Implementation-of-VGCN-BERT-for-Rumor-Detection

Further work about project "Improving Rumor Detection with User Comments". We use VGCN and BERT to extract global and local features from texts and apply Attention Mechanism for feature fusion to improve rumor detection performance even further.

• Previous Project

1. Project Introduction

• Overall Structure

  

• Model Details

  • Segmentation Rumor Text Feature Extraction Module （分段谣言文本特征提取模块）

    • Segementation of long text in datasets.
    • Build a vacabulary graph.
    • Extract local feature (by BERT) of rumor text segments.
    • Extract global feature (by VGCN) of rumor text segments.
    • Concatenate local and global features.

  • Rumor Features Fusion and Classification Module （谣言特征融合与分类模块）

    • Feature fusion with attention mechanism between local and global features.
    • Concatenate features in the order of original segments.
    • Perform rumor feature classification.

---

2. Usage

• Clean Redundant Data

• Format dataset from json to csv

  

1️⃣ data_reorganize_pheme.ipynb / data_reorganize_weibo.ipynb

• Dataset reorganize & segment long text

  

2️⃣ prepare_data.py

• Get Vocabulary Graph

  

3️⃣ train_vgcn_bert.ipynb

• Train VGCN_BERT on graph data and comments data

4️⃣ joint_vgcn_bert_model.ipynb

• Feature concatenate

  

• Apply trained VGCN_BERT model to get rumor detection results

3. Experiment Results

(to be continued...)