Zero-shot domain paraphrase with unaligned pre-trained language models

Abstract

AbstractAutomatic paraphrase generation is an essential task of natural language processing. However, due to the scarcity of paraphrase corpus in many languages, Chinese, for example, generating high-quality paraphrases in these languages is still challenging. Especially in domain paraphrasing, it is even more difficult to obtain in-domain paraphrase sentence pairs. In this paper, we propose a novel approach for domain-specific paraphrase generation in a zero-shot fashion. Our approach is based on a sequence-to-sequence architecture. The encoder uses a pre-trained multilingual autoencoder model, and the decoder uses a pre-trained monolingual autoregressive model. Because these two models are pre-trained separately, they have different representations for the same token. Thus, we call them unaligned pre-trained language models. We train the sequence-to-sequence model with an English-to-Chinese machine translation corpus. Then, by inputting a Chinese sentence into this model, it could surprisingly generate fluent and diverse Chinese paraphrases. Since the unaligned pre-trained language models have inconsistent understandings of the Chinese language, we believe that the Chinese paraphrasing is actually performed in a Chinese-to-Chinese translation manner. In addition, we collect a small-scale English-to-Chinese machine translation corpus in the domain of computer science. By fine-tuning with this domain-specific corpus, our model shows an excellent capability of domain-paraphrasing. Experiment results show that our approach significantly outperforms previous baselines regarding Relevance, Fluency, and Diversity.