Abstract
Nowadays, geographic information related to Twitter is crucially important for fine-grained applications. However, the amount of geographic information available on Twitter is low, which makes the pursuit of many applications challenging. Under such circumstances, estimating the location of a tweet is an important goal of the study. Unlike most previous studies that estimate the pre-defined district as the classification task, this study employs a probability distribution to represent richer information of the tweet, not only the location but also its ambiguity. To realize this modeling, we propose the convolutional mixture density network (CMDN), which uses text data to estimate the mixture model parameters. Experimentally obtained results reveal that CMDN achieved the highest prediction performance among the method for predicting the exact coordinates. It also provides a quantitative representation of the location ambiguity for each tweet that properly works for extracting the reliable location estimations.
Introduction
Geographic information is vital for real-time applications such as earthquake detection and disease tracking. However, fewer than 0.5% of tweets are natively geo-tagged, creating a significant data gap. Estimating a tweet's location from its text is inherently difficult due to a lack of explicit clues, linguistic ambiguity (e.g., "Portland" referring to multiple cities), and contextual mismatches where a user mentions a place they are not currently visiting.
Previous research has typically approached geolocation as either a classification task (assigning tweets to districts) or a regression task (predicting latitude/longitude). Standard regression models, particularly those using L2 loss, often suffer from "regression towards the mean." When faced with ambiguity—such as a tweet that could originate from two distant cities—these models minimize error by predicting a point halfway between them, often resulting in nonsensical locations in the middle of the ocean or unpopulated areas.
Methodology: Convolutional Mixture Density Network
To address the limitations of point-based regression, we propose the Convolutional Mixture Density Network (CMDN). Instead of a single coordinate pair, CMDN estimates the parameters of a Gaussian Mixture Model (GMM), representing the location as a probability distribution.
The architecture consists of:
- Feature Extraction: A Convolutional Neural Network (CNN) processes tweet text to extract semantic features.
- Density Estimation: The CNN's output is mapped to GMM parameters: mixture weights, means, variances, and correlations.
- Stable Training: To prevent gradient explosion common in Mixture Density Networks, we utilize
softplusactivations for variance andsoftsignfor correlation. - Prediction: The final location is determined by the mode (the highest peak) of the estimated probability distribution.
Dataset
The model was trained and evaluated on a massive corpus of 24,633,478 Japanese tweets collected between July 2011 and July 2012. The dataset was split into 20 million tweets for training, and 2 million each for development and testing.
Results and Discussion
CMDN demonstrates a significant improvement in geolocation accuracy, particularly in terms of median error. By modeling the location as a distribution, the network can be "aggressive"—assigning high probability to specific, likely locations rather than "conservative" averaging.
Performance Metrics
- Median Accuracy: CMDN achieved a median error distance of 10.7 km, vastly outperforming CNN-L1 (28.2 km) and CNN-L2 (80.5 km).
- Reliability Filtering: The likelihood score of the estimated distribution serves as a built-in confidence metric. By filtering out low-likelihood (high ambiguity) predictions, the system can extract a highly reliable subset of geo-located tweets for downstream applications.
- Mean vs. Median: While CMDN leads in median accuracy, its mean error (159.4 km) is slightly higher than CNN-L1 (147.5 km). This is due to occasional "aggressive" but incorrect guesses that act as outliers, skewing the mean.
Conclusion
The CMDN framework successfully leverages text data to estimate location probability distributions, offering a robust solution to the ambiguity inherent in social media text. While currently limited to text-only features and a flat coordinate approximation for Japan, future work will explore incorporating user metadata and expanding the model to a global scale. Integrating temporal user movement via Recurrent Neural Networks (RNNs) remains a promising direction for further improving geolocation precision.
Summary updated