Towards a Language for Natural Language Treebank Transductions

Proceedings of the 27th International Conference on Computational Linguistics, pages 1022–1032 Santa Fe, New Mexico, USA, August 20-26, 2018. 1022 Towards a Language for Natural Language Treebank Transductions Carlos A. Prolo Department of Informatics and Applied Mathematics - DIMAp Federal University of Rio Grande do Norte - UFRN Natal, RN, 59078-970, Brazil prolo@dimap.ufrn.br Abstract This paper describes a transduction language suitable for natural language treebank transforma- tions and motivates its application to tasks that have been used and described in the literature. The language, which is the basis for a tree transduction tool allows for clean, precise and con- cise description of what has been very confusingly, ambiguously, and incompletely textually described in the literature also allowing easy non-hard-coded implementation. We also aim at getting feedback from the NLP community to eventually converge to a de facto standard for such transduction language. 1 Introduction Linguists have always liked to use trees to theorize about the structure of natural language sentences (Marcus et al., 1983). In the 1990’s, computer scientists also started to grow excited with the possibility of getting those proposed structures reasonably accurately from a computer. The advent of large corpora with annotated syntactic structures, the treebanks, among which the Penn Treebank (PTB) (Marcus et al., 1993; Marcus et al., 1994) is a notable representative, made it possible to build statistical parsers with an accuracy that was not feasible before. From the early work of (Magerman, 1994; Magerman, 1995; Charniak, 1997; Collins, 1997), state-of-the-art accuracy has been progressively raising, achieving now scores above 92% in the English Penn Treebank, as reported, for instance, in (Shindo et al., 2012). As for dependency representations (Nivre et al., 2016; Nivre and Fang, 2017), although they capture syntactic structure in a different way, they are still generally represented as trees, even if not necessarily caring for the order among the children of a node. Of course, not everybody is expected to agree on any given tree-structure style. On the linguistic side, each practitioner has their own theoretical conceptions on what is an adequate syntactic structure for a sentence, which may in turn differ from that of a specific treebank. Often all the information that is needed to promote the intended change is already present in the tree, either with some structure different than the desired, or “latent” in the annotation (Chiang and Bikel, 2002), in the form of secondary attributes on the nodes. When the change to the desired structure can be mechanically accomplished by applying to the trees a precisely defined rule, we call it a transduction. In order to perform more complex transformations we may instead define sequences of simpler transductions to be applied as a pipeline. A few tree transducers have been proposed in the literature, among them the ones described in (Chiang and Bikel, 2002), based on context-free style rules, (Blahetta, 2003), based on Richard Pitto’s TGrep (see (Rohde, 2005)), Tsurgeon (Levy and Andrew, 2006) and TTT (Purtee and Schubert, 2012). There are also transducers for other purposes, perhaps …

Content truncated to 500 words