Building Sentiment Lexicons for Mainland Scandinavian Languages Using Machine Translation and Sentence Embeddings

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2816–2825 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2816 Building Sentiment Lexicons for Mainland Scandinavian Languages Using Machine Translation and Sentence Embeddings Peng Liu∗, Cristina Marco∗,†, Jon Atle Gulla Department of Computer Science, NTNU, Trondheim, Norway {peng.liu,jon.atle.gulla}@ntnu.no,cristinasmarco@gmail.com Abstract This paper presents a simple but effective method to build sentiment lexicons for the three Mainland Scandinavian languages: Danish, Norwegian and Swedish. This method benefits from the English Sentiwordnet and a thesaurus in one of the target languages. Sentiment information from the English resource is mapped to the target languages by using machine translation and similarity measures based on sentence embeddings. A number of experiments with Scandinavian languages are performed in order to determine the best working sentence embedding algorithm for this task. A careful extrinsic evaluation on several datasets yields state-of-the-art results using a simple rule-based sentiment analysis algorithm. The resources are made freely available under an MIT License. Keywords: Sentiment lexicon, Scandinavian languages, Machine translation, Sentence embedding 1. Introduction Ubiquitous and cognitive artificial intelligence is be- coming a reality that is shaping our world. By inter- acting through voice with conversational systems, it is now possible to make purchases online or follow a recipe while you are in the kitchen. Besides, knowl- edge derived from continuous text data sources coming from social media or online news sources plays an in- creasingly important role in political campaigns, finan- cial analysis, or analysis of medical records. The fine- grained analysis of sentiment and emotion is becoming essential for all these tasks. Automatically extracting the positive or negative orien- tation that a passage expresses toward some targets, so- called sentiment analysis or opinion mining, is one of the fundamental tasks of text categorization. As these opinions or sentiments are written in natural languages and considering the lack of training corpora in several languages, one of the most important tools needed to do such analysis is sentiment lexicons, which contain lists of positive, negative and, in some cases, neutral words. Despite the fact that English resources exist to do this task, sentiment lexicons or training corpora in other languages are often not easily available. To alleviate the problem raised above, we propose a simple but effective method to develop large gen- eral sentiment lexicons for the three Mainland Scan- dinavian languages, Danish, Norwegian and Swedish. Specifically, to create these resources, machine transla- tion and sentence embeddings methods are used to map the sentiment information from an already available English resource to a thesaurus in one of the target lan- guages. Machine translation is used to bridge this lex- icon into several languages. In particular, we focus on experimenting with Danish, Swedish and Norwegian. ∗Equal contribution. †This work was done prior to the author joining Amazon. Because there are a handful of evaluation datasets for these three under-resourced languages, we construct four new datasets based on Tripadvisor reviews as a new benchmark to evaluate the generated …

Content truncated to 500 words