Moving TIGER beyond Sentence-Level

Moving TIGER beyond Sentence-Level Agnieszka Falenska, Kerstin Eckart, Jonas Kuhn Institute for Natural Language Processing University of Stuttgart {falenska, eckartkn, jonas}@ims.uni-stuttgart.de Abstract We present TIGER 2.2-doc – a new set of annotations for the German TIGER corpus. The set moves the corpus to a document level. It includes a full mapping of sentences to documents, as well as additional sentence-level and document-level annotations. The sentence- level annotations refer to the role of a sentence in the document. They introduce structure to the TIGER documents by separating headers and meta-level information from article content. Document-level annotations recover information which has been neglected in the intermediate releases of the TIGER corpus, such as document categories and publication dates of the articles. Additionally, we introduce new document-level annotations: authors and their gender. We describe the process of corpus annotation, show statistics of the obtained data and present baseline experiments for lemmatization, part-of-speech and morphological tagging, and dependency parsing. Finally, we present two example use cases: sentence boundary detection and authorship attribution. These use cases take the data from TIGER into account and illustrate the usefulness of the new annotation layers from TIGER 2.2-doc. Keywords: Corpus Annotation, Treebank, Document Structure 1. Introduction The TIGER corpus (Brants et al., 2004) is an important treebank for German. It has been frequently employed over the years: for several shared tasks on multilingual de- pendency parsing (Buchholz and Marsi, 2006) and pars- ing morphologically rich languages (Seddah et al., 2013; Seddah et al., 2014); as part of the HamleDT compilation (Zeman et al., 2012); and as training data for processing pipelines for variety of tasks, including coreference and bridging resolution (Björkelund et al., 2014), theoretical linguistics (Haselbach et al., 2012), and building morpho- logical word-embeddings (Cotterell and Schütze, 2015). TIGER consists of roughly 900,000 words or 50,000 sen- tences from German newspaper articles and its focus is on token and sentence level: For the first release of the TIGER treebank, the corpus was semi-automatically annotated and corrected with part-of-speech tags and constituent struc- tures. The following versions enhanced the corpus with respect to the number of sentences and introduced morpho- logical and lemma annotations. Additionally, parts of the corpus were released as dependency treebanks, e.g. a de- pendency version of TIGER 2.2 (Seeker and Kuhn, 2012). During the early development of the corpus parts of the in- formation about document boundaries were lost. The cur- rent distribution 2.2 does not contain any grouping of sen- tences into documents, i.e., the corpus consists of one long sequence of sentences. Therefore, it is not possible to anno- tate this dataset with any document-level information, such as anaphora and coreference relations, or discourse trees.1 However, as opposed to datasets where contiguous docu- ments are not available due to technical design or copyright reasons (Faaß and Eckart, 2013; Schäfer, 2015), TIGER does contain full documents and the only information lost is a mapping between sentences and documents to which they belong. 1Cf. guidelines like Riester and Baumann (2017) and Riester et al. …

Content truncated to 500 words