NLP

Word Completion using LSTM

Real-time next-word suggestions from an LSTM trained on Shakespeare.

202490% Model accuracy

OVERVIEW

The Project

A word-level LSTM model for sentence completion, trained on Shakespeare's plays (111,396 lines; 11,000 used for training). A real-time interface suggests words as the user types, achieving ~90% accuracy with strong coherence and fluency.

Objectives

Predict the next word in a sequence from prior context.

Build a responsive interface for live word suggestions.

Tune hyperparameters for coherence, fluency and accuracy.

Tools & Technologies

PythonTensorFlow / KerasLSTM

FIGURES

Charts & Graphs

METHODOLOGY

The Approach

Clean and lowercase text; tokenize; build word sequences.

Model: input layer → LSTM layer(s) → dense softmax over vocabulary.

Train with the forget/input/output gate mechanics of LSTM.

Serve predictions via a UI that takes a seed sentence + word count.

OUTCOME

Results & Learnings

~90% training accuracy with closely-matched validation accuracy.

Coherent, fluent generations; clear improvement over a vanilla RNN.

Key Learnings

LSTM gating beats plain RNNs on long-term dependencies.
Balancing complexity prevents overfitting while keeping accuracy high.

Related Case Studies

NLP