Animacy Denoting German Nouns: Annotation and Classification

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1360–1364 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1360 Animacy Denoting German Nouns: Annotation and Classification Manfred Klenner & Anne G¨ohring Department of Computational Linguistics University of Zurich, Switzerland {klenner,goehring}@cl.uzh.ch Abstract In this paper, we introduce a gold standard for animacy detection comprising almost 14,500 German nouns that might be used to denote either animate entities or non-animate entities. We present inter-annotator agreement of our crowd-sourced seed annotations (9,000 nouns) and discuss the results of machine learning models applied to this data. Keywords: animacy detection, metonymy trigger, GermaNet 1. Introduction Animacy detection is meant to automatically distin- guish words which denote humans (or animals) from words used to denote non-humans (or non-animals). Reference to humans comprises among others the us- age of proper names (Anne), profession names (minis- ter), institution names (government), company names (Google), (capital) cities (Washington) and nation names (France). Having casted the task like this, metonymy detection is claimed to be part of it. metonymy is defined as reference to some entity by re-A ferring to another entity which is strongly related to it. In Washington cheats the world, Washington is under- stood as reference to the American government (not the city), i.e. a group of people with a particular po- litical function. Our ultimate goal is to detect expres- sions in a text that might be used to cast humans as very negative actors (like Washington in the example given). A logical first step in such a model is to identify animacy denoting nouns (incl. particular metonymy triggering words). Our research hypothesis is that the classification of these words can be learned solely on the basis of the embeddings of a medium sized gold standard. The gold standard must comprise both ani- macy and non-animacy denoting nouns. We have cre- ated such a gold standard on the basis of 15,000 can- didate nouns that come from the German wordnet Ger- maNet (Hamp and Feldweg, 1997), (Henrich and Hin- richs, 2010) and an inhouse newspaper corpus. We de- termined inter-annotator agreement and evaluated it in a cross-validation setting. Named entity recognition (NER) is a subtask of ani- macy detection. Person, nation and city names are cru- cial here. We exclude NER from our current investiga- tions, since we use an existing NER in our pipeline. 2. Related Work There seems to exist no approaches to animacy classifi- cation for German. The difference of our work to exist- ing approaches for English is that we use only the word embeddings of nouns and no other features. Since there is no gold standard for German where nouns or phrases in their context are annotated, our generic context ag- nostic approach is at least a feasible solution. Existing models use morpho-syntactic features or rely on lexical and semantic resources like WordNet. Some- times named entity recognition systems support the an- imacy detection. Bowman and Chopra (2012) build a classifier on three …

Content truncated to 500 words