emotion-classification-model

This model is a fine-tuned version of distilbert-base-uncased. It achieves the following results on the evaluation set:

• Loss: 0.1789
• Accuracy: 0.931

Model Description

The Emotion Classification Model is a fine-tuned version of the distilbert-base-uncased transformer architecture, adapted specifically for classifying text into six distinct emotions. DistilBERT, a distilled version of BERT, offers a lightweight yet powerful foundation, enabling efficient training and inference without significant loss in performance.

This model leverages the pre-trained language understanding capabilities of DistilBERT to accurately categorize textual data into the following emotion classes:

• Sadness
• Joy
• Love
• Anger
• Fear
• Surprise

By fine-tuning on the dair-ai/emotion dataset, the model has been optimized to recognize and differentiate subtle emotional cues in various text inputs, making it suitable for applications that require nuanced sentiment analysis and emotional intelligence.

Intended Uses & Limitations

Intended Uses

The Emotion Classification Model is designed for a variety of applications where understanding the emotional tone of text is crucial. Suitable use cases include:

• Sentiment Analysis: Gauging customer feedback, reviews, and social media posts to understand emotional responses.
• Social Media Analysis: Tracking and analyzing emotional trends and public sentiment across platforms like Twitter, Facebook, and Instagram.
• Content Recommendation: Enhancing recommendation systems by aligning content suggestions with users' current emotional states.
• Chatbots and Virtual Assistants: Enabling more empathetic and emotionally aware interactions with users.

Limitations

While the Emotion Classification Model demonstrates strong performance across various tasks, it has certain limitations:

• Bias in Training Data: The model may inherit biases present in the dair-ai/emotion dataset, potentially affecting its performance across different demographics, cultures, or contexts.
• Contextual Understanding: The model analyzes text in isolation and may struggle with understanding nuanced emotions that depend on broader conversational context or preceding interactions.
• Language Constraints: Currently optimized for English, limiting its effectiveness with multilingual or non-English inputs without further training or adaptation.
• Emotion Overlap: Some emotions have overlapping linguistic cues, which may lead to misclassifications in ambiguous text scenarios.
• Dependence on Text Quality: The model's performance can degrade with poorly structured, slang-heavy, or highly informal text inputs.

Training and Evaluation Data

Dataset

The model was trained and evaluated on the dair-ai/emotion dataset, a comprehensive collection of textual data annotated for emotion classification.

Dataset Statistics

• Total Samples: 20,000
  • Training Set: 16,000 samples
  • Validation Set: 2,000 samples
  • Test Set: 2,000 samples

Data Preprocessing

Prior to training, the dataset underwent the following preprocessing steps:

1. Tokenization: Utilized the DistilBertTokenizerFast from the distilbert-base-uncased model to tokenize the input text. Each text sample was converted into token IDs, ensuring compatibility with the DistilBERT architecture.
2. Padding & Truncation: Applied padding and truncation to maintain a uniform sequence length of 32 tokens. This step ensures efficient batching and consistent input dimensions for the model.
3. Batch Processing: Employed parallel processing using all available CPU cores minus one to expedite the tokenization process across training, validation, and test sets.
4. Format Conversion: Converted the tokenized datasets into PyTorch tensors to facilitate seamless integration with the PyTorch-based Trainer API.

Evaluation Metrics

The model's performance was assessed using the following metrics:

• Accuracy: Measures the proportion of correctly predicted samples out of the total samples.

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

• Learning Rate: 6e-05
• Training Batch Size: 16 per device
• Evaluation Batch Size: 32 per device
• Number of Epochs: 2
• Weight Decay: 0.01
• Gradient Accumulation Steps: 2 (effectively simulating a batch size of 32)
• Mixed Precision Training: Enabled (Native AMP) if CUDA is available

Optimization Strategies

• Mixed Precision Training: Utilized PyTorch's Native AMP to accelerate training and reduce memory consumption when a CUDA-enabled GPU is available.
• Gradient Accumulation: Implemented gradient accumulation with 2 steps to effectively increase the batch size without exceeding GPU memory limits.
• Checkpointing: Configured to save model checkpoints at the end of each epoch, retaining only the two most recent checkpoints to manage storage efficiently.

Training Duration

• Total Training Time: Approximately 2.40 minutes

Logging and Monitoring

• Logging Directory: ./logs
• Logging Steps: Every 10 steps
• Reporting To: TensorBoard
• Tools Used: TensorBoard for real-time visualization of training metrics, including loss and accuracy.

Training Results

After training, the model achieved the following performance metrics:

• Validation Accuracy: 93.10%
• Test Accuracy: 93.10%