NusaX: Multilingual Parallel Sentiment Dataset for 10 Indonesian Local Languages

by   Genta Indra Winata, et al.

Natural language processing (NLP) has a significant 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 Chinese, while it remains inaccessible to many languages due to the unavailability of data resources and benchmarks. In this work, we focus on developing resources for languages in Indonesia. Despite being the second most linguistically diverse country, most languages in Indonesia are categorized as endangered and some are even extinct. We develop the first-ever parallel resource for 10 low-resource languages in Indonesia. Our resource includes datasets, a multi-task benchmark, and lexicons, as well as a parallel Indonesian-English dataset. We provide extensive analyses and describe the challenges when creating such resources. We hope that our work can spark NLP research on Indonesian and other underrepresented languages.


page 1

page 2

page 3

page 4


One Country, 700+ Languages: NLP Challenges for Underrepresented Languages and Dialects in Indonesia

NLP research is impeded by a lack of resources and awareness of the chal...

Masakhane – Machine Translation For Africa

Africa has over 2000 languages. Despite this, African languages account ...

Location-based Twitter Filtering for the Creation of Low-Resource Language Datasets in Indonesian Local Languages

Twitter contains an abundance of linguistic data from the real world. We...

Language Resources and Technologies for Non-Scheduled and Endangered Indian Languages

In the present paper, we will present a survey of the language resources...

The first large scale collection of diverse Hausa language datasets

Hausa language belongs to the Afroasiatic phylum, and with more first-la...

Challenges of language technologies for the indigenous languages of the Americas

Indigenous languages of the American continent are highly diverse. Howev...

What a Creole Wants, What a Creole Needs

In recent years, the natural language processing (NLP) community has giv...

1 Introduction

Indonesia is one of the most populous and diverse countries, with more than 700 languages spoken across the country aji-etal-2022-one; ethnologue. However, while many of these languages are spoken by millions of people they receive little attention by the NLP community. Public datasets are practically non-existent, which prevents the global research community from exploring these languages. To this end, we introduce NusaX, a high-quality multilingual parallel corpus that covers 10 local languages in Indonesia, namely Acehnese, Balinese, Banjarese, Buginese, Madurese, Minangkabau, Javanese, Ngaju, Sundanese, and Toba Batak.

The NusaX dataset is created by translating SMSA purwarianti2019improving

— an existing Indonesian sentiment analysis dataset containing comments and reviews from the IndoNLU benchmark 

wilie2020indonlu — using competent bilingual speakers, coupled with additional human-assisted quality assurance to ensure the quality of the translation. Sentiment analysis is one of the most popular NLP tasks arxiv-najjasenti and has many practical uses in industry cambria2017affective. Sentiment analysis has been explored in many applications in Indonesia, including but not limited to presidential elections ibrahim2015buzzer; budiharto2018prediction, product reviews fauzi2019word2vec, stock prices cakra2015stock; sagala2020stock, and COVID-19 nurdeni2021sentiment. By translating existing text, we additionally produce a parallel corpus, which is useful for building and evaluating translation systems. As we translate from a related high-resource language (Indonesian), we ensure that the topics and entities reflected in the data are culturally relevant to the other languages, which is not the case when translating an English dataset conneau2018xnli; ponti-etal-2020-xcopa.

We apply the corpus to two downstream tasks: sentiment analysis and machine translation. We create a new benchmark and evaluate the performance of existing Indonesian language models, multilingual language models, and classical machine learning methods in few-shot and full-data settings.

Figure 1: Language taxonomy of the studied Indonesian languages according to Ethnologue ethnologue. The color represents the language category level in the taxonomy. Purple denotes language and other colors denote language family.

Our contributions are as follows:

  • We propose NusaX111The code and dataset is released at, the first high-quality human annotated parallel corpus in 10 languages from Indonesia, and corresponding parallel data in Indonesian and English.

  • We introduce datasets in these local languages covering two tasks: sentiment analysis and machine translation.

  • We provide an extensive evaluation covering deep learning and classical NLP and machine learning methods on downstream tasks in few-shot and full-data settings.

  • We conduct comprehensive analyses regarding the similarity of the languages under study from both a linguistics and empirical perspective, the cross-lingual transferability of existing monolingual and multilingual pre-trained models, and an efficiency analysis of different approaches for NLP tasks in extremely low-resource languages.

