Multi-Lingual Theme Prediction for Customer Reviews Using Deep Pre-Trained Embeddings and Siamese Networks Description Train a category prediction model on the Amazon product review dataset based on LASER embeddings per review The model must be based on a Siamese Network and a reasonable sampling strategy Evaluate on German without training on German data Fine-tune and evaluate on our Organic Dataset for relevance, entity, and attribute classification (each task independently) Team Size 3 References

LASER One Shot Learning and Siamese Networks in Keras One Shot Learning with Siamese Networks in PyTorch Triplet Loss and Online Triplet Mining in TensorFlow Sampling Matters in Deep Embedding Learning

Data

Organic Dataset (annotated) Amazon product review data Multi-lingual version of Amazon product review dataset

Contributors:- Vindhya Singh Chinmay Prabhakar Wasiq Rumaney *

• All members have equal contribution.