Automatically Generated Affective Norms of Abstractness, Arousal, Imageability and Valence for 350 000 German Lemmas

Automatically Generated Affective Norms of Abstractness, Arousal, Imageability and Valence for 350 000 German Lemmas Maximilian K¨oper Sabine Schulte im Walde Institut f¨ur Maschinelle Sprachverarbeitung Universit¨at Stuttgart, Germany {maximilian.koeper,schulte}@ims.uni-stuttgart.de Abstract This paper presents a collection of 350 000 German lemmatised words, rated on four psycholinguistic affective attributes. All ratings were obtained via a supervised learning algorithm that can automatically calculate a numerical rating of a word. We applied this algorithm to abstractness, arousal, imageability and valence. Comparison with human ratings reveals high correlation across all rating types. The data discussed in the present paper is accessible at: http://www.ims.uni-stuttgart.de/data/affective_norms/ Keywords: affective norms, abstractness, arousal 1. Introduction Abstract words refer to things that we cannot perceive di- rectly with our senses (idea, politics, ...), concrete words on the other hand refer to things that we can perceive (image, scent, ...,). A large subset of concrete words have a high imageability, these are words that refers to things that we can actually see. Valence determines the pleasantness of a word (gift vs. punishment). Arousal describes the inten- sity of emotion provoked by a stimulus (alert vs. calm). These psycholinguistic attributes are often called affective norms. Information about the affective meaning of words is used by researchers working in multiple fields, such as sen- timent analysis, metaphor detection, word processing and lexical decision. Currently, a number of databases offers affective norms for words in different languages, including English (Altarriba et al., 1999; Bradley and Lang, 1994; Stevenson et al., 2007; Warriner et al., 2013), German (V˜o et al., 2006; V˜o et al., 2009; Lahl et al., 2009; Kanske and Kotz, 2010; Schmidtke et al., 2014), Spanish (Redondo et al., 2007), and Finnish (Eilola and Havelka, 2010). There is a number of resources that focuses on a single rating type only: Brys- baert et al. (2014) collected 40.000 abstractness ratings for English. The MRC Psycholinguistic Database1 contains roughly eight thousand abstractness ratings. However all of these resources cover only a small proportion of a lan- guage due to the fact that human annotators are required to obtain ratings. Especially the German resources contain usually less than 3000 words. In addition most resources focus on nouns only and therefore lack other word classes such as verbs or adjectives. A notable exception is the work by Turney et al. (2011), who used a method that overcome these limitations by applying an algorithm from Turney and Littman (2003). This algorithm uses distributional seman- tics (vector representations of words) and learns to assign a rating score to unseen words based on other known labelled training instances. Using this method, they were able to learn abstractness ratings for 114 501 English words. In this work we apply the same method as in Turney et al. (2011) and: (i) we learn ratings for German, (ii) we learn four dif- 1http://www.psych.rl.ac.uk/ ferent rating types, (iii) recent advances in the field of word representation learning (namely the methods of Mikolov et al. (2013)) allow us to learn ratings for a large vocabulary (≈350k …

Content truncated to 500 words