No. of L1
No. of
Acehnese (ace) Threatened
(2000 census)
7 Latin
Balinese (ban) Developing
(2010 census)
Banjarese (bjn)
(2015 UNSD)
2 Latin, Arabic
Toba Batak (bbc) Threatened
(1991 UBS)
Buginese (bug)
(2010 census)
Indonesian (ind) National
(2015 UNSD)
Javanese (jav) Educational
(2015 UNSD)
Madurese (mad) Developing
(2015 UNSD)
6 Latin
Minangkabau (min) Developing
(2015 UNSD)
12 Latin
Ngaju (nij)
7 Latin
Sundanese (sun) Developing
(2015 UNSD)
Table 1: Language data according to Ethnologue ethnologue. We list the canonical language name, ISO 639-2 code, status, number of L1 speakers in Indonesia, number of dialects, and writing systems. The language status information can be found at

2 Focus Languages

In this work, we focus on 10 languages, namely Acehnese, Balinese, Banjarese, Buginese, Madurese, Minangkabau, Javanese, Ngaju, Sundanese, and Toba Batak. Figure 1 shows the taxonomy of these languages (including Indonesian) according to Ethnologue ethnologue. The languages covered belong to the Austronesian language family under the Malayo-Polynesian subgroup. Table 1 provides additional information such as their status, number of L1 speakers, number of dialects, and writing systems.

Acehnese (ace) is a language spoken mainly in the Aceh province. Although it is the de facto language of provincial identity of Aceh, language use is shifting to Indonesian in urban areas. Acehnese has features typical of the Mon-Khmer languages of mainland Southeast Asia, a result of its former status as part of the early Chamic dialect continuum on the coast of Vietnam. It has at least ten contrasting vowels and as many distinct diphthongs, as well as voiceless aspirated stops and murmured voiced stops blust2013austronesian. In addition to the large number of diphthongs, it has a high percentage of monosyllabic root morphemes. Prefixes and infixes play an active role while suffixes are absent Durie:1985. It is of the ‘active’ or so-called ‘Split-S’ type: some intransitive verbs take arguments, which have the properties of ‘transitive subjects’ while others take arguments with the properties of ‘transitive objects’ durie1988preferred.

Balinese (ban) is a language spoken mainly in the Bali province and in the West Nusa Tenggara province. It has three main dialects: Highland Balinese, Lowland Balinese, and Nusa Penida. It is mainly written in the Latin script since the early 20th century although it has its own Balinese script. The word order in Balinese is SVO. It is non-tonal and has 17 consonant and 6 vowel phonemes. Stress is on the penultimate syllable. It has three sociolinguistic registers. Regarding patterns of verb affixation, Balinese is an ‘active’ or ‘split-S’ language: verbs with Undergoer-like subject arguments are marked in one way (with a ‘zero prefix’), while verbs with Actor-like subject arguments—intransitive or transitive—are marked in another (either with the nasal prefix ‘N-’, or with ‘ma-’) arka2003balinese.


(bjn) is a language spoken in Kalimantan (Central, East, South, and West Kalimantan provinces). It became a language of wider communication through trade in the market, in business, and in media. It is dominant in the South Kalimantan Province and also growing rapidly in the Central and Eastern Kalimantan provinces. It has two main dialects: Kuala and Hulu dialects. Although it is a Malayic language, it has many Javanese loanwords, probably acquired during the Majapahit period from the late thirteenth century until the fifteenth century 

blust2013austronesian. It has 73% of lexical similarity with Indonesian222i.e., 73% of its words also occur in Indonesian. and it is written in Arabic and Latin scripts ethnologue.

Toba Batak (bbc) is a language spoken in the North Sumatra province. Similarly to Acehnese, it is slowly being replaced by Indonesian in urban and migrant areas. It used to be written in the Batak script but is mainly written in Latin script now. The Batak languages are predicate-initial, and have verb systems reminiscent of Philippine languages, although they differ from them in many details blust2013austronesian.

Buginese (bug) is a language spoken mainly in the South Sulawesi, Southeast Sulawesi, Central Sulawesi, and West Sulawesi provinces. The word order is SVO. Verb affixes are used to mark persons. It is non-tonal and has 19 consonant and 6 vowel phonemes. Stress is on the penultimate syllable. It was written in the Buginese script in the past (derived from Brahmi script) but is mainly written in Latin script now ethnologue. In Buginese, the pronoun ‘I’ has three forms: the independent form ‘iyya’, the ergative form ‘-ka’, and the absolutive form/clitic ‘u-’. Buginese employs sentence patterns, pronouns, and certain terms to express politeness weda2016syntactic.

