Abstract:
The article outlines the principal challenges encountered in the automation of annotating implicit discourse relations, analyzes the underlying causes of these challenges, and suggests possible solutions. The article examines the main stages of the process: ($i$) the extraction of examples with implicit discourse relations; ($ii$) the delimitation of relational argument boundaries; and ($iii$) the selection of features for annotation of the extracted fragments. The results of applying the method of search with exclusion in parallel texts are presented along with a critical assessment of its limitations. Two factors significantly hindering the automation of argument identification in text spans with implicit discourse relations are analyzed: the considerable variability in argument length and the noncontiguous nature of arguments, which may be interrupted by intervening tokens. A comprehensive analysis of methods for automating feature selection for the linguistic data is provided. It has been demonstrated that even the processing of formal features may require the involvement of experts. Furthermore, while some semantic features are amenable to partial automation, others currently require manual annotation. The conclusions are illustrated by examples from the corpus.
Fulltext:
PDF file (169 kB)