SufiSent - Universal Sentence Representations Using Suffix Encodings.

Computing universal distributed representations of sentences is a fundamentaltask in natural language processing. We propose a method to learn suchrepresentations by encoding the suffixes of word sequences in a sentence andtraining on the Stanford Natural Language Inference (SNLI) dataset. Wedemonstrate the effectiveness of our approach by evaluating it on the SentEvalbenchmark, improving on existing approaches on several transfer tasks.