Three Types of Average Dependency Distances of Sentences in a Multilingual Parallel Corpus

Three Types of Average Dependency Distances of Sentences in a Multilingual Parallel Corpus Masanori Oya Meiji University Masanori_oya2019@meiji.ac.jp Abstract This paper is an exploration of three types of dependency distances of the sentences in a multilingual parallel corpus, in order to verify the expectation that we can obtain an objective, quantitative measure to indicate cross-linguistic variation of the syntactic- structural setting of human languages. The results indicated that pair-wise average dependency distances seem to categorize languages into several groups, and type- wise average dependency distances seem to provide us with fine-grained quantification of syntactic properties of individual natural languages. 1. Introduction Dependency distance is the linear distance between two words in dependency relationship within the same sentence. It has been investigated as one of the important measures of memory burden and syntactic complexity (Gibson, 1998, 2000; Gildea and Temperley, 2010; Grodner and Gibson, 2005; Liu 2007, 2008; Liu et al., 2017; Oya, 2013). For example, in the sentence “David read three articles yesterday,” the noun David depends on read as the subject, and the dependency distance between them is 1, and the noun articles depends on read as the object, and the dependency distance between them is 2 (the direction of dependency is ignored here, and dependency distances are given as absolute values). In the example sentence above, the dependency distance between the subject and the verb is shorter than that between the object and the verb. Investigations of dependency distances have been conducted from the viewpoint of finding out the general properties of human languages in general, and it has been pointed out that natural languages show a certain preference for shorter dependency distances; for example, Gibson (2000) proposes the Dependency Locality Theory, which basically states that the preference for shorter dependency distances in natural languages is due to the limit of short-term memory during the integration of words into a larger structure. Liu (2008) shows that, based on the corpus-based investigation of 20 languages, there is a threshold of dependency distances which is 4, and the average dependency distance of these natural languages is shorter than that of an artificial language in which the dependency relationships of words are randomized. Futrell et al. (2015) also reported the similar result based on the corpus data of 37 natural languages. Provided that shorter dependency distances are preferred across different languages and they are below a certain threshold which has been found in previous research, we can assume that it is worth searching for cross-linguistic variations of dependency distances, not only by focusing on the dependency distances in a language as a whole, but also by focusing on the difference of average dependency distances of language pairs, and also on the dependency distances of different dependency types, such as the dependency distance between the object and the verb which is longer than that between the subject and the verb, as shown in the example above. By doing this, it is expected that we can obtain an objective, quantitative measure to indicate …

Content truncated to 500 words