Cross-lingual Emotion Detection

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6948–6958 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6948 Cross-lingual Emotion Detection Sabit Hassan1, Shaden Shaar2, Kareem Darwish3 1University of Pittsburgh, 2Cornell University, 3aiXplain Inc. 1Pittsburgh, PA 15260, 2Ithaca, NY 14850, 3Los Gatos, CA 95032 sah259@pitt.edu, ss2753@cornell.edu, kareem.darwish@aixplain.com Abstract Emotion detection can provide us with a window into understanding human behavior. Due to the complex dynamics of human emotions, however, constructing annotated datasets to train automated models can be expensive. Thus, we explore the efficacy of cross-lingual approaches that would use data from a source language to build models for emotion detection in a target language. We compare three approaches, namely: i) using inherently multilingual models; ii) translating training data into the target language; and iii) using an automatically tagged parallel corpus. In our study, we consider English as the source language with Arabic and Spanish as target languages. We study the effectiveness of different classification models such as BERT and SVMs trained with different features. Our BERT-based monolingual models that are trained on target language data surpass state-of-the-art (SOTA) by 4% and 5% absolute Jaccard score for Arabic and Spanish respectively. Next, we show that using cross-lingual approaches with English data alone, we can achieve more than 90% and 80% relative effectiveness of the Arabic and Spanish BERT models respectively. Lastly, we use LIME to analyze the challenges of training cross-lingual models for different language pairs. Keywords: cross-lingual, multi-lingual, knowledge transfer, emotion detection 1. Introduction Detecting emotions in text has a wide variety of ap- plications ranging from identifying anger in customer responses to evaluating the emotional well-being of in- dividuals and societies. Hence, there has been much recent work on automatically detecting emotions in text, with special emphasis on social media posts. Despite contention (Ortony and Turner, 1990), much work on psychology suggests that basic emotions are universal (Ekman, 1999; Scarantino and Griffiths, 2011). How- ever, the expressions and perceptions of emotions may exhibit cross-cultural variations (Hareli et al., 2015), which may complicate cross-lingual learning. Nonethe- less, the ability to perform such cross-lingual classifica- tion of emotions effectively is advantageous as it would avoid the laborious task of building emotion detection datasets, which involves collecting representative tweets and properly tagging them. While cross-lingual methods have shown success for some tasks such as cross-lingual search (Chin et al., 2014), named entity recognition (Darwish, 2013), and sentiment analysis (Zhou et al., 2016), lesser success has been observed for other tasks such as parsing (Guo et al., 2015) and offensive language detection (Pamungkas and Patti, 2019), due to the divergences in linguistic and cul- tural specificities. In this paper, we explore the efficacy of using an existing tagged emotion detection corpus from one language to another, and we perform rigor- ous analysis of results to ascertain the sources of errors. Specifically, we attempt to extend the SemEval2018 English emotion detection dataset (Mohammad et al., 2018a) to Arabic and Spanish. Though the effectiveness …

Content truncated to 500 words