IT Support Ticket Classification Using DistilBERT
This project implements an IT Support Ticket Classification system using DistilBERT.
The objective is to automatically classify IT support tickets into predefined categories based on the text description of the ticket.
The project uses the pretrained DistilBERT (distilbert-base-uncased) transformer model and fine-tunes it for multi-class text classification.
Project Information
| Information | Details |
|---|---|
| Project Title | IT Support Ticket Classification Using DistilBERT |
| Author | Sonu Kumar |
| USN | SCA25MSD066 |
| Base Model | distilbert-base-uncased |
| Task | Multi-class Text Classification |
| Number of Classes | 8 |
Support Categories
The model classifies IT support tickets into the following categories:
- Access
- Administrative rights
- HR Support
- Hardware
- Internal Project
- Miscellaneous
- Purchase
- Storage
Dataset
The dataset contains IT service and support tickets represented as text along with their corresponding topic/category.
Main Columns
| Column | Description |
|---|---|
Document |
IT support ticket description |
Topic_group |
Support category |
Data Preprocessing
The preprocessing pipeline includes:
- Removing missing ticket descriptions.
- Removing missing category values.
- Removing duplicate records.
- Encoding category labels.
- Splitting the dataset into training, validation, and test sets.
- Tokenizing the text using the DistilBERT tokenizer.
- Applying a maximum sequence length of 128 tokens.
Model
The project uses:
Base Model: distilbert-base-uncased
Architecture: DistilBERT
Task: Multi-class text classification
Maximum Sequence Length: 128
Training Configuration
| Parameter | Value |
|---|---|
| Model | distilbert-base-uncased |
| Number of Classes | 8 |
| Maximum Length | 128 |
| Training Epochs | 2 |
| Learning Rate | 2e-5 |
| Weight Decay | 0.01 |
| Training Batch Size | 16 |
| Evaluation Batch Size | 32 |
| Random Seed | 42 |
| Data Split | 80/10/10 |
Evaluation Metrics
The model is evaluated using:
- Accuracy
- Precision
- Recall
- F1 Score
Final metric values should be added after completing model evaluation.
| Metric | Score |
|---|---|
| Accuracy | Add Result |
| Precision | Add Result |
| Recall | Add Result |
| F1 Score | Add Result |
Example Prediction
Input
I am unable to access my company email account.
Predicted Category
Access
Another example:
Input
I need administrator privileges to install required software.
Predicted Category
Administrative rights
Using the Model
from transformers import AutoTokenizer, AutoModelForSequenceClassification
model_id = "YOUR_USERNAME/it-support-ticket-distilbert"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
Replace YOUR_USERNAME with your Hugging Face username.
Making a Prediction
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
model_id = "YOUR_USERNAME/it-support-ticket-distilbert"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
ticket = "I cannot access my company email."
inputs = tokenizer(
ticket,
return_tensors="pt",
truncation=True,
padding=True,
max_length=128
)
with torch.no_grad():
outputs = model(**inputs)
prediction = torch.argmax(outputs.logits, dim=-1).item()
print("Predicted Label:", prediction)
Technologies Used
Python
PyTorch
Hugging Face Transformers
Hugging Face Datasets
Scikit-learn
Pandas
NumPy
Matplotlib
Seaborn
DistilBERT
Applications
This model can be used for:
Automatic IT ticket routing
IT helpdesk automation
Support ticket categorization
Helpdesk workflow automation
Automatic assignment of support categories
Reducing manual ticket classification
Intelligent IT service management
Limitations
The model's performance depends on the quality and distribution of the training dataset.
The model may perform poorly on tickets that are significantly different from the training data.
Predictions should be treated as automated classification suggestions rather than guaranteed decisions.
This project is intended for educational and research purposes.
Future Improvements
Possible improvements include:
Increasing the training dataset
Adding more IT support categories
Hyperparameter tuning
Training for more epochs
Using larger transformer models
Adding ticket priority prediction
Adding urgency detection
Building a Streamlit web application
Creating a FastAPI prediction API
Integrating the model with an IT helpdesk system
Project Structure
it-support-ticket-distilbert/
│
├── README.md
├── config.json
├── model.safetensors
├── tokenizer.json
├── tokenizer_config.json
├── special_tokens_map.json
├── vocab.txt
├── label_mapping.json
├── notebook/
│ └── it_support_ticket_classification.ipynb
└── test_predictions.csv
Author
Sonu Kumar
USN: SCA25MSD066
License
This project is intended for educational and academic purposes.
- Downloads last month
- 43