CAOSS and Transcendence – On the Representation and Processing of Noun Compounds

名词复合词，如键位置，由两个成分组成：中心名词(这里：位置)和修饰语(这里：键)。然而，根据形态先验假说，这些成分的意义并不直接影响名词复合词的心理表征和认知加工。相反，假设这些成分根据它们在复合词(修饰语或中心语)中的位置，在更常用的情况下形成改变的、位置约束的表示。例如，通过考虑诸如键位置或键时刻之类的情况，可以清楚地看出，key-如果用作修饰语，其含义与key的自由词含义不同。形态超越性假说认为，这些位置约束表征在复合词的表征或加工中起着重要作用。对于高度流行的修饰语，如德语的Haupt-(head-)，甚至可以观察到，这些修饰语实际上失去了自由词义，起到了类似词缀的作用。这可以认为是修饰语在表示形式上的进一步变化和超越，超越了位置限制的表示，变成了词缀。利用分布语义模型的组合方法，这些不同的表示可以直接量化。在分布语义模型中，词义被概念化为高维数值向量。这些向量是从自然语言的大型语料库中的单词分布中获得的。对于成分模型，例如CAOSS模型，名词复合词的向量表示可以从其成分的向量表示计算出来。关键是，在CAOSS模型中，作为中间步骤计算成分的位置约束向量表示，然后将其相加以获得复合向量表示。有了这些手边成分的位置约束向量表示，根据形态超越假设的一系列重要预测可以在一系列计算和实验研究中得到经验性的检验，如项目描述中详细描述的。另一种成分模型，功能方法，将修饰符概念化为线性函数，应用于头部向量表示来计算复合向量表示。这种函数方法可以成功地用作表示词缀的模型(例如，通过将错误的线性函数应用于财富的矢量表示来计算不幸的含义)。在该项目的第二部分，计划进行一系列的计算和实验研究，以检验这种功能方法是否是获得复合意义的更合适的模式，因为修饰语的使用更频繁，并形成了词缀式的表征。

Noun compounds, such as key position, consist of two constituents: the head noun (here: position) and the modifier (here: key). However, according to the morphological transcendence hypothesis, the meaning of these constituents does not directly play a role in the mental representation and cognitive processing of noun compounds. It is rather assumed that these constituent develop altered, positionally bound representations as they are more commonly used, depending on their position in the compound (modifier or head). By considering cases such as key position or key moment, for example, it is clear that the meaning of key-, if used as a modifier, differs from the free word meaning of key. The morphological transcendence hypothesis assumes that these positionally bound representations play an important role in the representation or processing of compounds. For highly prevalent modifiers, such as the German Haupt- (head-), it is even observed that these modifiers virtually lose their free word meanings and play an affix-like role. This can be considered as a further change and transcendence in the representation of the modifiers, beyond positionally bound representations and into affixes.Using compositional methods for models of distributional semantics, these different representations can be quantified directly. In distributional semantics models, word meanings are conceptualized as high-dimensional numerical vectors. These vectors are obtained from word distributions in large corpora of natural language. With compositional models, such as the CAOSS model, vector representations for noun compounds can be computed from the vector representations for their constituents. Crucially, in the CAOSS model, positionally bound vector representations for the constituents are computed as an intermediate step, which are then added up to obtain the compound vector representation. With these positionally bound vector representations for the constituents at hand, a series of important predictions that follow from the morphological transcendence hypothesis can be examined empirically in a series of computational and experimental studies, as described in detail in the Project Description.Another compositional model, the functional approach, conceptualizes the modifier as a linear function that is applied to the head vector representation to compute a compound vector representation. Such a functional approach can be employed successfully as a model for representing affixed words (for example, the meaning of misfortune is computed by applying the linear function for mis- to the vector representation for fortune). In the second part of the project proposed here, a series of computational and experimental studies is planned to examine whether such a functional approach is a more adequate model for obtaining compound meanings as a modifier is more frequently used and develops an affix-like representation.