How syntactic analysis influences the calculation of mean dependency distance: Evidence from the enhanced dependency representation

How syntactic analysis influences the calculation of mean dependency distance: Evidence from the enhanced dependency representation Tsy Yih Jianwei Yan Department of Linguistics Department of Linguistics Zhejiang University Zhejiang University yezi_leafy@hotmail.com yanjianwei@aliyun.com Haitao Liu✉ Department of Linguistics Zhejiang University lhtzju@gmail.com Abstract Based on Yan & Liu (2022), the present paper further explores how syntactic analysis, or the annotation scheme of dependency treebanks, affects mean dependency distance (MDD). By comparing the treebanks of 16 languages with both basic (BUD) and enhanced universal dependency (EUD) representations, we find that the MDD measured by the EUD representation is statistically larger than that of BUD. The main distinction between the two representations lies in the treatment of three constructions: coordinate structures, relative clauses, and pivotal constructions. However, a closer look at the data reveals that these three analyses in EUD do not necessarily have longer MDD in single sentences: The enhanced analysis of coordinate structures and pivotal constructions statistically have a significant contribution to the increase of MDD, while that of relative clauses contributes the least. We conclude with the factors that may affect the change of MDD, including both internal, structural ones and external ones, such as the context of the text (in the form of stochastically intervening dependents) and the language type (in the form of word order type). 1 Introduction Mean Dependency Distance (MDD), defined as the sum of all dependency distances divided by the number of dependency relations, is a measure based on the dependency structures of sentences. It has received much attention in the last two decades (Liu, 2008; Futrell et al., 2015; Jiang & Liu, 2015). Previous studies have shown a general tendency for natural languages to possess statistically smaller MDD than languages generated by a number of random baselines (Liu, 2008; Futrell et al., 2020). Therefore, it is generally considered to be a metric to reflect syntactic complexity and the limit of human memory, and many studies have attempted to explain its inner mechanism (Temperley, 2008; Gildea & Temperley, 2010; Liu et al., 2017), such as being the result of the Principle of Least Efforts (Zipf, 1949). MDD is known to be subject to many factors, such as language type (Liu, 2008), sentence length (Jiang & Liu, 2015), chunking (Lu et al., 2016), genre (Wang & Liu, 2017), annotation scheme (Yan & Liu, 2022), etc. Among all these factors, the annotation scheme differs from others in that it influences the observed value of the MDD, rather than the real value of the variable itself. An analogy is to measure the temperature with different scales, and one would have different values. For instance, a certain temperature might be measured to have the value 40 under degree Celsius, and show the value of 104 under degree Fahrenheit. Turning back to the linguistic issue here, an annotation scheme reflects the choice of syntactic analysis. Since there are various versions of syntactic structural analysis in the linguistic literature, it is thus also important to pay attention to such effect. As we …

Content truncated to 500 words