Text matching is a fundamental and critical problem in natural language understanding (NLU), where multi-level semantics matching is the most challenging task. Human beings can always leverage their semantic knowledge, while neural computer systems first learn sentence semantic representations and then perform text matching based on learned representation. However, without sufficient semantic information, computer systems will not perform very well. To bridge the gap, we propose a novel Frame-based Multi-level Semantics Representation (FMSR) model, which utilizes frame knowledge to extract multi-level semantic information within sentences explicitly for the text matching task. Specifically, different from existing methods that only rely on the sophisticated architectures, FMSR model, which leverages both frame and frame elements in FrameNet, is designed to integrate multi-level semantic information with attention mechanisms to learn better sentence representations. Our extensive experimental results show that FMSR model performs better than the state-of-the-art technologies on two text matching tasks.

文本匹配是自然语言理解 (NLU) 中的一个基本且关键的问题，其中多级语义匹配是最具挑战性的任务。人类总是可以利用他们的语义知识，而神经计算机系统首先学习句子语义表示，然后根据学习到的表示进行文本匹配。然而，如果没有足够的语义信息，计算机系统将不会表现得很好。为了弥补差距，我们提出了一种新颖的基于框架的多级语义表示(FMSR) 模型，它利用框架知识为文本匹配任务显式地提取句子中的多级语义信息。具体而言，与仅依赖复杂架构的现有方法不同，FMSR 模型利用 FrameNet 中的框架和框架元素，旨在将多级语义信息与注意机制相结合，以学习更好的句子表示。我们广泛的实验结果表明，FMSR 模型在两个文本匹配任务上的表现优于最先进的技术。