Analysis of Coordinating Constructions in a Dependency Treebank

This paper summarizes results of automatic analysis of coordinating constructions and appositions in the Prague Dependency Treebank using a method of analysis by reduction. Experiments are performed on a large subset of the treebank. This subset is obtained as a result of a query providing a set of more than 4,300 suitable sentences and their tree structures containing coordinations and appositions. The automatic procedure is complemented by a manual analysis of reasons why certain sentences (trees) were not fully reduced. This analysis helps to gain a better insight into the phenomena of coordination and apposition and their formal properties. Dependency trees have a long tradition in linguistics, especially in the description of Slavic languages of Central and Eastern Europe. Although the history of linguistics witnessed many heated discussions between the followers of the tradition of constituent trees and the linguists adoring dependency trees (both sides usually unable to persuade the opponents about the advantages of their type of trees), it seems that the dependency notation has recently been recognized as an efficient and transparent data type for the description of syntactic relations in treebanks for a number of languages. One of the dependency treebanks which became quite popular among linguists due to the thoroughness of its annotation is the Prague Dependency Treebank (PDT), see (Bejček et al. 2013), the corpus exploited in this paper. One of the difficulties faced by the dependency notation is the necessity to express within a dependency tree not only dependencies, but also relations which are naturally not of a dependency nature. This is a well-known issue, a member of the Prague Linguistic Circle, Lucien Tesnière (Tesnière 1959) already distinguished between structural relations which we nowadays call as dependency (‘connexion’), and between the coordinating relationships (‘junction’).1 In this paper we study primarily the relationships of coordination and apposition, which pose a great challenge to any dependency formalism. The analysis is performed Copyright c © 2015, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. However, such conception is not accepted without reservations – there are influential approaches which capture coordination as a (type of) a dependency relation, see esp. (Mel’čuk 1988). by means of analysis by reduction (Lopatková, Plátek, and Kuboň 2005; Lopatková, Plátek, and Sgall 2007), a procedure which naturally defines governing and dependent words in a dependency relationship; we enrich the procedure to capture also relations of coordination and apposition. We have applied the analysis by reduction automatically to dependency trees of selected sentences from the Prague Dependency Treebank. The results obtained by this method were then manually analyzed. This approach actually brings two kinds of results – it helps to gain better insight into the problem of coordination and its annotation in the corpus, and, on top of that, it also helps to identify potential annotation inconsistencies in the corpus. The theory we adhere to in our investigations, dependency-based Functional Generative Description (FGD), is described primarily in (Sgall, Hajičová, and Panevová 1986). Analysis by Reduction The original method of analysis by reduction (AR) makes it possible to formulate the relationship between dependency and word order (Lopatková, Plátek, and Kuboň 2005). This approach is beneficial especially for modeling the syntactic structure of languages with a high degree of free word order, where the dependency structure and word order are only loosely related. Let us now describe the ideas behind the method used for sentence analysis. Analysis by reduction is based on a stepwise simplification of an analyzed sentence. It defines possible sequences of reductions (deletions) in the sentence – each step of AR is represented by deleting of at least one word of the input sentence; in specific cases, deleting is accompanied by a shift of a word form to different word order position. Let us stress the basic constraints imposed on the analysis by reduction, namely: (i) the obvious constraint on preserving individual word forms, their morphological characteristics and/or their surface dependency relations, and (ii) the constraint on preserving the correctness (a grammatically correct sentence must remain correct after its simplification). The basic principles of AR can be illustrated by the following Czech sentence (1).