Polar Quantification of Actor Noun Phrases for German

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1376–1380 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1376 Polar Quantification of Actor Noun Phrases for German Anne G¨ohring, Manfred Klenner Department of Computational Linguistics University of Zurich {goehring,klenner}@cl.uzh.ch Abstract In this paper, we discuss work that strives to measure the degree of negativity - the negative polar load - of noun phrases, especially those denoting actors. Since no gold standard data is available for German for this quantification task, we generated a silver standard and used it to fine-tune a BERT-based intensity regressor. We evaluated the quality of the silver standard empirically and found that our lexicon-based quantification metric showed a strong correlation with human annotators. Keywords: polarity intensity, quantification metric, BERT-based regressor 1. Introduction The polar intensity of words, noun phrases, sentences and texts has been focused on for quite some time, e.g. Polanyi and Zaenen (2006),Taboada et al. (2011) and more recently Huang et al. (2020). In contrast to pre- vious approaches, that are just interested in a proper prediction of the intensity, we also are interested in the discourse function of a particular type of noun phrases, namely noun phrases that could be used to refer to an agent of some action, an actor. Besides our theoret- ical interest in such an exploration, there is also an application-oriented aspect. One of the strategies of hate speech is defamation and vilification. Verbally, we could use various means to conceptualize somebody as a negative person, a negative actor. Among them are definite descriptions (‘the foolish Merkel’), predicative statements (‘Merkel is a fool’), or role fillers of par- ticular verbs (‘Merkel cheats us’). If a text casts an actor as highly negative, then it might be an instance of hate speech. Our quantification approach, thus, could be used to identify candidates of hate speech. Since no gold standard data are available for German, we generated a lexicon-based silver standard (Manfred Klenner, Anne G¨ohring, 2022) and carefully evaluated the resulting data set. Although the lexicon-based met- ric turned out to strongly correlate with human anno- tations, the intention was to get rid of the need for lexicon-based modelling, since this is limited due to lexicon gaps. The research question then was: how well does the learned model generalize. The main contribution of this work is: we introduce the first approach to quantification (possibly not only for German) where a general model is derived in a boot- strapping manner from existing lexical resources. 2. Actor Noun Phrase Quantification For the detection of hate speech the identification of highly negatively conceptualized actors could be use- ful. Thus, a regression model for the quantification of negativity (of actor noun phrases) is needed. Cur- rently, there is no gold standard with quantified Ger- man noun phrases available. The annotation of nega- tive noun phrases with concrete strength values might turn out to be challenging. actor noun phrase 1 die l¨ugnerische Merkel (mendacious) 2 die sturre …

Content truncated to 500 words