SB-CH: A Swiss German Corpus with Sentiment Annotations

SB-CH: A Swiss German Corpus with Sentiment Annotations Ralf Grubenmann, Don Tuggener, Pius von D¨aniken, Jan Deriu, Mark Cieliebak Zurich University of Applied Sciences, Winterthur, Switzerland SpinningBytes AG, K¨usnacht, Switzerland rg@spinningbytes.com, tuge@zhaw.ch, piusvd@gmail.com, deri@zhaw.ch, mc@spinningbytes.com Abstract We present the SB-CH corpus, a novel Swiss German corpus with annotations for sentiment analysis. It consists of more than 200,000 phrases (approx. 1 Mio tokens) from Facebook comments and online chats. Additionally, we provide sentiment annotations for almost 2000 Swiss German phrases. We describe the methodologies used in the collection and annotation of the data, and provide the first baseline results for Swiss German sentiment analysis. Keywords: Swiss German, Sentiment Analysis, Emotion Corpus, Language Resources 1. Introduction Swiss German denotes a collection of Alemannic dialects spoken in the German-speaking part of Switzerland. Al- though not one of the official languages, it is spoken on a daily basis by an estimated 4,5 million speakers, i.e. by more than three fifths of the Swiss population (Coray and Bartels, 2017). Swiss German is almost exclusively used for private communications between native speakers, while Swiss Standard German, which closely resembles the writ- ten German in Germany, is used for official and public com- munications. Due to its non-official status and the phonetic differences between the dialects, there is no standardized spelling for Swiss German, and it has been hardly used in written form (Baumgartner, 2003). Through the advent of social media and messaging systems, Swiss German has increasingly expanded to written form (Stark et al., 2015, e.g.). However, resources for written Swiss German are still sparse, and it can be considered a low-resourced language. Combined with the fluent nature of the (mostly) oral lan- guage, it is non-trivial to find a large collection of written dialect. Furthermore, there exists no Swiss German corpus for Sentiment Analysis. Contributions In this paper, we present the following1: • A new corpus SB-CH composed of more than 200,000 Swiss German phrases2 and approx. 1 Mio tokens. • Sentiment annotations for 1843 phrases with labels positive, negative, or neutral. • Baselines for Swiss German sentiment classification. 2. Related Work Language Resources for Swiss German At present, we are aware of the following resources for Swiss German: • Swiss SMS Corpus (Stark et al., 2015): a collec- tion of SMS which were sent by the Swiss public in 1The corpus including the annotations is available here: https://www.spinningbytes.com/resources/ swissgermansentiment 2The term phrases is used to describe a unit of the dataset, which is typically a sentence or a short paragraph of text. 2009/2010. It contains 10,708 SMS in Swiss German, together with demographic information about the au- thor. • NOAH’s Corpus of Swiss German Dialects (Hol- lenstein and Aepli, 2015): a compilation of 7,453 Swiss German texts from various genres, collected be- tween 2010-2014. Text sources include the Alemannic Wikipedia, blog posts, novels by Viktor Schobinger, newspaper articles from ”Blick am Abend”, and the Swatch Annual Business Report. The texts were man- ually annotated with part-of-speech tags for 106,987 tokens. • Sprachatlas der Schweiz …

Content truncated to 500 words