Indonesian (ind) is the national language of Indonesia in 1945 Constitution, Article 36  indonesia2002undang. Its lexical similarity to Standard Malay is over 80%. The word order is SVO. It is non-tonal and has 19 consonants, 6 vowels, and 3 diphthongs. The stress is on the penultimate syllable. It has a rich affixation system, including a variety of prefixes, suffixes, circumfixes, and reduplication. Most of the affixes in Indonesian are derivational. It is developed from literary ‘Classical Malay’ of the Riau-Johor sultanate sneddon2003 and has regional variants. It is written mainly in Latin script.

Madurese (mad) is a language spoken in the East Java province, mainly on Madura Island, south and west of Surabaya city, Bawean, Kangean, and Sapudi islands. It has vowel harmony, gemination, rich affixation, three types of reduplication, and SVO basic word order davies2010grammar.


(min) is a language spoken mainly in West Sumatra and other provinces on Sumatra Island such as Bengkulu and Riau. Although it is classified as Malay, it is not intelligible with Indonesian. The word order is SVO written in Latin script. Standard Minangkabau voice can be characterised as an Indonesian-type system whereas colloquial Minangkabau voice is more effectively characterised as a Sundic-type system 


Javanese (jav) is a language spoken mainly in Java island. It is the de facto language of provincial identity in central and eastern Java. The word order is SVO. It has 21 consonants and 8 vowels. It used to be written in Javanese script but since 20th century is mostly written in Latin script. Javanese differs from most other languages of western Indonesia in contrasting dental and retroflex stops, and in the feature of breathy voice or murmur as a phonetic property of its voiced obstruents. Javanese also differs from most languages of the Philippines and western Indonesia in allowing a number of word-initial consonant clusters. It has an elaborate system of speech levels blust2013austronesian.

Ngaju (nij) is a language spoken in the Central Kalimantan province. It is widely used as a language of wider communication for trade in much of Kalimantan, from the Barito to the Sampit river. It is used in many domains (church, school, village-level government, market, etc.). It has various affixes and reduplication, similar to Indonesian. The active voice is marked by prefix ‘maN-’ and the passive voice is marked by prefix ‘iN-’. The word order is similar to the one in Indonesian. The pronouns have enclitic forms to mark possessors in a noun phrase or agents in a passive sentence UchiboriShibata1988.

Sundanese (sun) is a language spoken mainly in the Banten and West Java provinces. It is the de facto language of provincial identity in western Java. The main dialects are Bogor (Krawang), Pringan, and Cirebon. It is non-tonal and has 18 consonant and 7 vowel phonemes. The stress is on the penultimate syllable. It has elaborate coding of respect levels. It is written in Latin script since the middle of the 19th century but was previously written in Arabic, Javanese, and Sundanese scripts. Sundanese is a predominantly SVO language. It has voice marking and incorporates some (optional) actor-verb agreement, i.e., number and person kurniawan2013sundanese.

3 Data Construction

Our data collection process consists of several steps. First, we take an existing dataset in a related high-resource language (Indonesian) as a base for expansion to the other ten languages. Then, we ask human annotators to translate the data. To ensure the quality of the final translation, we run quality assurance with additional human annotators.

3.1 Annotator Recruitment

Eliciting or annotating data in underrepresented languages generally requires working with local language communities in order to identify competent bilingual speakers Nekoto2020. In the Indonesian setting, this challenge is compounded that by the fact that most languages have several dialects (cf. Table 1). As dialects in Indonesian languages may have significant differences in word usage and meaning aji-etal-2022-one, it is important to recruit annotators who speak the same or similar dialects to ensure that translations are mutually intelligible.

In this work, we employ at least 2 native speaker annotators for each local language who still practice the local language in daily communication in addition to Indonesian (see Appendix A). We run a preliminary test to filter the annotators to be recruited: We first ask annotators to translate three samples. We then conduct a peer review by asking asking annotators whether they can understand the translations of other annotators for the same language. The hired annotators work as translators, as well as translation validators. For English translations, we hire annotators based on their English proficiency test with IELTS 6.5 or TOEFL PBT 600.

3.2 Data Filtering and Sampling

We take an existing sentence-level Indonesian sentiment analysis data, SMSA, the largest publicly available Indonesian sentiment analysis dataset containing comments and reviews from the IndoNLU benchmark purwarianti2019improving; wilie2020indonlu as the data source to build the multilingual parallel corpus. SMSA is a sentence-level multi-domain expertly-annotated sentiment analysis dataset consisting of more than 11,000 data instances of comments and reviews collected from several online platforms such as Twitter, Zomato, TripAdvisor, etc. We filter the data that contains abusive language and personal information by manually inspecting all sentences. Finally, we randomly select 1,000 samples via stratified sampling for translation, ensuring that the label distribution is balanced.

