Predicting Part of Speech and Next Word in Scottish Gaelic

Overview:

Developed two neural network models for Scottish Gaelic, a low-resource language, in collaboration with Edinburgh Napier University.

Part-of-Speech Tagger

• Built a BiLSTM-based neural network to tag 87K tokens from the ARCOSG dataset
• Achieved 89% accuracy across 223 unique PoS tags
• Visualized architecture with input, hidden, and output layers

Next-Word Prediction (Small Language Model)

• Preprocessed and trained on 1M tokens using an LSTM model from the Institutional Books Corpus
• Predicted next words in Gaelic phrases

Tools:

PyTorch, FastText, spaCy, Jupyter Notebook, Python

Future Plans:

• Scale training data from 1M to 40M tokens for improved predictions
• Integrate the PoS model into spaCy for open-source usage

CompassQA – AI-powered Student Policy Assistant

Built a RAG system that enables students to ask questions about Cornell College’s student policies, including academic, residential, and financial rules.

Tools:

LangChain, Ollama, Streamlit