Abstract:News event detection involves the identification and categorization of significant happenings or occurrences from social media data. Recent work has typically relied on datasets collected solely based on event-related keywords. However, datasets collected with such keywords tend to oversimplify the task. They reduce the contribution of non-text modalities and do not fully capture real-world scenarios where news events involve diverse expressions and media forms. To address these limitations, we introduce a News Event Detection (NED) dataset, comprising 17,366 posts with text-image pairs from Twitter, annotated with 40 real-world events. Unlike previous datasets, our NED dataset is collected based on users’ hashtags rather than keywords. This method captures more diverse event-related content and avoids the limitations of keyword searches. Additionally, we propose a Multi-modal Fusion with External Knowledge (MFEK) model to address the out-of-distribution (OOD) issue commonly encountered in news event detection. The MFEK model features a text enrichment module that leverages image semantic information to enhance the textual content, a knowledge extraction module that extracts explicit and implicit external knowledge to mitigate the OOD issue, and a knowledge-aware feature fusion module that employs a co-attention mechanism to integrate external knowledge, text, and images, while filtering out irrelevant information. Extensive experiments validate the superior performance of our MFEK model on the NED dataset for the task of news event detection, achieving a 5.48% increase in the F1 score compared to the current state-of-the-art model. The NED dataset is available at https://github.com/RetrainIt/NED.

Image Enhanced Event Detection in News Articles

Image Enchanced Event Detection in News Articles. Meihan Tong, Shuai Wang

Training Multimedia Event Extraction With Generated Images and Captions

Multi-modal news event detection with external knowledge

Context Event Features and Event Embedding Enhanced Event Detection

OntoED: Low-resource Event Detection with Ontology Embedding

Enhancing Unimodal Features Matters: A Multimodal Framework for Building Extraction.

Discriminative Multimodal Embedding for Event Classification

Enhancing Video Event Recognition Using Automatically Constructed Semantic-Visual Knowledge Base.

Cross-Modal Event Retrieval: A Dataset and a Baseline Using Deep Semantic Learning.

Exploiting Semantic Space to Enhance Event Detection Combined with Event Knowledge

Columbia-UCF TRECVID2010 Multimedia Event Detection: Combining Multiple Modalities, Contextual Concepts, and Temporal Matching.

Improving Event Detection via Open-domain Trigger Knowledge

Reliable Shot Identification for Complex Event Detection Via Visual-Semantic Embedding.

Learning Discriminative Neural Representations for Event Detection

Event extraction as machine reading comprehension with question-context bridging

EDSC: Event Detection with Superior Conceptual Information

Incorporating Image Matching into Knowledge Acquisition for Event-Oriented Relation Recognition.

Extracting Biomedical Event with Dual Decomposition Integrating Word Embeddings

Learning Event Representations for Zero-Shot Detection via Dual-Contrastive Prompting

Dual-path Rare Content Enhancement Network for Image and Text Matching