technique laid out here can act as a basis for similar work in other under-served languages in NLP.