[Paper Review] Cross-lingual Transfer Learning for COVID-19 Outbreak Alignment

Sharon Levy, William Yang Wang|arXiv (Cornell University)|Jun 5, 2020

Topic Modeling14 references4 citations

TL;DR

This paper proposes cross-lingual transfer learning using multilingual BERT and tweet features to align social media activity with COVID-19 outbreak trends across countries. By training on Italian Twitter data, the model achieves up to 0.85 Spearman correlation in predicting case counts in other countries, demonstrating that linguistic and semantic patterns in social media can generalize across nations during pandemics.

ABSTRACT

The spread of COVID-19 has become a significant and troubling aspect of society in 2020. With millions of cases reported across countries, new outbreaks have occurred and followed patterns of previously affected areas. Many disease detection models do not incorporate the wealth of social media data that can be utilized for modeling and predicting its spread. In this case, it is useful to ask, can we utilize this knowledge in one country to model the outbreak in another? To answer this, we propose the task of cross-lingual transfer learning for epidemiological alignment. Utilizing both macro and micro text features, we train on Italy's early COVID-19 outbreak through Twitter and transfer to several other countries. Our experiments show strong results with up to 0.85 Spearman correlation in cross-country predictions.

Motivation & Objective

• To investigate whether social media patterns in one country can predict outbreak trends in another through cross-lingual transfer learning.
• To evaluate the effectiveness of multilingual sentence embeddings in aligning epidemiological trends across languages and countries.
• To determine if macro-level (e.g., tweet frequency) and micro-level (e.g., semantic embeddings) features can jointly improve cross-country outbreak prediction.
• To assess the generalization capability of models trained on early outbreak data from Italy to predict trends in countries with later outbreaks.
• To explore the impact of training data timing and sample size on cross-lingual prediction performance.

Proposed method

• The study uses the multilingual BERT (mBERT) model to generate cross-lingual sentence embeddings from tweets in Italian, Thai, Japanese, Turkish, and Indonesian.
• It combines macro-level features (tweet frequency per day) with micro-level features (semantic embeddings) to represent social media activity.
• A linear regression model is trained on Italy’s early outbreak data (February–March 2020) to predict total and daily confirmed cases.
• The trained model is transferred to predict case trends in other countries using only the same model weights, without fine-tuning on target-country data.
• The model is evaluated using Spearman rank correlation between predicted and actual case counts across multiple time settings and countries.
• Five time settings are used: training on February only, February–March, February–April, March only, and February–April (upper bound), to assess generalization under varying data availability.

Experimental results

Research questions

• RQ1 Can social media activity in one country, particularly from Twitter, be used to predict the trajectory of a COVID-19 outbreak in another country through cross-lingual transfer?
• RQ2 How well do multilingual sentence embeddings from mBERT align social media content with epidemiological case counts across different languages and countries?
• RQ3 Does combining macro-level (frequency) and micro-level (semantic) tweet features improve cross-lingual outbreak prediction compared to using either alone?
• RQ4 How does the timing of training data (e.g., pre-peak vs. post-peak) affect the generalization performance of the model to other countries?
• RQ5 Can models trained on Italy’s early outbreak data generalize to countries with different outbreak curves, such as Indonesia, which was still rising in April 2020?

Key findings

• The model achieved a Spearman correlation of up to 0.859 with total case counts when training on Italy’s data and testing on Thailand, indicating strong cross-lingual predictive power.
• For new daily case counts, the highest correlation reached 0.742 in Indonesia when training on February–April data, showing strong alignment with post-peak trends.
• Training on February and March data yielded the highest correlations (up to 0.856 for Japan in total cases), suggesting that linear growth phases are easier to predict than peak or declining phases.
• Indonesia showed stronger alignment with new daily case counts than other countries, likely due to its ongoing increase in cases by April 2020, which the model had not seen during training.
• The upper-bound setting (training on February–April) achieved the highest correlations across all countries and case types, confirming that broader outbreak coverage improves generalization.
• Lower correlations in time settings I and IV suggest that models trained on limited or non-linear data (e.g., only February) struggle to generalize to countries with different outbreak dynamics.