YatishW79/cnn-lstm-goemotions

CNN-LSTM + GloVe — GoEmotions Emotion Classifier

CNN trigram extractor followed by Bidirectional LSTM with additive attention for multi-label emotion classification on GoEmotions (28 labels).

CS6120 NLP Project — Northeastern University

Architecture

• GloVe 100d pre-trained embeddings
• Conv1d (kernel=3, 128 filters) for local feature extraction
• 2-layer Bidirectional LSTM (hidden: 256)
• Additive attention pooling
• Per-class threshold tuning on validation set

Performance (GoEmotions test set, 5,427 samples)

Metric	Score
Micro-F1	0.5371
Macro-F1	0.4428
Subset Accuracy	0.3230
Hamming Accuracy	0.9543

Files

• cnn_lstm_model.pt — model weights + config (load with torch.load)
• vocab.pkl — vocabulary dictionary
• thresholds_cnn_lstm.pkl — per-class decision thresholds (numpy array, shape 28)

Usage

import torch, pickle
from lstm_model import CNNLSTMClassifier  # from the Space repo

ckpt = torch.load("cnn_lstm_model.pt", map_location="cpu")
cfg  = ckpt["config"]
model = CNNLSTMClassifier(**cfg)
model.load_state_dict(ckpt["state_dict"])
model.eval()