UniLMv2: Pseudo-Masked Language Models for Unified Language Model Pre-Training


We propose to pre-train a unified language model for both autoencoding and partially autoregressive language modeling tasks using a novel training procedure, referred to as a pseudo-masked language model (PMLM). Given an input text with masked tokens, we rely on conventional masks to learn inter-relations between corrupted tokens and context via autoencoding, and pseudo masks to learn intra-relations between masked spans via partially autoregressive modeling. With well-designed position embeddings and self-attention masks, the context encodings are reused to avoid redundant computation. Moreover, conventional masks used for autoencoding provide global masking information, so that all the position embeddings are accessible in partially autoregressive language modeling. In addition, the two tasks pre-train a unified language model as a bidirectional encoder and a sequence-to-sequence decoder, respectively. Our experiments show that the unified language models pre-trained using PMLM achieve new state-of-the-art results on a wide range of natural language understanding and generation tasks across several widely used benchmarks.

Publication Downloads

UniLM – Unified Language Model Pre-training

October 1, 2019

We develop pre-trained models for natural language understanding (NLU) and generation (NLG) tasks. ***** New October 1st, 2019: UniLM v1 release ***** UniLM v1 (September 30th, 2019): the code and pre-trained models for the NeurIPS 2019 paper entitled "Unified Language Model Pre-training for Natural Language Understanding and Generation". UniLM (v1) achieves the new SOTA results in NLG (especially sequence-to-sequence generation) tasks/benchmarks, including abstractive summarization (the Gigaword and CNN/DM dataset), question generation (the SQuAD QG dataset), etc. UniLM v2: the new pre-training protocol and implementation scheme (coming soon).