Multilingual NLU System — Intent Classification + Slot Filling

Streamlit app for intent classification and slot filling across 51 languages, built with XLM-RoBERTa and trained on the MASSIVE dataset.

Features

• Intent classification: 60+ intents (alarm_set, play_music, weather_query, etc.)
• Slot filling: 40+ slot types (time, place_name, song_name, etc.)
• 51 languages via zero-shot cross-lingual transfer
• Streamlit UI: Enter an utterance and see intent + extracted slots

Local setup

1. Clone and enter the repo

   git clone https://github.com/AdithyaReddyGeeda/Intent-classification-and-slot-filling.git
   cd Intent-classification-and-slot-filling

2. Create a virtual environment and install dependencies

   python -m venv .venv
   source .venv/bin/activate   # Windows: .venv\Scripts\activate
   pip install -r requirements.txt

3. Add the trained models (not in the repo due to size):

   • Intent model: Place your saved checkpoint in results_intent_classification/checkpoint-71965/ (or checkpoint-57572/). The app expects a Hugging Face AutoModelForSequenceClassification checkpoint (e.g. XLM-RoBERTa) with config.json containing id2label.
   • Slot model: Place your saved model in slot_filling_model_crf/final_model/ with tokenizer files and id2label.json.

4. Run the app

   streamlit run app.py

   Open http://localhost:8501.

Deploy on Streamlit Cloud

1. Push this repo to GitHub (code only; models are in .gitignore).

2. Host the models (required for the app to run):

   • Option A: Upload the model folders to Hugging Face Hub (as a private or public model repo), then in app.py replace the local paths with from_pretrained("your-username/your-intent-model") and from_pretrained("your-username/your-slot-model").
   • Option B: Use Streamlit’s “Secrets” to point to a URL or path where the app can download the model files at startup (you’d add a small download step in app.py).

3. Deploy:

   • Go to share.streamlit.io, sign in with GitHub, and select this repo.
   • Set Main file path to app.py.
   • Add any secrets (e.g. HF token) if you use private models.
   • Deploy. The first run may be slow while dependencies and models load.

Project layout

• app.py — Streamlit UI and inference (intent + slot filling)
• requirements.txt — Python dependencies
• results_intent_classification/ — Intent model checkpoints (add locally; gitignored)
• slot_filling_model_crf/final_model/ — Slot filling model (add locally; gitignored)
• *.ipynb — Notebooks for training and evaluation

Requirements

• Python 3.8+
• PyTorch, Transformers, Streamlit (see requirements.txt)