Syntactic similarity of the sentences in a multi-lingual parallel corpus based on the Euclidean distance of their dependency trees

Syntactic similarity of the sentences in a multi-lingual parallel corpus based on the Euclidean distance of their dependency trees Masanori Oya Meiji University masanori_oya2019@meiji.ac.jp Abstract This study proposes the idea that the difference between the syntactic structures of a sentence and its translation pair in another language can be numerically represented by their Euclidean distance, calculated on the basis of the degree centralities and closeness centralities of the syntactic dependency trees of the sentences. The mean distances thus calculated for a set of translation pairs of two languages can be used as a measure of the similarity/difference between these two languages. A corpus analysis using a multi- lingual parallel corpus reveals that mean Euclidean distances thus calculated seem to reflect the typological tendencies of the differences between languages within and between language families. 1 Introduction Sentence similarity measuring has recently attracted the attention of many researchers because it is required for various natural language applications, such as those related to question answering (De Boni and Manandhar, 2003), plagiarism detection (Alzahrani et al., 2012), and semantic searching (Farouk et al., 2018). Sentence similarity measuring has been conducted through a variety of techniques, yet the majority of them emphasize lexico-semantic similarity between sentences. Syntactic similarity measures in this context are typically employed only to augment the accuracy of semantic similarity measurement (e.g., Batanovic and Bojic, 2015; Lee et al., 2014; Ma and Suel, 2016), possibly based on the observation that the same meaning can be expressed by a variety of sentences with different syntactic structures and, inversely, that sentences with the same syntactic structure but different words can have completely different meanings. This one-to- many correspondence between sentential meaning and syntactic structures situates lexical or semantic similarity between sentences as equivalent to sentence similarity and syntactic similarity as playing a secondary, supplemental role in sentence similarity. In spite of the success of sentence similarity measuring, its trend with an emphasis on lexico- semantic similarity should not distract us from investigating the purely syntactic similarities or differences between sentences, which remains worthy of extensive linguistics research with high quality data. The most appropriate data for this purpose is multilingual parallel corpora that consist of a large number of sentences and their translations in several languages. Since these translation pairs are semantically equivalent and the lexico-semantic differences between them are somewhat controlled, we can focus on their syntactic similarities/differences. In analyzing syntactic similarities/differences between translation pairs in a multilingual parallel corpus, it is important to take a quantitative approach for the following reasons. First, quantifying the syntactic similarity of a given translation pair of two languages included in a multilingual parallel corpus allows it to represent the pure syntactic similarity of the translation pair, and thus to be applied to the types of natural language processing applications mentioned above as an auxiliary measure for sentence similarity. Second, and more importantly, many existing similarity analyses were subjectively conducted by individual researchers, such as those focused on the differences and similarities in the syntactic structures of …

Content truncated to 500 words