3.3 Human Translation

We instruct the annotators to retain the meaning of the text and to keep entities such as persons, organizations, locations, and time with no target language translation the same. Specifically, we instruct them to (1) retain the entity; (2) maintain the complete information content of the original sentence; (3) maintain the sentence’s sentiment polarity. Initially, we also asked the translators to maintain the typography. Most sentences from the original dataset are written in an informal tone, with non-standard spelling, e.g. elongated vowel and punctuations. When the sentence is translated into the target local language, direct translation can sound unnatural. For example: ‘kangeeeen’ (ind) (translation in eng: ‘miss’). The translation in Minangkabau is ‘taragak’ (as it is quite unnatural to express it as ‘taragaaak’). Similarly, the original sentence may also contain typos. Due to the difficulty of accurately assessing typographical consistency of translations, we remove this as a criterion. We instruct annotators to use the Latin script as it is the most used writing system among locals.

3.4 Human-Assisted Quality Assurance

We conduct quality control (QC) between two annotators where annotator A is asked to check the translation result of annotator B, and vice versa. We include the corrected translations in our dataset. To ensure the quality assurance is performed well, we randomly perturb 5% of the sentences by removing a random sequence of words. The quality assurance annotators are then expected to notice the perturbed sentences and to fix them.

We analyze the quality assurance edits for Balinese (ban), Minangkabau (min), Sundanese (sun), and Javanese (jav), which are spoken by the authors of this paper. For each language, we randomly sample 100 translations that have been edited by a QC annotator. We classify edits along the following categories:

Typos and Mechanics: Edit that involves correcting typos, punctuation, casing, white spaces/dashes, and numerical formatting.

Orthography: Edit that changes the spelling of words due to orthographic variation in local languages without a standard orthography. The word sounds and means the same before and after editing and both are being used by natives. The QC annotator might feel that one writing variant is more natural/commonly used, hence making this change.

Translation: The words used by the translator are still in Indonesian and the QC annotator translates them to the local language.

Word edit: The QC annotator paraphrases a word/phrase. This also includes adding/removing words and morpheme changes.

Major changes: Other edits that significantly alter the original translation.

Category ban sun jav
Typos & Mechanic 31 14 42
Orthography 14 6 12
Translation 22 55 10
Word edit 67 65 61
Major changes 3 0 1
Table 2: Statistics of QC edits per category.

The result is shown in Table 2. Generally, word edits make up the majority of QC modifications, which involves changing a word/phase to its synonym or altering the morpheme slightly. In contrast, major changes are extremely rare. We also see some changes of the orthography around 10% of the time. Other types of edits vary between languages. Sundanese has significantly less typos compared to other languages, but a considerably higher number of translation edits. We suspect this is because code-switching with Indonesian happens regularly in Sundanese, which results in many Indonesian words being adopted despite the existence of equivalent Sundanese translations.

4 NusaX Benchmark

4.1 Tasks

Model ace ban bbc bjn bug eng ind jav mad min nij sun avg
Naive Bayes 72.5 72.6 73.0 71.9 73.7 76.5 73.1 69.4 66.8 73.2 68.8 71.9 72.0
SVM 75.7 75.3 76.7 74.8 77.2 75.0 78.7 71.3 73.8 76.7 75.1 74.3 75.4
LR 77.4 76.3 76.3 75.0 77.2 75.9 74.7 73.7 74.7 74.8 73.4 75.8 75.4
IndoBERT 75.4 74.8 70.0 83.1 73.9 79.5 90.0 81.7 77.8 82.5 75.8 77.5 78.5
IndoBERT 76.3 79.5 74.0 83.2 70.9 87.3 90.2 85.6 77.2 82.9 75.8 77.2 80.0
IndoLEM 72.6 65.4 61.7 71.2 66.9 71.2 87.6 74.5 71.8 68.9 69.3 71.7 71.1
mBERT 72.2 70.6 69.3 70.4 68.0 84.1 78.0 73.2 67.4 74.9 70.2 74.5 72.7
XLM-R 73.9 72.8 62.3 76.6 66.6 90.8 88.4 78.9 69.7 79.1 75.0 80.1 76.2
XLM-R 75.9 77.1 65.5 86.3 70.0 92.6 91.6 84.2 74.9 83.1 73.3 86.0 80.0
Table 3: Sentiment analysis result in marco-F1(%). Models were trained and evaluated on each language independently.

