Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Contrastive Learning for Few-shot Geolocation of Social Posts

Menglin Li, Kwan Hui Lim

TL;DR

This paper tackles the problem of geolocating social posts when labeled data are scarce and location distributions are highly imbalanced. It introduces ContrastGeo, a few-shot framework that combines Tweet-Location Contrastive Learning (TLC) to align tweet and location representations with a Tweet-Location Matching (TLM) objective that uses online hard negative mining, all within a flexible fusion module to form joint representations. The approach is evaluated on three public datasets and shows consistent, substantial improvements over state-of-the-art baselines in 1–16 shot settings, aided by fine-tuning a pre-trained language model and careful design choices such as pooling and prompts. The work highlights the effectiveness of contrastive learning for cross-entity alignment in geolocation and suggests broader applicability to other social analysis tasks in data-scarce regimes.

Abstract

Social geolocation is an important problem of predicting the originating locations of social media posts. However, this task is challenging due to the need for a substantial volume of training data, alongside well-annotated labels. These issues are further exacerbated by new or less popular locations with insufficient labels, further leading to an imbalanced dataset. In this paper, we propose \textbf{ContrastGeo}, a \textbf{Contrast}ive learning enhanced framework for few-shot social \textbf{Geo}location. Specifically, a Tweet-Location Contrastive learning objective is introduced to align representations of tweets and locations within tweet-location pairs. To capture the correlations between tweets and locations, a Tweet-Location Matching objective is further adopted into the framework and refined via an online hard negative mining approach. We also develop three fusion strategies with various fusion encoders to better generate joint representations of tweets and locations. Comprehensive experiments on three social media datasets highlight ContrastGeo's superior performance over several state-of-the-art baselines in few-shot social geolocation.

Leveraging Contrastive Learning for Few-shot Geolocation of Social Posts

TL;DR

This paper tackles the problem of geolocating social posts when labeled data are scarce and location distributions are highly imbalanced. It introduces ContrastGeo, a few-shot framework that combines Tweet-Location Contrastive Learning (TLC) to align tweet and location representations with a Tweet-Location Matching (TLM) objective that uses online hard negative mining, all within a flexible fusion module to form joint representations. The approach is evaluated on three public datasets and shows consistent, substantial improvements over state-of-the-art baselines in 1–16 shot settings, aided by fine-tuning a pre-trained language model and careful design choices such as pooling and prompts. The work highlights the effectiveness of contrastive learning for cross-entity alignment in geolocation and suggests broader applicability to other social analysis tasks in data-scarce regimes.

Abstract

Social geolocation is an important problem of predicting the originating locations of social media posts. However, this task is challenging due to the need for a substantial volume of training data, alongside well-annotated labels. These issues are further exacerbated by new or less popular locations with insufficient labels, further leading to an imbalanced dataset. In this paper, we propose \textbf{ContrastGeo}, a \textbf{Contrast}ive learning enhanced framework for few-shot social \textbf{Geo}location. Specifically, a Tweet-Location Contrastive learning objective is introduced to align representations of tweets and locations within tweet-location pairs. To capture the correlations between tweets and locations, a Tweet-Location Matching objective is further adopted into the framework and refined via an online hard negative mining approach. We also develop three fusion strategies with various fusion encoders to better generate joint representations of tweets and locations. Comprehensive experiments on three social media datasets highlight ContrastGeo's superior performance over several state-of-the-art baselines in few-shot social geolocation.
Paper Structure (17 sections, 4 equations, 4 figures, 8 tables)

This paper contains 17 sections, 4 equations, 4 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Contrastive Learning
  4. Social Geolocation
  5. Social Geolocation with Contrastive Learning
  6. Few-Shot Social Geolocation
  7. Proposed Method
  8. Background: Contrastive Learning
  9. Model Architecture
  10. Tweet-Location Contrastive Learning
  11. Tweet-Location Matching
  12. Fusion Module
  13. Experiments
  14. Evaluation Setting
  15. Few-Shot Social Geolocation
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Cumulative distribution of tweet counts against locations from the Twitter-SG dataset.
  • Figure 2: Model Architecture of ContrastGeo.
  • Figure 3: Three types of the fusion module. For Sum type and Concat type, the fusion encoder can be Multi-Head Attention (MHA), a Bottle-Neck Adapter (BNA), Transformer Encoder (TE), and Multi-Layer Perceptron (MLP).
  • Figure 4: Few-shot performance comparison between ContrastGeo, ClassGeo, and representative geolocation models, on Twitter-Mel, Flickr-Mel, Twitter-SG in terms of $accuracy$. Our ContrastGeo shows consistent superiority to other models under 1, 2, 4, 6, 8, 10, 12, 14, 16-shot settings.