Data for "Based on Data Balancing and Model Improvement for Multi-Label Emotion Recognition"

Dataset Description Data for "Based on Data Balancing and Model Improvement for Multi-Label Emotion Recognition". This repository contains the comprehensive data and experimental results supporting our study on multi-label emotion recognition using the GoEmotions dataset. The dataset and materials are shared under the CC-BY 4.0 license. Core Dataset balanced_emotion_dataset.csv Final balanced multi-label sentiment dataset used for training and evaluation. Renamed from final_balanced_df_output.csv. Columns: text, sentiment (list of emotion labels). goemotions_original.csv Original GoEmotions data after removing example_very_unclear. sentiment140_auto_labels.csv Sentiment140 tweets labeled into the 28 GoEmotions categories. Columns include text and model_labels. gpt4mini_generated_texts.csv GPT-4 mini generated texts with target emotion prompts. Original Submission Data (Version 1) Data for Figures balanced_label_counts.csv Renamed from fig2_balanced_label_counts.csv. Counts of each of the 28 emotion labels in the final balanced dataset. Columns: Sentiment Labels, Counts. training_history.csv Training history log for figures (loss and accuracy per epoch). Columns: epoch, accuracy, loss, val_accuracy, val_loss. Source Code model_pipeline.ipynb Renamed from model (1).ipynb. Full notebook for data processing, model training, and evaluation. Updated Experimental Results (Version 2) In response to reviewer feedback, we conducted ablation studies and baseline comparisons. The following ablation archives are included: ablation_unbalanced_attn.tar.gz CNN + BiLSTM + Attention on original unbalanced GoEmotions. ablation_unbalanced_noattn.tar.gz CNN + BiLSTM (no attention) on original unbalanced GoEmotions. ablation_balanced_attn.tar.gz CNN + BiLSTM + Attention on oversampled balanced GoEmotions. Key Updates in Version 2 Extended Training: all models trained for 34 epochs (no early stopping). Validation-Only Threshold Optimization: thresholds tuned on validation only. Comprehensive Metrics: Subset accuracy Jaccard index Hamming loss Micro/Macro Precision, Recall, F1-score Macro AUC Per-label metrics for all 28 emotion categories File Structure (inside each ablation archive) *_loss.png *_precision.png *_recall.png per_label_metrics_thr0.5.csv per_label_metrics_thr_opt.csv f1_thr05.png f1_thr_opt.png summary.json Quality Control and Audits sentiment140_audit.csv Audit samples for Sentiment140 auto-labels. gpt4mini_annotations.csv Five-annotator labels with majority vote. gpt4mini_audit.csv Audit samples for GPT-4 mini generated texts. Transformer Baseline (Version 3) transformer_baseline_train.py DistilRoBERTa baseline training script. transformer_baseline_requirements.txt Python dependencies for the baseline. transformer_baseline_summary.json Overall metrics at threshold 0.5 and optimized thresholds. transformer_baseline_per_label_thr0.5.csv Per-label metrics at threshold 0.5. transformer_baseline_per_label_thr_opt.csv Per-label metrics under validation-tuned thresholds. transformer_baseline_thresholds_opt.csv Optimized thresholds per label. Scripts data_balancing_pipeline.py Data integration, filtering, and balancing logic. cnn_bilstm_training.py CNN + BiLSTM + Attention training and evaluation script. transformer_baseline_train.py Transformer baseline training script. Notes Split protocol: 80/10/10 with MultilabelStratifiedShuffleSplit, random_state=42. Threshold optimization: per-label grid search from 0.05 to 0.95 (step 0.05). File names are normalized and do not include timestamps or parentheses. Citation If you use this dataset in your research, please cite the paper associated with this repository. Contact For questions about the data or experiments, please contact the corresponding author.