Abstract Much of the patterned use of language occupies a poorly charted middle ground of usages that are neither frozen, one-off items listable in dictionaries nor products of maximally general rules found in grammars. Similarly, these usages fly below the radar of modular theories of language that make a strict distinction between items in a lexicon and the rules of syntax for combining them. Early constructionist approaches showed this neglected territory to be teeming with conventional form-meaning pairings, i.e., lexico-grammatical constructions. While corpora have been seen as a source for investigating these constructions, they entail a fundamental but seldom-noted limitation: constructions are constituted by both syntagmatic and paradigmatic relations, but corpora lack the paradigmatic dimension. Thus corpora can reveal multiword items but not the relations among them that constitute constructions. We elaborate on an alternative, illustrating how the design of an existing machine-readable language model affords discovery of lexico-grammatical constructions by capturing bottom up the relations they contract with other patterns. The network is noise-ridden, but we exploit the noise as the necessary background against which constructions can be set into relief and thus made discoverable.

中文翻译：

---

构造与发现问题：以范式为例

摘要语言的许多模式使用都占据了图表使用的中间位置，它们既不是冻结的，一次性可在词典中列出的项目，也不是语法中存在的最大通用规则的产品。类似地，这些用法在模块化语言理论的雷达之下飞速发展，该语言理论严格区分了词典中的项目和组合它们的语法规则。早期的建构主义方法表明，这一被忽视的领域充满了传统的形式意义配对，即词汇语法构造。尽管语料库已被视为研究这些结构的来源，但它们却具有一个基本但鲜为人知的局限性：结构是由同义词关系和范式关系构成的，但是语料库缺乏范式维度。因此，语料库可以揭示多词项目，但不能揭示构成结构的它们之间的关系。我们详细介绍了另一种方法，它说明了现有的机器可读语言模型的设计如何通过捕获自底向上的它们与其他模式的联系来提供词汇语法构造的发现。该网络充满了噪音，但是我们将噪音作为必要的背景，可以根据需要对结构进行调整，从而使其易于发现。