Scholarlink

<p>The ubiquitous nature of social media platforms has generated vast volumes of textual data, offering a rich source of information for various applications. Such social media platforms are employed by users to post about an event or activity that often describes their surroundings. Also, these posts are often associated with location attributes. In the realm of geolocation extraction, leveraging this textual content presents promising avenues for understanding users’ locations and behaviors. This research introduces a novel approach for extracting users’ locations at the time of posting from textual content. Leveraging embeddings to discern context, along with feature expansion techniques such as deconvolutions, the study employs an attention mechanism to emphasize pertinent information critical for location determination. This study proposes the Geographic Attention model for Extraction (GAttE) that incorporates these techniques by transforming text into sentence, word, and character embeddings. These combined embeddings are subsequently fed through deconvolutional layers and a multi-head attention mechanism of the model, to conclude the extraction of users’ current location information. Comparative analysis against prior methodologies and a baseline underscores the model’s superior performance, achieving an F1-score of 0.55 and an accuracy of 0.54, which is a 49% relative improvement in accuracy compared to BERT. Moreover, GAttE demonstrates lower spatial error, reducing the average distance error to 998.98&#xa0;km, compared to 5470.69&#xa0;km for BERT and 10,116.21&#xa0;km for Spacy. Component-level evaluation and testing samples further confirm the model’s consistency across metrics. In conclusion, GAttE demonstrates significant gains in both extraction accuracy and spatial precision, establishing it as a strong solution for location inference from social media text.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GAttE: geographic attention model for extraction of users’ current locations from social media texts

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Ashish Kumar,
  • Manisha Dubey,
  • Sangeeta Sharma

摘要

The ubiquitous nature of social media platforms has generated vast volumes of textual data, offering a rich source of information for various applications. Such social media platforms are employed by users to post about an event or activity that often describes their surroundings. Also, these posts are often associated with location attributes. In the realm of geolocation extraction, leveraging this textual content presents promising avenues for understanding users’ locations and behaviors. This research introduces a novel approach for extracting users’ locations at the time of posting from textual content. Leveraging embeddings to discern context, along with feature expansion techniques such as deconvolutions, the study employs an attention mechanism to emphasize pertinent information critical for location determination. This study proposes the Geographic Attention model for Extraction (GAttE) that incorporates these techniques by transforming text into sentence, word, and character embeddings. These combined embeddings are subsequently fed through deconvolutional layers and a multi-head attention mechanism of the model, to conclude the extraction of users’ current location information. Comparative analysis against prior methodologies and a baseline underscores the model’s superior performance, achieving an F1-score of 0.55 and an accuracy of 0.54, which is a 49% relative improvement in accuracy compared to BERT. Moreover, GAttE demonstrates lower spatial error, reducing the average distance error to 998.98 km, compared to 5470.69 km for BERT and 10,116.21 km for Spacy. Component-level evaluation and testing samples further confirm the model’s consistency across metrics. In conclusion, GAttE demonstrates significant gains in both extraction accuracy and spatial precision, establishing it as a strong solution for location inference from social media text.