Multilingual POS Tagging & Context-Aware Spell Correction

Year

2024

Tech & Technique

Python, HMM / Viterbi, RNN, LSTM, BiLSTM

Description

Part-of-speech tagging and spelling correction across 3 languages (English, Japanese, Bulgarian). Compared a hidden Markov model tagger (Viterbi decoding) with RNN, LSTM, and BiLSTM taggers. A bigram HMM reached a 5.4% English word error rate.

Key Facts:

🧠 Models: HMM with Viterbi decoding; RNN, LSTM, BiLSTM.
🌍 Languages: English, Japanese, Bulgarian.
📊 Result: 5.4% English word error rate for a bigram HMM.
✍️ Spelling correction.

My Role

Project — POS tagging and spelling correction

🧠 Implemented and compared HMM/Viterbi, RNN, LSTM, and BiLSTM taggers.
📊 A bigram HMM reached a 5.4% English word error rate; evaluated on 3 languages.

POS Tagging Example

Illustration of part-of-speech tags on an example sentence.

spaCy Example (illustrative)

Illustrative spaCy dependency-parse snippet — not code from this project.

Learning Curves

Error rate by word and by sentence vs training-set size for the HMM tagger.