Sentence semantic matching (SSM) always plays a critical role in natural language processing. Measuring the intrinsic semantic similarity among sentences is very challenging and has not been substantially addressed. The latest SSM research usually relies on a shallow text representation and interaction between sentence pairs, which might not be enough to capture the complex semantic features and lead to limited performance. To capture more semantic context features and interactions, we propose a hierarchical encoding model (HEM) for sentence representation, further enhanced by a hierarchical matching mechanism for sentence interaction. Given two sentences, HEM generates intermediate and final representations in encoding layer, which are further handled by a novel hierarchical matching mechanism to capture more multi-view interactions in matching layer. The comprehensive experiments demonstrate that our model is capable to capture more sentence semantic features and interactions, which significantly outperforms the existing state-of-the-art neural models on the public real-world dataset.

句子语义匹配（SSM）在自然语言处理中始终扮演着至关重要的角色。测量句子之间的内在语义相似性非常具有挑战性，尚未得到实质性解决。最新的SSM研究通常依赖于浅层的文本表示和句子对之间的交互，这可能不足以捕获复杂的语义特征并导致性能受限。为了捕获更多语义上下文特征和交互作用，我们提出了一种用于句子表示的分层编码模型（HEM），并通过用于句子交互作用的分层匹配机制进一步增强了该结构。给定两个句子，HEM在编码层生成中间和最终表示，新颖的分层匹配机制进一步处理了这些问题，以在匹配层中捕获更多的多视图交互。全面的实验表明，我们的模型能够捕获更多的句子语义特征和交互作用，大大优于公共现实世界数据集上现有的最新神经模型。