It’s All in the Name: A Character-Based Approach to Infer Religion

Abstract

Large-scale microdata on group identity are critical for studies on identity politics and violence but remain largely unavailable for developing countries. We use personal names to infer religion in South Asia—where religion is a salient social division, and yet, disaggregated data on it are scarce. Existing work predicts religion using a dictionary-based method and, therefore, cannot classify unseen names. We provide character-based machine-learning models that can classify unseen names too with high accuracy. Our models are also much faster and, hence, scalable to large datasets. We explain the classification decisions of one of our models using the layer-wise relevance propagation technique. The character patterns learned by the classifier are rooted in the linguistic origins of names. We apply these to infer the religion of electoral candidates using historical data on Indian elections and observe a trend of declining Muslim representation. Our approach can be used to detect identity groups across the world for whom the underlying names might have different linguistic roots.

Bridging or Breaking: Impact of Intergroup Interactions on Religious Polarization

“…In Appendix B, we discuss how we select the sample of names for manual annotation and choose a more suitable classification threshold. Chaturvedi and Chaturvedi [13] report an 𝐹 1 score of 95% on their test set which reduces to around 85% when only one name part is available (i.e., only the first or last name)-as is often the case on Twitter; we find comparable values (sensitivity 84% and specificity 86.5%). The misclassifications lead to measurement error, potentially underestimating polarization; though qualitatively we expect the temporal patterns to remain the same.…”

“…Since the data on religious identity is not available on Twitter, we infer this from usernames. In India, names are highly predictive of group identity and we leverage character sequence-based machine learning models from our earlier work [13] to obtain religion estimates. 4 Previous research uses these to examine disparities in allocation of publicly provided goods [14], impact of government surveillance on minority voter turnout [2], and potential election irregularities [16].…”

“…Though such an algorithm can be used by nefarious actors, it can also help researchers highlight systematic targeted harassment or, as in our case, monitor trends in polarization across identity groups. See [13] for a more general discussion on ethical concerns.…”

“…Our final data comprises 692,559 unique users with 3,072,503 tweets. 6 The model from [13] is trained on names obtained from a nationally representative sample, while Twitter usernames tend to be noisy. The name classification exercise also depends on the distribution of names in the specific domain the algorithm was trained on.…”

“…We compute the n-day mean of 𝐺𝐶𝑆 𝑖 for each user 𝑖 over the pre-event and post-event windows. 13 Finally, we take Δ𝐺𝐶𝑆 𝑖 = 𝐺𝐶𝑆 𝑖,𝑝𝑜𝑠𝑡 − 𝐺𝐶𝑆 𝑖,𝑝𝑟𝑒 as the difference in the averages post and pre-event. We choose the window size to be large enough to balance the daily fluctuations and small enough to rule out other events influencing the outcome.…”