This repository contains the implementation of the paper "Connecting Supervised and Unsupervised Sentence Embeddings" presented at the 3rd Workshop on Representation Learning for NLP (rep4NLP) held in conjuction with ACL 2018. Our work is based on the InferSent (Conneau et al., 2017) and based on their implementation.
While (Conneau et al., 2017) uses only a supervised loss, we add to their framework additional unsupervised losses that act as regularization terms. We evaluate our results using SentEval and show improved results on various downstream tasks.
(Conneau et al., 2017) Alexis Conneau, Douwe Kiela, Holger Schwenk, Loı̈c Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. pages 670–680.