Towards Generalizable Sentence Embeddings

In this work, we evaluate different sentence encoders with emphasis on examining their embedding spaces. Specifically, we hypothesize that a “high-quality” embedding aids in generalization, promoting transfer learning as well as zero-shot and one-shot learning. To investigate this, we modify Skipthought vectors to learn a more generalizable space by exploiting a small amount of supervision. The aim is to introduce an additional notion of similarity in the embeddings, rendering the vectors informative for different tasks requiring less adaptation. Our embeddings capture human intuition on similarity favorably than competing models, while we also show positive indications of transfer from the task of natural language inference to paraphrase detection and paraphrase ranking. Further, our model’s behaviour on paraphrase detection when trained with an increasing amount of labelled data is indicative of a generalizable model. Finally, we support our hypothesis on generalizability of our embeddings through inspection of their statistics.