A Lightweight Modeling Middleware for Corpus Processing

A Lightweight Modeling Middleware for Corpus Processing Markus G¨artner Jonas Kuhn Institute for Natural Language Processing University of Stuttgart {markus.gaertner,jonas.kuhn}@uni-stuttgart.de Abstract Present-day empirical research in computational or theoretical linguistics has at its disposal an enormous wealth in the form of richly annotated and diverse corpus resources. Especially the points of contact between modalities are areas of exciting new research. However, progress in those areas in particular suffers from poor coverage in terms of visualization or query systems. Many limitations for such tools stem from the non-uniform representations of very diverse resources and the lack of standards that address this problem from the perspective of processing or querying. In this paper we present our framework for modeling arbitrary multi-modal corpus resources in a unified form for processing tools. It serves as a middleware system and combines the expressiveness of general graph-based models with a rich metadata schema to preserve linguistic specificity. By separating data structures and their linguistic interpretations, it assists tools on top of it so that they can in turn allow their users to more efficiently exploit corpus resources. Keywords: data model, multi-modal, metadata 1. Introduction The availability of richly annotated multi-layer corpora is one of the cornerstones of modern theoretical and compu- tational linguistics. With a steady flow of newly created corpus resources and tools to annotate or process those cor- pora comes also the need for exploratory support tools, es- pecially visualization and query systems, to get a better un- derstanding of the resources at hand. The context of this contribution is the successor project of the ICARUS (G¨artner et al., 2013) platform, an interactive visualization and query tool for corpora with dependency syntax, coreference structures (G¨artner et al., 2014) or rich annotations for intonation (G¨artner et al., 2015). This project aims at providing researchers a platform to explore and query even more diverse multi-layer and multi-modal corpus resources and artifacts through a single interactive interface. Our target resources include, but are not limited to, text and speech data annotated for a variety of different layers on varying linguistic levels and granularities, such as the SFB732 Silver Standard Collection, of which a first overview has been provided by Eckart and G¨artner (2016). A strong focus lies hereby on connecting linguistic layers or modalities which so far have not received much attention in their particular combination. This is to support interest- ing research questions, such as the interaction of informa- tion status and prosodic realization(Baumann and Riester, 2013) or the incorporation of prosodic information into the task of automatic coreference resolution on spoken text as done by Roesiger and Riester (2015). As a result we faced the challenge of modeling access from a single processing software to a very diverse set of cor- pus resources. And while several standards around the Lin- guistic Annotation Framework (LAF)(Ide and Suderman, 2014) exist, they primarily take the point of view of prepa- ration or curation of data. The ongoing standardization ef- fort for a Corpus Query Lingua Franca (CQLF, ISO/DIS 24623-1)(Banski …

Content truncated to 500 words