NusaX: Multilingual Parallel Sentiment Dataset for 10 Indonesian Local Languages

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 815–834 May 2-6, 2023 ©2023 Association for Computational Linguistics NusaX: Multilingual Parallel Sentiment Dataset for 10 Indonesian Local Languages Genta Indra Winata1∗, Alham Fikri Aji2∗, Samuel Cahyawijaya3∗, Rahmad Mahendra4,5∗, Fajri Koto2,6∗, Ade Romadhony5,7∗, Kemal Kurniawan5,6∗, David Moeljadi8, Radityo Eko Prasojo9, Pascale Fung3, Timothy Baldwin2,6, Jey Han Lau6, Rico Sennrich10, Sebastian Ruder11 1Bloomberg 2MBZUAI 3HKUST 4Universitas Indonesia 5INACL 6The University of Melbourne 7Telkom University 8Kanda University of International Studies 9Kata.ai 10University of Zurich 11Google Research Abstract Natural language processing (NLP) has sig- nificant impact on society via technologies such as machine translation and search engines. Despite its success, NLP technology is only widely available for high-resource languages such as English and Mandarin Chinese, and re- mains inaccessible to many languages due to the unavailability of data resources and bench- marks. In this work, we focus on developing re- sources for languages of Indonesia. Despite be- ing the second most linguistically-diverse coun- try, most languages in Indonesia are catego- rized as endangered and some are even extinct. We develop the first-ever parallel resource for 10 low-resource languages in Indonesia. Our resource includes sentiment and machine trans- lation datasets, and bilingual lexicons. We provide extensive analysis, and describe chal- lenges for creating such resources. Our hope is that this work will spark more NLP research on Indonesian and other underrepresented lan- guages. 1 Introduction Indonesia is one of the most populous and linguistically-diverse countries in the world, with more than 700 languages spoken across the coun- try (Aji et al., 2022; Eberhard et al., 2021). How- ever, while many of these languages are spoken by millions of people they have received little atten- tion from the NLP community. There are very few public datasets, preventing the global research com- munity from exploring these languages. To this end, we introduce NusaX,1 a high-quality multilingual parallel corpus that covers 10 local languages from Indonesia: Acehnese, Balinese, Banjarese, Bugi- nese, Madurese, Minangkabau, Javanese, Ngaju, Sundanese, and Toba Batak. ∗ These authors contributed equally. 1The dataset is released at https://github.com/ IndoNLP/nusax. The NusaX dataset was created by translating SmSA (Purwarianti and Crisdayanti, 2019) — an existing Indonesian sentiment analysis dataset con- taining comments and reviews from the IndoNLU benchmark (Wilie et al., 2020) — using compe- tent bilingual speakers, coupled with additional human-assisted quality assurance. Sentiment anal- ysis is one of the most popular NLP tasks, and has been explored in many applications in Indone- sia, including presidential elections (Ibrahim et al., 2015; Budiharto and Meiliana, 2018), product re- views (Fauzi, 2019), stock forecasting (Cakra and Trisedya, 2015; Sagala et al., 2020), and COVID- 19 monitoring (Nurdeni et al., 2021). By translating an existing text, we additionally produce a parallel corpus, which is useful for building and evaluat- ing translation systems. As we translate from a regional high-resource language (Indonesian), we ensure that the topics and entities reflected in the data are culturally relevant to the other languages, which is generally …

Content truncated to 500 words