What it does
This English to French translator takes an English sentence as input and outputs the French translation.
How we built it
With an Enlish-French dataset from Kaggle, I first preprocessed the data by "cleaning" the sentences: put everything in lowercase, removed punctuations, etc. Then using the keras library, I tokenized the French and English sentences and performed word embeddings. I then encoded sentences into vectors and fed these vectors into a single-layered unidirectional LSTM Encoder-Decoder, also built using the keras library. I used this model to predict French sentences/translations from English sentences.
Challenges we ran into
The model was evaluated using BLEU-scores (BLEU-1 to BLEU-4). These scores were not great (BLEU-1: 0.480895, BLEU-2: 0.370989, BLEU-3: 0.304639, BLEU-4: 0.174916). I tried to improve them by changing my model to be multi-layered and bidirectional but the scores remained the same or decreased.
Future Directions
Future directions include trying to improve the performance of the model, potentially by preprocessing the data differently, using a larger training set, and modifying a different model structure.
Log in or sign up for Devpost to join the conversation.