We develop two tasks (sentiment analysis and machine translation) based on the datasets covering 12 languages, including Indonesian, English, and 10 local languages.

4.1.1 Sentiment Analysis

Sentiment analysis is an NLP task that aims to extract the sentiment from a given text document. The sentiment is commonly categorized into 3 classes, i.e., positive, negative, and neutral. We focus our dataset construction on sentiment analysis because it is one of the most widely explored tasks in Indonesia aji-etal-2022-one due to its broad industry applications such as competitive and marketing analysis, and detection of unfavorable rumors for risk management socher2013recursive.

After translating 1,000 instances from the sentiment analysis dataset (SMSA), we have a sentiment analysis dataset for each translated language. For each language, we split the dataset into 500, 100, and 400 train, validation, and test examples. In total, our dataset contains 6,000 train, 1,200 validation, and 4,800 test instances for all 12 languages (Indonesian, English and the 10 local languages).

4.1.2 Machine Translation

Indonesia consists of 700+ languages covering three different language families  aji-etal-2022-one. Despite its linguistic diversity, existing machine translation systems only cover a small fraction of Indonesian languages, mainly Indonesian (the national language), Sundanese, and Javanese. To broaden the coverage of existing machine translation systems for underrepresented local languages, we construct a machine translation dataset using our translated sentiment corpus, which results in a parallel corpus between all language pairs. In other words, we have 132 possible parallel corpora each with 1,000 samples (500 train, 100 validation, and 400 test instances) which can be used to train machine translation models. Compared to many other MT evaluation datasets, our data is in the review domain and is not English-centric.

4.2 Baselines

4.2.1 Classical Machine Learning

Classical machine learning approaches are still widely used by local Indonesian researchers and institutions, due to their efficiency nityasya2021costs

. The trade-off between performance and compute cost is particularly important in situations with limited compute that are common for work with low-resource languages. We therefore use classical methods as baselines for our comparison. Namely, we use Naive Bayes, SVM, and logistic regression for the classification tasks. For MT, we employ a naive baseline that simply copies the original Indonesian text, a dictionary-based substitution method, and a phrase-based MT system based on Moses 


4.2.2 Pre-trained Local Language Models

Recent developments in neural pre-trained language models (LMs) have brought substantial improvements in various NLP tasks. Despite the lack of resources in Indonesian and local languages, there are several efforts in developing large pre-trained language models for Indonesian and major local languages. IndoBERT wilie2020indonlu and SundaneseBERT wongso2022pre are two popular models for natural language understanding (NLU) tasks in Indonesian and Sundanese. IndoBART and IndoGPT have also been introduced for natural language generation (NLG) tasks in Indonesian, Sundanese, and Javanese cahyawijaya-etal-2021-indonlg. We employ these pretrained models as baselines to assess their adaptability to other languages.

4.2.3 Pre-trained Massively Multilingual LMs

We additionally consider large pre-trained multilingual language models to further understand their applicability to low-resource languages. Specifically, we experiment with mBERT devlin2019bert and XLM-R conneau-etal-2020-unsupervised for NLU tasks and mBART liu-etal-2020-multilingual-denoising and mT5 xue-etal-2021-mt5 models for NLG tasks. We provide the hyper-parameters of all models in Appendix B.

