A Large Wordnet-based Sentiment Lexicon for Polish

Proceedings of Recent Advances in Natural Language Processing, pages 721–730, Hissar, Bulgaria, Sep 7–9 2015. A Large Wordnet-based Sentiment Lexicon for Polish Monika Za´sko-Zieli´nska University of Wrocław Wrocław, Poland monika.zasko-zielinska@uwr.edu.pl Maciej Piasecki Wrocław University of Technology Wrocław, Poland maciej.piasecki@pwr.edu.pl Stan Szpakowicz University of Ottawa Ottawa, Ontario, Canada szpak@eecs.uottawa.ca Abstract The applications of plWordNet, a very large wordnet for Polish, do not yet in- clude work on sentiment and emotions. We present a pilot project to annotate plWordNet manually with sentiment po- larity values and basic emotion values. We work with lexical units, plWordNet’s ba- sic building blocks.1 So far, we have an- notated about 30,000 nominal and adjecti- val LUs. The resulting lexicon is already one of the largest sentiment and emotion resources, in particular among those based on wordnets. We opted for manual an- notation to ensure high accuracy, and to provide a reliable starting point for fu- ture semi-automated expansion. The pa- per lists the principal assumptions, out- lines the annotation process, and intro- duces the resulting resource, plWordNet- emo. We discuss the selection of the ma- terial for the pilot study, show the distri- bution of annotations across the wordnet, and consider the statistics, including inter- annotator agreement and the resolution of disagreement. 1 Introduction The Polish wordnet, plWordNet (Piasecki et al., 2009; Maziarz et al., 2013), is very large and comprehensive, with well over 150,000 synsets 1The term lexical unit will be abbreviated to LU through- out this paper. and 200,000 LUs at the time of writing. It has many applications, e.g., text similarity (Siemi´nski, 2012), terminology extraction and clustering (Mykowiecka and Marciniak, 2012), extraction of opinion attributes from product descriptions (Wawer and Gołuchowski, 2012), addition of fea- tures for text mining (Maciołek and Dobrowolski, 2013), or a mapping between a lexicon and an on- tology (Wróblewska et al., 2013). It is fast becom- ing a go-to resource in Polish lexical semantics. So far, however, it has not supported applications in the crucially important area of sentiment analy- sis and opinion mining. That area requires annota- tion: a word or word sense either does or does not carry sentiment, emotion or affect. That is why we have recently set out to annotate plWordNet with sentiment polarity and basic emotions. Automatic annotation of lexical material is not a viable option. Wordnets are reference resources, relied upon for the absence of lexical errors. In fact, all widely published sentiment-marked and emotion-marked lists of lexical items have been created manually, sometimes by crowdsourcing. Now, plWordNet is much too large for com- plete, affordable manual annotation, but a reliable core of as little as 10% of the wordnet annotated makes it entirely possible to continue with semi- automatic expansion. Our pilot project manually annotated around 30,000 LUs (15% of plWord- Net)2 with sentiment and basic emotions, so we have ample material to also compare fully manual and semi-automatic annotation. 2This annotation is already on a scale several times larger than SentiWordNet (Esuli and Sebastiani, 2006). 721 2 Sentiment and Affect …

Content truncated to 500 words