Multi-level analysis and recognition of the text sentiment on the example of consumer opinions

Proceedings of Recent Advances in Natural Language Processing, pages 559–567, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_066 559 Multi-level analysis and recognition of the text sentiment on the example of consumer opinions Jan Koco´n Wrocław University of Science and Technology Wrocław, Poland jan.kocon @pwr.edu.pl Monika Za´sko-Zieli´nska University of Wrocław Institute of Polish Studies Wrocław, Poland monika.zasko-zielinska @uwr.edu.pl Piotr Miłkowski Wrocław University of Science and Technology Wrocław, Poland piotr.milkowski @pwr.edu.pl Abstract In this article, we present a novel multi- domain dataset of Polish text reviews, an- notated with sentiment on different lev- els: sentences and the whole documents. The annotation was made by linguists in a 2+1 scheme (with inter-annotator agree- ment analysis). We present a preliminary approach to the classification of labelled data using logistic regression, bidirec- tional long short-term memory recurrent neural networks (BiLSTM) and bidirec- tional encoder representations from trans- formers (BERT). 1 Introduction Linguistic research on sentiment recognition in- volves two approaches: from the perspective of analysing the occurrence of emotional words and from the perspective of the entire document. The first attempt is usually a consequence of the cre- ation of the sentiment lexicon, e.g. manual anno- tation of the WordNet (Baccianella et al., 2010). The second results from the analysis of the spe- cific text content in which we see that the senti- ment of a word or phrase changes under the in- fluence of the surrounding context (Taboada et al., 2008). This change may vary depending on the domain of the text. As a research material for our research we have chosen online customer reviews from four domains: • S – School – students’ reviews on the lectur- ers1, • M – Medicine – patients’ opinions on doc- tors2, 1https://polwro.pl/ 2https://www.znanylekarz.pl/ • H – Hotels – customer reviews of hotels3, • P – Products – buyers’ opinions on prod- ucts4. In the introduction we focus mainly on the influ- ence of discourse on the classification of the doc- ument sentiment. We only briefly present an ap- proach based on the analysis of emotional words. The rest of the article concerns the description of the corpus used in our analysis, guidelines for the description of the text with sentiment for annota- tors, the results of the pilot stage of annotation, the proper annotation and experiments with automatic recognition of the text polarity. 2 Related work One approach for recognising polarity of text is to use a dictionary of emotional words – senti- ment lexicons, e.g. WordNet annotated with po- larity (Kamps et al., 2004; Takamura et al., 2005) and emotions (Janz et al., 2017). Usually the task is to determine the number of occurrence of such words with a specific polarity in the text or use a simple bag of words method (Wang and Manning, 2012). Such a solution has a number of limita- tions: simple methods cannot cope with irony, sar- casm, negation and more complex text structures that modify the sound of the words that make them up (Wallace et al., 2015). The second …

Content truncated to 500 words