5 Results and Discussion

5.1 Overall Results

Sentiment Analysis
idn x
Model ace ban bbc bjn bug eng jav mad min nij sun avg
Copy 5.89 10.00 4.28 15.99 3.45 0.56 9.29 5.11 18.10 7.52 9.24 8.13
Word Substitution 7.60 10.31 5.99 17.51 3.57 0.76 14.75 7.58 22.34 9.76 12.38 10.23
PBSMT 20.47 26.48 18.18 42.08 10.84 7.73 39.08 33.26 52.21 29.58 36.04 28.72
IndoGPT 9.60 14.17 8.20 22.23 5.18 5.89 24.05 14.44 26.95 17.56 23.15 15.58
IndoBARTv2 19.21 27.08 18.41 40.03 11.06 11.53 39.97 28.95 48.48 27.11 38.46 28.21
mBART-50 17.21 22.67 17.79 34.26 10.78 3.90 35.33 28.63 43.87 25.91 31.21 24.69
mT5 14.79 18.07 18.22 38.64 6.68 11.21 33.48 0.96 45.84 13.59 33.79 21.39
Table 4: Results of the machine translation task from Indonesian to other languages (idn x) in SacreBLEU.
x idn
Model ace ban bbc bjn bug eng jav mad min nij sun avg
Copy 5.88 9.99 4.28 15.99 3.44 0.57 9.29 5.11 18.10 7.51 9.24 8.13
Word Substitution 7.33 12.30 5.02 16.17 3.52 1.67 17.34 7.89 24.17 12.07 15.38 11.17
PBSMT 25.17 41.22 20.94 47.80 15.21 6.68 46.99 38.39 60.56 32.86 41.79 34.33
IndoGPT 7.01 13.23 5.27 19.53 1.98 4.26 27.31 13.75 23.03 10.83 23.18 13.58
IndoBARTv2 24.44 40.49 19.94 47.81 12.64 11.73 50.64 36.10 58.38 33.50 45.96 34.69
mBART-50 18.45 34.23 17.43 41.73 10.87 17.92 39.66 32.11 59.66 29.84 35.19 30.64
mT5 18.59 21.73 12.85 42.29 2.64 12.96 45.22 32.35 58.65 25.61 36.58 28.13
Table 5: Results of the machine translation task on from other languages to Indonesian (x idn) in SacreBLEU.

Table 3 shows the sentiment analysis performance of various models across different local languages trained and evaluated using the same language data. Fine-tuned large LMs such as IndoBERT or XLM-R generally achieve the best performance. XLM-R models achieve strong performance on some languages, such as Indonesian (idn), Banjarese (bjn), English (eng), Javanese (jav), Minangkabau (min). We conjecture that this might be because these languages are included among the supported languages in the multilingual model. An exception is Banjarese, although it is very similar to Indonesian and Minangkabau. IndoBERT models, despite only being pre-trained on Indonesian, also show good performance across some local languages, suggesting transferability from Indonesian to the local languages.

The classic approaches are surprisingly competitive compared to the neural methods. Logistic regression even outperforms both IndoBERT and XLM-R on Acehnese (ace), Buginese (bug) and Toba Batak (bbc). These results indicate that both Indonesian and multilingual pre-trained language models cannot transfer well to these languages, which is supported by the fact that these languages are very distinct from Indonesian, Sundanese, Javanese, or Minangkabau—the languages covered by IndoBERT and XLR-R.

Machine Translation

We show the results on machine translation in Table 4 (idn x) and Table 5 (x idn) in SacreBLEU post-2018-call. As some local languages are similar to Indonesian, we observe that the Copy baseline (which does not do any translation) performs quite well. Since these local languages also share similar grammatical structure with Indonesian, when we perform dictionary-based substitution (Word Substitution) we also see an improvement.

Both PBSMT and fine-tuned LMs demonstrate encouraging performance despite the limited training data, which we attribute to the target languages’ similarity to Indonesian. We generally see a good translation performance across local languages. Thus, there is an opportunity to utilize translation models to create new synthetic datasets in local languages via translation from a related high-resource language, not only for Indonesian local languages but also other underrepresented languages. However, note that even for language pairs where BLEU is very high, we observe translation deficiencies stemming from the small amount of training data: rare words may just be copied with PBSMT, and mistranslated with NMT.

5.2 Cross-lingual Capability of LMs

From a linguistic perspective, local languages in Indonesia share similarities based on their language family. Many local languages share a similar grammatical structure and have some vocabulary overlap. Following prior works that show positive transfer between closely related languages  winata2021language; cahyawijaya-etal-2021-indonlg; hu2020xtreme; khanuja2020gluecos, we analyze the transferability between closely related languages in the Malayo-Polynesian language family.

Empirically, Figure 2 shows the cross-lingual capability of the best performing model (XLM-R) in the zero-shot cross-lingual setting for sentiment analysis. In general, most languages, except for Buginese (bug) and Toba Batak (bbc), can be used effectively as the source language, reaching on average 70-75% F1 compared to 80% average F1 in the monolingual setting (cf. XLM-R in Table 3). This empirical result aligns with the fact that both Buginese (bug) and Toba Batak (bbc) possess a very low vocabulary overlap with Indonesian aji-etal-2022-one. Interestingly, despite coming from a completely different language family, English can also be effectively used as the source language for all 10 local languages.

