Universal Sentence Encoder

TLDR

The paper introduces sentence‑encoding models designed to facilitate transfer learning for downstream NLP tasks. Two variants of the encoder are proposed, differing in accuracy–compute trade‑offs, and their performance is evaluated against word‑embedding baselines across varying data sizes and resource budgets. The models achieve efficient, accurate transfer performance, surpassing word‑embedding baselines, performing well even with limited training data, and demonstrating low bias on WEAT, with pretrained models publicly released.

Abstract

We present models for encoding sentences into embedding vectors that specifically target transfer learning to other NLP tasks. The models are efficient and result in accurate performance on diverse transfer tasks. Two variants of the encoding models allow for trade-offs between accuracy and compute resources. For both variants, we investigate and report the relationship between model complexity, resource consumption, the availability of transfer task training data, and task performance. Comparisons are made with baselines that use word level transfer learning via pretrained word embeddings as well as baselines do not use any transfer learning. We find that transfer learning using sentence embeddings tends to outperform word level transfer. With transfer learning via sentence embeddings, we observe surprisingly good performance with minimal amounts of supervised training data for a transfer task. We obtain encouraging results on Word Embedding Association Tests (WEAT) targeted at detecting model bias. Our pre-trained sentence encoding models are made freely available for download and on TF Hub.