my420254/spear-ecpe

SPEAR — Emotion-Cause Pair Extraction

Decoupled Biaffine Scorer with R-Drop regularization and Sliding Window for Emotion-Cause Pair Extraction (ECPE).

Model Details

Item	Value
Base model	bert-base-chinese
Task	Emotion-Cause Pair Extraction
Dataset	Chinese ECPE (Xia & Ding, ACL 2019)
Evaluation	10-fold cross-validation
Avg Precision	79.07%
Avg Recall	75.59%
Avg F1	77.24%
This checkpoint	Fold 7 (F1 = 86.92%)

Architecture

• Encoder: BERT-base-Chinese for clause representation
• GNN: Graph Attention Network (192 dims, 4 heads) for inter-clause reasoning
• Predictor: Decoupled biaffine scorer for emotion-cause pair ranking
• Regularization: R-Drop for robust training
• Sliding Window: Window size K=12 for candidate pair generation

Usage

import torch
from config import Config
from networks.rank_cp import Network

configs = Config()
model = Network(configs)
model.load_state_dict(torch.load("best_model.pt", map_location="cpu"))
model.eval()

Citation

If you use this model, please cite the SPEAR paper.