NoReC: The Norwegian Review Corpus

NoReC: The Norwegian Review Corpus Erik Velldal, Lilja Øvrelid, Eivind Alexander Bergem, Cathrine Stadsnes, Samia Touileb, Fredrik Jørgensen† Language Technology Group, Department of Informatics, University of Oslo {erikve, liljao, eivinabe, cathsta, samiat}@ifi.uio.no †Schibsted Media Group fredrik.jorgensen@schibsted.com Abstract This paper presents the Norwegian Review Corpus (NoReC), created for training and evaluating models for document-level sentiment analysis. The full-text reviews have been collected from major Norwegian news sources and cover a range of different domains, including literature, movies, video games, restaurants, music and theater, in addition to product reviews across a range of categories. Each review is labeled with a manually assigned score of 1–6, as provided by the rating of the original author. This first release of the corpus comprises more than 35,000 reviews. It is distributed using the CoNLL-U format, pre-processed using UDPipe, along with a rich set of metadata. The work reported in this paper forms part of the SANT initiative (Sentiment Analysis for Norwegian Text), a project seeking to provide open resources and tools for sentiment analysis and opinion mining for Norwegian. Keywords: Sentiment Analysis, Opinion Mining, Corpus, Norwegian, Reviews 1. Introduction For Norwegian, training and evaluation data are still lack- ing for many core NLP tasks. The current work aims to fill the gap for the particular task of sentiment analysis. The SANT project – Sentiment Analysis for Norwegian Text – seeks to create, and make publicly available, resources and tools for sentiment analysis for Norwegian. The SANT ef- fort described in the current paper marks the release of the Norwegian Review Corpus1 (NoReC). The dataset com- prises more than 35,000 full-text reviews from a range of different domains, collected from several of the major Nor- wegian news sources. Each review is rated with a numer- ical score on a scale of 1–6, and can be used for training and evaluating models for document-level sentiment anal- ysis, i.e., the task of predicting overall positive or negative polarity for a given text. 1.1. Rating by dice A particularity of review journalism in Norway, is the wholesale adoption of dice rolls (‘terningkast’) as a stan- dard rating scale: The item under review is rated on a scale of 1–6, commonly visualized by the face of a die with a corresponding number of ‘dots’ or pips. The practice is thought to have been introduced already in 1952 when reviewing movies in the newspaper Verdens Gang (VG). Given that the result of a die roll is otherwise associated with randomness, it is somewhat surprising that it would catch on as a symbol for summarizing reviews – something one would typically hope to construe as a well-informed and deliberate judgment of merit and quite the opposite of chance or luck. Nontheless, by now it has found widespread use in all sorts of arts and consumer journalism and is used when reviewing everything from books, theater and music, to home electronics, restaurants, and children’s clothing. The rating practice described above has several benefits for the goal of document-level SA: (i) It eliminates the …

Content truncated to 500 words