Abstract
Progress on the UN Sustainable Development Goals (SDGs) is hampered by a persistent lack of data regarding key social, environmental, and economic indicators, particularly in developing countries. For example, data on poverty — the first of seventeen SDGs — is both spatially sparse and infrequently collected in Sub-Saharan Africa due to the high cost of surveys. Here we propose a novel method for estimating socioeconomic indicators using open-source, geolocated textual information from Wikipedia articles. We demonstrate that modern NLP techniques can be used to predict community-level asset wealth and education outcomes using nearby geolocated Wikipedia articles. When paired with nightlights satellite imagery, our method outperforms all previously published benchmarks for this prediction task, indicating the potential of Wikipedia to inform both research in the social sciences and future policy decisions.
Introduction: Addressing the Data Gap
Tracking progress toward the UN Sustainable Development Goals (SDGs) is severely hampered by a significant "data gap." In many developing nations, especially in Sub-Saharan Africa, socioeconomic data is gathered through door-to-door surveys that are expensive, infrequent (often every five years), and spatially sparse.
While previous research has used satellite imagery, such as "nightlights," to estimate wealth, these methods face a "floor" effect: in the poorest regions, the lack of artificial light makes it impossible to distinguish between different levels of poverty. This research explores whether the qualitative information in Wikipedia—articles about local schools, hospitals, infrastructure, and history—can serve as a high-resolution proxy for economic development.
Methodology
The researchers developed a multi-modal architecture to fuse textual data with satellite imagery:
- Textual Representation (Doc2Vec): Over 1 million geolocated Wikipedia articles were processed. Using the Doc2Vec method, the unstructured text of each article was converted into a 300-dimensional vector capturing its semantic meaning.
- Wikipedia Embedding (WE) Model: For any target coordinate, the model identifies the ten nearest geolocated Wikipedia articles. It concatenates their embeddings and physical distances, feeding them into a 3-layer Multi-Layer Perceptron (MLP) to regress wealth or education scores.
- Multi-Modal (MM) Model: This approach combines the WE model with a Convolutional Neural Network (CNN) processing 5km x 5km patches of VIIRS night-time satellite imagery. Features from both text and image branches are concatenated to produce a final prediction.
Datasets
The study utilized several key data sources:
- Socioeconomic Ground Truth: Asset Wealth Index (AWI) data from Demographic and Health Surveys (DHS) and education indices from Afrobarometer.
- Wikipedia Corpus: A June 2018 dump containing approximately 1.02 million geolocated articles globally, with about 50,000 located in Africa.
- Satellite Imagery: Night-time light intensity data from the VIIRS sensor.
Key Insights
- Text as a Proxy for Infrastructure: Wikipedia articles act as data-rich proxies. For instance, an article about a new light rail system or a specific hospital provides qualitative evidence of development that raw pixel values might miss.
- Complementary Strengths: While nightlights identify urban-rural divides, Wikipedia provides the "fine print," helping the model distinguish between varying levels of poverty in areas that appear dark on satellite maps.
- Semantic Clustering: Using Principal Component Analysis (PCA), the researchers found that the model naturally associates wealthier areas with categories like "university" or "company," while poorer areas cluster around general "settlement" articles.
Results and Performance
The models were evaluated across five countries: Malawi, Nigeria, Tanzania, Uganda, and Ghana.
- Poverty Prediction: The Multi-Modal (MM) model achieved Pearson’s $r^2$ values between 0.64 and 0.76, a significant improvement over nightlight-only models (averaging 0.45–0.55).
- State-of-the-Art Comparison: The MM model outperformed previous benchmarks by an average of 31.2%, even those using expensive high-resolution daytime imagery.
- Education Prediction: The model successfully predicted non-economic social indicators, achieving $r^2$ values between 0.30 and 0.39 for education levels.
- Generalization: A model trained on data from one country (e.g., Ghana) could accurately rank wealth in another (e.g., Tanzania), suggesting it learns universal features of development.
Limitations and Future Directions
Despite its success, the approach has limitations. Wikipedia coverage is not uniform, creating "blind spots" in extremely remote or marginalized areas. Additionally, this study focused on English Wikipedia; incorporating local-language versions could provide deeper cultural context. Future research may also explore how the temporal evolution of Wikipedia text correlates with economic growth over time.
Summary updated