A Multilingual Collection of Facebook Comments on the Moro Identity and Armed Conflict in the Southern Philippines

Context

In the information age, social media posts and comments on current events provide a vast resource for insights into public opinion and online political discourse (Acerbi, 2020). The Philippines, which consistently ranks amongst the countries with the highest degrees of social media use (Baclig, 2022) with more than 87 million Facebook users as of 2023 (NapoleonCat, 2023), thus presents an apt target for data collection; however, there are few publicly accessible datasets that centre political comments on social media.

This dataset attempts to capture a wide spectrum of political comments on the transition between the former political entity of the Autonomous Region in Muslim Mindanao (ARMM) and the Bangsamoro Autonomous Region in Muslim Mindanao (BARMM), representing a critical turning point in the aspirations of Moros (the collective term for various Muslim peoples in the southern Philippines) for greater political autonomy. Three key events in the last decade are the focus of this data collection: 1) the Mamasapano Clash (2015); 2) the Marawi Siege (2017) representing widely reported encounters between the Philippine police or military and armed groups in ARMM; and 3) the establishment of BARMM (2019). This data collection further contributes to online resources for studying online views on Islam and Muslims in Southeast Asia, which is often overshadowed by the amount of research on the topic in Western settings (Hashmi, Radzuwan, & Ahmad, 2021).

Steps

The comments in this dataset were gathered through Facepager (Jünger and Keyling, 2019), an open-source third-party software, and came from the official Facebook pages of widely-distributed national newspapers: The Philippine Daily Inquirer, Manila Bulletin, The Philippine Star, and The Manila Times (see Chua, 2023), and regional newspapers such as Sunstar Cebu, Cebu Daily News, The Freeman, Sunstar Davao, MindaNews and The Mindanao Times (Table 1).

Sampling strategy

The dataset is part of a larger corpus of more than 7.6 million words containing Facebook posts and comments of the newspapers above. Data was collected from the month in which the target event occurred and its succeeding month, comprising the main corpus, and posts and comments from two random months in the same year, forming the side corpus, amongst other variables (Table 2).

A list of target words for filtering the corpus was developed by expanding a core set of terms² through fastText (Bojanowski et al., 2016), a library for the learning of word embeddings. The resulting list of 184 related terms in Filipino, English, and Cebuano was used to extract relevant comments from the corpus to form the present dataset. Due to the presence of multilingualism and codeswitching in the comments, a multilabel classifier trained on a set of 16,156 manually-encoded comments was used for language detection (see Cruz & Kestemont, 2022).

The classifier revealed three codeswitched varieties in the dataset: Taglish (Filipino and English), Bislish (Cebuano and English) and Bislog (Cebuano and Filipino), alongside the dominant languages Filipino, English, and Cebuano, with most of the comments written in Filipino or English (Figure 1), although language use varied significantly based on newspaper (Figure 2). The distribution of comments in the dataset shows more comments from months associated with the target events (Figure 3), the Luzon region (Figure 4), and the Aquino administration (Figure 5).

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Quality control

The original and pre-processed messages do not include any markup for tags or hyperlinks, and all duplicates (based on message content and creation time) were removed. While only public pages are identified in the column “from_name”, all comments from public pages save for those made by the newspapers are anonymized as ‘NAME’. Additionally, some level of error in language detection is present, as language labels were computationally derived and based on the pre-processed version of the original message, with a detection accuracy of 92.44% (Cebuano), 92.77% (Filipino/Tagalog) and 96.27% (English) as described in Cruz & Kestemont (2022). These discrepancies can be checked against the original message for comparison.

Repository name

Zenodo

Object name

MMB151719SOCMED

Format names and versions

CSV

Creation dates

2021-03-17–2023-06-03

Dataset creators

Frances Cruz (Conceptualization, Curation, Methodology)

Language

Tagalog (Filipino), Cebuano, English, Taglish, Bislog, Bislish

License

CC BY-NC-SA 4.0 DEED

Publication date

2024-06-05