To what extent does sentence-internal realisation reflect discourse context? A study on word order

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 767–776, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics To what extent does sentence-internal realisation reflect discourse context? A study on word order Sina Zarrieß Jonas Kuhn Institut f¨ur maschinelle Sprachverarbeitung University of Stuttgart, Germany zarriesa,jonas@ims.uni-stuttgart.de Aoife Cahill Educational Testing Service Princeton, NJ 08541, USA acahill@ets.org Abstract We compare the impact of sentence- internal vs. sentence-external features on word order prediction in two generation settings: starting out from a discrimina- tive surface realisation ranking model for an LFG grammar of German, we enrich the feature set with lexical chain features from the discourse context which can be robustly detected and reflect rough gram- matical correlates of notions from theoreti- cal approaches to discourse coherence. In a more controlled setting, we develop a con- stituent ordering classifier that is trained on a German treebank with gold corefer- ence annotation. Surprisingly, in both set- tings, the sentence-external features per- form poorly compared to the sentence- internal ones, and do not improve over a baseline model capturing the syntactic functions of the constituents. 1 Introduction The task of surface realization, especially in a rel- atively free word order language like German, is only partially determined by hard syntactic con- straints. The space of alternative realizations that are strictly speaking grammatical is typically con- siderable. Nevertheless, for any given choice of lexical items and prior discourse context, only a few realizations will come across as natural and will contribute to a coherent text. Hence, any NLP application involving a non-trivial generation step is confronted with the issue of soft constraints on grammatical alternatives in one way or another. There are countless approaches to modelling these soft constraints, taking into account their interaction with various aspects of the discourse context (givenness or salience of particular refer- ents, prior mentioning of particular concepts). Since so many factors are involved and there is further interaction with subtle semantic and prag- matic differentiations, lexical choice, stylistics and presumably processing factors, theoretical ac- counts making reliable predictions for real cor- pus examples have for a long time proven elusive. As for German, only quite recently, a number of corpus-based studies (Filippova and Strube, 2007; Speyer, 2005; Dipper and Zinsmeister, 2009) have made some good progress towards a coherence- oriented account of at least the left edge of the German clause structure, the Vorfeld constituent. What makes the technological application of theoretical insights even harder is that for most relevant factors, automatic recognition cannot be performed with high accuracy (e.g., a coreference accuracy in the 70’s means there is a good deal of noise) and for the higher-level notions such as the information-structural focus, interannotator agreement on real corpus data tends to be much lower than for core-grammatical notions (Poesio and Artstein, 2005; Ritz et al., 2008). On the other hand, many of the relevant dis- course factors are reflected indirectly in proper- ties of the sentence-internal material. Most no- …

Content truncated to 500 words