DeltaLM: Encoder-Decoder Pre-training for Language Generation and Translation by Augmenting Pretrained Multilingual Encoders
This work addresses the problem of adapting pretrained encoders for generation tasks in natural language processing, offering a novel approach that is incremental but effective for multilingual applications.
The authors tackled the gap between pretrained encoders and natural language generation tasks by introducing DeltaLM, a multilingual encoder-decoder model that augments existing encoders with a decoder and pre-trains it using span corruption and translation span corruption, resulting in outperformance of strong baselines on tasks like machine translation and text summarization.
While pretrained encoders have achieved success in various natural language understanding (NLU) tasks, there is a gap between these pretrained encoders and natural language generation (NLG). NLG tasks are often based on the encoder-decoder framework, where the pretrained encoders can only benefit part of it. To reduce this gap, we introduce DeltaLM, a pretrained multilingual encoder-decoder model that regards the decoder as the task layer of off-the-shelf pretrained encoders. Specifically, we augment the pretrained multilingual encoder with a decoder and pre-train it in a self-supervised way. To take advantage of both the large-scale monolingual data and bilingual data, we adopt the span corruption and translation span corruption as the pre-training tasks. Experiments show that DeltaLM outperforms various strong baselines on both natural language generation and translation tasks, including machine translation, abstractive text summarization, data-to-text, and question generation. The code and pretrained models are available at \url{https://aka.ms/deltalm}.