Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TriMod Fusion for Multimodal Named Entity Recognition in Social Media

Mosab Alfaqeeh

TL;DR

This work tackles Named Entity Recognition on social media by leveraging multiple modalities—text, images, and hashtags—through a Transformer-based fusion framework (TriMod). The method combines textual embeddings and character-aware representations, image-derived captions, and hashtag processing, then fuses these via attention and decodes with a CRF to capture label dependencies. Evaluated on a Twitter dataset with four entity categories, TriMod achieves state-of-the-art performance, significantly outperforming both text-only and prior multimodal approaches in precision, recall, and F1. The approach demonstrates strong cross-modal grounding and robustness across entity types, with practical implications for social media mining and information extraction.

Abstract

Social media platforms serve as invaluable sources of user-generated content, offering insights into various aspects of human behavior. Named Entity Recognition (NER) plays a crucial role in analyzing such content by identifying and categorizing named entities into predefined classes. However, traditional NER models often struggle with the informal, contextually sparse, and ambiguous nature of social media language. To address these challenges, recent research has focused on multimodal approaches that leverage both textual and visual cues for enhanced entity recognition. Despite advances, existing methods face limitations in capturing nuanced mappings between visual objects and textual entities and addressing distributional disparities between modalities. In this paper, we propose a novel approach that integrates textual, visual, and hashtag features (TriMod), utilizing Transformer-attention for effective modality fusion. The improvements exhibited by our model suggest that named entities can greatly benefit from the auxiliary context provided by multiple modalities, enabling more accurate recognition. Through the experiments on a multimodal social media dataset, we demonstrate the superiority of our approach over existing state-of-the-art methods, achieving significant improvements in precision, recall, and F1 score.

TriMod Fusion for Multimodal Named Entity Recognition in Social Media

TL;DR

This work tackles Named Entity Recognition on social media by leveraging multiple modalities—text, images, and hashtags—through a Transformer-based fusion framework (TriMod). The method combines textual embeddings and character-aware representations, image-derived captions, and hashtag processing, then fuses these via attention and decodes with a CRF to capture label dependencies. Evaluated on a Twitter dataset with four entity categories, TriMod achieves state-of-the-art performance, significantly outperforming both text-only and prior multimodal approaches in precision, recall, and F1. The approach demonstrates strong cross-modal grounding and robustness across entity types, with practical implications for social media mining and information extraction.

Abstract

Social media platforms serve as invaluable sources of user-generated content, offering insights into various aspects of human behavior. Named Entity Recognition (NER) plays a crucial role in analyzing such content by identifying and categorizing named entities into predefined classes. However, traditional NER models often struggle with the informal, contextually sparse, and ambiguous nature of social media language. To address these challenges, recent research has focused on multimodal approaches that leverage both textual and visual cues for enhanced entity recognition. Despite advances, existing methods face limitations in capturing nuanced mappings between visual objects and textual entities and addressing distributional disparities between modalities. In this paper, we propose a novel approach that integrates textual, visual, and hashtag features (TriMod), utilizing Transformer-attention for effective modality fusion. The improvements exhibited by our model suggest that named entities can greatly benefit from the auxiliary context provided by multiple modalities, enabling more accurate recognition. Through the experiments on a multimodal social media dataset, we demonstrate the superiority of our approach over existing state-of-the-art methods, achieving significant improvements in precision, recall, and F1 score.
Paper Structure (19 sections, 18 equations, 2 figures, 5 tables)

This paper contains 19 sections, 18 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Named Entity Recognition
  4. Multimodal Named Entity Recognition
  5. Problem Statement
  6. Methodology
  7. Textual Feature Extractor:
  8. Visual Feature Extractor:
  9. Hashtag Feature Extraction:
  10. Processing the Output
  11. Fusion:
  12. Output layer
  13. EXPERIMENT
  14. Implementation Details
  15. Comparison With Existing Methodes
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Mappings between visual objects, processed hashtags, and corresponding named entities in the text.
  • Figure 2: The fused representation $f_s$, which integrates information from text, images, and hashtags using Transformer-attention.