Neural sentence simplification method based on sequence-to-sequence framework has become the mainstream method for sentence simplification (SS) task. Unfortunately, these methods are currently limited by the scarcity of parallel SS corpus. In this paper, we focus on how to reduce the dependence on parallel corpus by leveraging a careful initialization for neural SS methods from paraphrase corpus. Our work is motivated by the following two findings: (1) Paraphrase corpus includes a large proportion of sentence pairs belonging to SS corpus. (2) We can construct large-scale pseudo parallel SS data by keeping these sentence pairs with a higher complexity difference. Therefore, we propose two strategies to initialize neural SS methods using paraphrase corpus. We train three different neural SS methods with our initialization, which can obtain substantial improvements on the available WikiLarge data compared with themselves without initialization.

翻译：基于序列到序列框架的神经句子简化方法已成为句子简化任务的主流方法。然而，这些方法目前受到平行简化语料库稀缺的限制。本文聚焦于如何通过利用释义语料库对神经句子简化方法进行精心初始化，从而降低对平行语料库的依赖。我们的工作基于以下两个发现：（1）释义语料库中包含大量属于句子简化语料库的句对；（2）通过保留复杂度差异较大的句对，我们可以构建大规模的伪平行简化数据。因此，我们提出两种利用释义语料库初始化神经句子简化方法的策略。采用我们的初始化方法训练三种不同的神经句子简化模型，与未初始化的模型相比，可在现有WikiLarge数据集上获得显著提升。