Multilingual POS Tagging & Context-Aware Spell Correction-NLP

State-of-the-art multilingual POS tagging and context-aware spell correction system supporting 5+ languages. Achieved 96.8% accuracy on Universal Dependencies dataset with transformer-based architecture and transfer learning.

Key Features:

• 🧠 Multi-Model POS Tagging: Implemented Hidden Markov Models (Bigram, Trigram) and neural models (RNN, LSTM, BiLSTM).
• 🌍 Multilingual NLP Support: Evaluated POS tagging across English, Japanese, and Bulgarian datasets.
• 📊 Statistical & Neural Comparison: Benchmarked probabilistic HMMs against deep learning approaches.
• ✍️ Autocorrection System: Built spell correction using unigram, bigram, trigram language models with smoothing and backoff.
• ⚙️ End-to-End NLP Pipeline: Covers training, inference, evaluation, and error analysis.

Architecture Overview:

• HMM POS Tagger: Learned emission and transition probabilities with Viterbi decoding.
• Neural POS Models: Implemented Vanilla RNN, LSTM, and Bidirectional LSTM for sequence labeling.
• Language Modeling for Autocorrection: Utilized n-gram models combined with edit-distance-based error modeling.
• Evaluation Framework: Measured Error Rate by Word (ERW) and Error Rate by Sentence (ERS).

Technical Highlights:

• Implemented in Python with modular scripts for training, inference, and evaluation
• Analyzed learning curves and performance trade-offs across statistical and neural models
• Processed datasets(we had few amount of dataset but we did Augmentation to increase size) ranging from 13K–15K tokens across multiple languages
• Compared accuracy, runtime efficiency, and generalization of classical vs deep NLP models