📄 BCCWJ-DepPara: A Syntactic Annotation Treebank on the ‘Balanced Corpus of Contemporary Written Japanese'

Summary

In Japanese NLP, syntax is usually described as dependencies between bunsetsu, i.e., base phrasal units. However, coordinate constructions fit poorly into a dependency tree. The authors propose a new annotation scheme that moves coordination and apposition outside the tree. For this purpose it uses "segments" (spans of a word sequence) and "groups" (sets of segments), which form a separate layer above ordinary dependencies. This makes it possible to describe nested coordination, coordination of conjuncts that do not form a constituent, and forward sharing, where a single left dependent relates to all conjuncts. The tree structure is not distorted in the process. The scheme also defines attachment rules for subordinate clauses based on Minami's classes. In addition, it has labels for correcting errors in earlier layers, in particular wrong bunsetsu and sentence boundaries. The annotation was applied to the core data of the BCCWJ corpus: about a million words in 1980 samples from six registers. Over seven years it was manually checked three times. The result is the BCCWJ-DepPara treebank, which the authors compare with the KC and CSJ schemes. It is a genre-balanced syntactic resource for Japanese in which parataxis is separated from dependencies.

Key points

• Arcs have only four labels: D — an ordinary dependency; B — merging into a longer bunsetsu; F — no dependency, e.g., for fillers or emoticons; Z — end of sentence, with nested sentences allowed. Non-projective and right-to-left arcs are not marked separately and are also labeled D.
• There are three kinds of groups: Parallel (coordination), Apposition (apposition in the narrow sense) and Generic (apposition in the broad sense). Segment boundaries are drawn at the level of short unit words, so they may fall inside a bunsetsu. Predicate coordination is not marked, whereas KC/CSJ have the label P for it.
• The data are the BCCWJ core data from six registers: newspapers (PN), books (PB), magazines (PM), government reports, the so-called white papers (OW), Yahoo! Answers (OC) and Yahoo! Blogs (OY). The size is given as ~1 million words in the abstract and as ~1.2 million in the introduction.
• Statistics:
  • a coordinate group has 2.19–2.35 members on average, an appositive group about 2.00;
  • Parallel groups are most frequent in the OW register;
  • the F label is most frequent in OW because of list item markers and in OY because of emoticons.
• The attachment of subordinate clauses is determined by Minami's three classes (A, B, C). For example, the topicalizer -ha is allowed only inside class C clauses. The classes themselves are not yet annotated in the corpus. The data are available to owners of the BCCWJ DVD edition. A parser that predicts both dependencies and coordination via dual decomposition was proposed by Iwatate (2012).