In addition, we observe high BLEU scores across languages in the MT task for Banjarese (bjn) and Minangkabau (min) languages despite not being included in XLM-R’s pre-training data. This may be due to Banjarese being similar to Malay nasution-2021-plan while Minangkabau shares some words and syntax with Indonesian koto-koto-2020-towards, suggesting positive transfer between closely related languages in the Malayo-Polynesian language family.

We can take advantage of language similarity by transferring knowledge from Indonesian and other local languages to perform zero-shot or a few-shot classification in closely related languages. New datasets for underrepresented languages that are closely related to high-resource languages thus do not necessarily need to be large, which may make the development of NLP datasets in low-resource languages more affordable than it may appear initially.

Figure 2: Zero-shot cross-lingual task results in the sentiment analysis task with XLM-R. The model is trained with the language data on the x-axis and evaluated in all languages. The score represents the average among all languages.

6 Related Work

Multilingual Parallel Corpus

Several multilingual parallel corpora have been developed to support studies on machine translation such as: the Global Communication Plan (GCP) corpus imamura2018multilingual, Leipzig corpora goldhahn-etal-2012-building, JRC Acquis steinberger2006jrc, TUFS Asian Language Parallel Corpus nomoto2018tufs, Intercorp ek2012case, DARPA LORELEI strassel2016lorelei, Asian Language Treebank riza2016introduction, FLORES guzman2019flores, the Bible Parallel Corpus resnik1999bible; black2019cmu, and JW-300 agic-vulic-2019-jw300. The languages covered in the GCP corpus are Japanese, English, Spanish, French, and several Asian languages including Indonesian. While the GCP corpus covers sentence pairs in tourism related domains, JRC Acquis contains mostly legal documents in 20 official EU languages. Similar to the GCP corpus, TUFS Asian Corpus (TALPCo) covers Japanese sentences and their translations to English, Burmese, Malay, and Indonesian. Intercorp is a multilingual corpus based on Czech covering 27 other languages. DARPA LORELEI and the Asian Language Treebank cover several low resources languages, while FLORES contains sentences in two low resource languages, parallel with English. Building high-quality parallel corpora is expensive and time-consuming, since it requires translation by native or fluent translators. guzman2019flores describe the procedure to generate high-quality translations as part of FLORES. The translation quality checks consist of two types of filtering: automatic and manual. Similar to FLORES, we also conducted quality checks of the translation.

Emerging Language Benchmarks

Recently, datasets and benchmarks in underrepresented languages have emerged, such as MasakhaNER tacl_masakhaner, AmericasNLI americasnli, PMIndia haddow2020pmindia, Samanantar tacl_samanantar, and NaijaSenti arxiv-najjasenti. MasakhaNER covers ten African languages, while AmericasNLI covers ten indigenous languages of Americas and NaijaSenti covers four Nigerian languages. Particularly, for Indonesian languages, we have seen NLP benchmarks, such as IndoNLU wilie2020indonlu, IndoLEM koto2020indolem, IndoNLG cahyawijaya-etal-2021-indonlg, and IndoNLI mahendra2021indonli. Other task-specific benchmarks include En-Id machine translation guntara2020benchmarking, paraphrasing aji2021paracotta, and low-cost sequence classification and labeling in nityasya2021costs; nityasya2022student. These benchmarks mostly focus on Indonesian, except for IndoNLG, which also includes Javanese and Sundanese. PMIndia covers 13 Indic languages where each language pair (English with one of languages of India) contains a maximum of 56,000 sentence pairs, sourced from the website of India’s prime minister. Samanantar covers 11 Indic languages, with a total of 49.7 million sentence pairs where one source corpus is PMIndia.

Datasets for Indonesian Languages

Only a limited number of labeled datasets exist for local languages in Indonesia. WikiAnn pan-etal-2017-cross

, a weakly supervised named entity recognition dataset, covers Acehnese, Javanese, Minangkabau, and Sundanese.

Putri_etal_2021_abusive built a multilingual dataset for abusive language and hate speech task involving Javanese, Sundanese, Madurese, Minangkabau, and Musi languages. Few datasets exist for individual languages, e.g., sentiment analysis and machine translation in Minangkabau koto-koto-2020-towards and emotion classification in Sundanese putra2020sundanese. Finally, some datasets focus on colloquial Indonesian mixed with local languages in the scope of morphological analysis wibowo2021indocollex and style transfer wibowo2020semi.

7 Conclusion

In this paper, we propose NusaX, the first parallel corpus for 10 low-resource Indonesian languages. We create a new benchmark of sentiment analysis and machine translation in few-shot and full-data settings. We present a comprehensive analysis of the language similarity of these languages from both linguistics and empirical perspective by assessing the cross-lingual transferability of existing Indonesian and multilingual pre-trained models.

