Location Prediction for Tweets.

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Mining

Background:

• Geographic information is valuable for data mining and social media analysis.
• User reluctance to share location data stems from privacy and convenience concerns.
• Existing methods struggle with long-term information modeling and explainability.

Purpose of the Study:

• To develop a deep learning solution for predicting geographic information from tweets.
• To address limitations of current approaches in modeling long-term information and enhancing model explainability.

Main Methods:

• Utilized a multi-head self-attention model for advanced text representation.
• Incorporated subword features to improve performance on informal language.
• Jointly trained the model with city and country labels for comprehensive information integration.

Main Results:

• Achieved competitive accuracy compared to state-of-the-art systems on the W-NUT 2016 Geo-tagging task.
• Demonstrated superior distance measurements over existing geo-prediction methods.
• The model effectively predicts geographic information from textual data.

Conclusions:

• The proposed deep learning model offers a robust solution for tweet geo-prediction.
• The multi-head self-attention and subword feature integration enhance model performance and explainability.
• This approach provides a valuable tool for leveraging location insights from social media data.