Creation of a Balanced State-of-the-Art Multilayer Corpus for NLU

Creation of a Balanced State-of-the-Art Multilayer Corpus for NLU Normunds Gruzitis, Lauma Pretkalnina, Baiba Saulite, Laura Rituma, Gunta Nespore-Berzkalne, Arturs Znotins and Peteris Paikens University of Latvia, Institute of Mathematics and Computer Science, Raina blvd. 29, Riga, Latvia Latvian Information Agency LETA, Marijas street 2, Riga, Latvia {normunds.gruzitis,lauma.pretkalnina,baiba.valkovska}@lumii.lv, {peteris.paikens}@leta.lv Abstract This paper presents a work in progress to create a multilayered syntactically and semantically annotated text corpus for Latvian. The broad application area we address is natural language understanding (NLU), while more specific applications are abstractive text summarization and knowledge base population, which are required by the project industrial partner, Latvian information agency LETA, for the automation of various media monitoring processes. Both the multilayered corpus and the downstream applications are anchored in cross-lingual state-of-the-art representations: Universal Dependencies (UD), FrameNet, PropBank and Abstract Meaning Representation (AMR). In this paper, we particularly focus on the consecutive annotation of the treebank and framebank layers. We also draw links to the ultimate AMR layer and the auxiliary named entity and coreference annotation layers. Since we are aiming at a medium-sized still general-purpose corpus for a less-resourced language, an important aspect we consider is the variety and balance of the corpus in terms of genres, authors and lexical units. Keywords: UD, FrameNet, PropBank, AMR, Latvian 1. Introduction Natural language understanding (NLU) systems for ab- stractive text summarization, knowledge base population, and many other tasks rely, explicitly or implicitly, on full- stack syntactic and semantic parsing, including semantic role labeling, named entity recognition and linking, and coreference resolution. State-of-the-art parsers, in turn, rely on supervised machine learning which requires substantial language resources – syntactically and semantically anno- tated text corpora and extensive linked lexicons. In the industry-oriented research project “Full Stack of Lan- guage Resources for Natural Language Understanding and Generation in Latvian”, we are creating a balanced text cor- pus with multilayer annotations, adopting widely acknowl- edged and cross-lingually applicable representations: Uni- versal Dependencies (UD) (Nivre et al., 2016), FrameNet (Fillmore et al., 2003), PropBank (Palmer et al., 2005) and Abstract Meaning Representation (AMR) (Banarescu et al., 2013). Figure 1 outlines the inter-layer relationships. First, the UD representation is automatically derived from a more elabo- rated manually annotated hybrid dependency-constituency representation (see Section 3.). This also ensures that para- graphs, sentences and tokens are correctly and uniformly split, and represented in the standard CoNLL-U data for- mat.1 Thus, all the annotation layers can be afterwards merged based on the document, paragraph, sentence and token identifiers. Second, the FrameNet annotations are manually added, guided by the underlying UD annotations (see Section 5.1.). Third, the PropBank layer is automati- cally derived from the FrameNet and UD layers (see Sec- tion 5.2.). Fourth, the semi-automatic annotation of named entities, as well as named entity linking, is done in paral- lel to and independently from the annotation of semantic frames (see Section 4.). Coreference annotations are added 1http://universaldependencies.org/format.html afterwards, on top of the named entity annotations, consult- ing the underlying UD tree if necessary. These two auxil- iary …

Content truncated to 500 words