8 Ethical Considerations

We cover low-resource languages which increase the accessibility of NLP research towards marginalize community. NusaX is created based on sentiment sentences taken from social media, reviews, and forums, therefore may contain biases towards certain groups or entities. Similarly, as our dataset was translated, there may be some translationese artifact on the resulting corpus. We argue that further study on the potential bias and the translationese effect of our dataset is needed.

Most of our best-performing baselines are based on large language models which require massive computational power to train and deploy. This computational cost is more significant given the lack of resources for local communities. We provide alternative classical-based approaches that are significantly more efficient, which still achieve competitive results.


We thank Dea Adhista and all annotators who help us in building the corpus. We are grateful to Alexander Gutkin for feedback on a draft of this manuscript. This work has been partially funded by (001/SD/YGI-NLP/1/2022) and PF20-43679 Hong Kong PhD Fellowship Scheme, Research Grant Council, Hong Kong.


Appendix A Data Statement for NusaX

a.1 General Information

Dataset title


Dataset curators

Genta Indra Winata (Bloomberg), Alham Fikri Aji (Amazon), Ade Romadhony (Telkom University, Indonesia), Rahmad Mahendra (Universitas Indonesia), Fajri Koto (University of Melbourne), Samuel Cahyawijaya (Hong Kong University of Science and Technology), Kemal Kurniawan (University of Melbourne)

Dataset version

1.0 (May 2022)

Dataset citation


Data statement author

Kemal Kurniawan (University of Melbourne)

Data statement version

1.0 (February 2022)

Data statement citation

Kurniawan, Kemal. (2022). Data Statement for NusaX. Version 1.0. University of Melbourne. TODO url.

a.2 Executive Summary

NusaX is a multilingual parallel corpus across 10 local languages in Indonesia: Acehnese, Balinese, Banjarese, Buginese, Madurese, Minangkabau, Javanese, Ngaju, Sundanese, and Toba Batak. The data was translated obtained by human translation from Indonesian and human-assisted quality assurance.

a.3 Curation Rationale

The goal of the dataset creation process is to provide gold-standard sentiment analysis corpora for Indonesian local languages. The Indonesian data is sampled from SmSA purwarianti2019improving, an Indonesian sentiment analysis corpus. SmSA is chosen among other corpora (e.g., HoASA azhar2019multi based on (1) the agreement of our manual re-annotation of a small and randomly selected samples and (2) manual inspection to ensure that the topics are diverse. After sampling, the data is edited and/or filtered to remove harmful contents and maintain quality. Several criteria are used in this process:

  1. [noitemsep]

  2. Is the sentiment label correct?

  3. Does the sentence contain multiple sentiments?

  4. Does the sentence contain harmful content that discriminates against race, religion, or other protected groups?

  5. Does the sentence contain an attack toward an individual or is abusive?

  6. Is the sentence politically charged?

  7. Is the sentence overly Bandung/Sunda-centric?333Bandung is the capital city of West Java, in which Sunda is the ethnic group.

  8. Will the sentence be difficult to translate into local languages?

  9. Are there any misspellings?

a.4 Documentation for Source Datasets

NusaX is obtained by translating SmSA purwarianti2019improving, an Indonesian sentiment analysis dataset.

a.5 Language Variety

NusaX covers a total of 10 local languages spoken in Indonesia (ID) as shown in Table 6.

Language ISO 639-3 Dialect
Acehnese ace Banda Aceh
Balinese ban Lowland
Toba Batak bbc Toba, Humbang
Banjarese bjn Hulu, Kuala
Buginese bug Sidrap
Javanese jav Matraman
Madurese mad Situbondo
Minangkabau min Padang, Agam
Ngaju nij Kapuas, Kahayan
Sundanese sun Priangan
Table 6: Local languages spoken in Indonesia (ID) that are covered in NusaX.

a.6 Speaker Demographic

The SmSA dataset was obtained from social media and online forums: Twitter, Zomato, TripAdvisor, Facebook, Instagram, Qraved. We can assume the users’ age ranges from 25 to 34 years, which is the age range of the majority of Indonesian social media users444

a.7 Annotator Demographic

A total of 28 translators are employed in the translation process. All translators are Indonesian and recruited by via either online surveys or personal contacts. They are then selected based on (1) the self-reported fluency in the local language into which they would be translating and (2) the highest education level achieved. Those who (a) are native speakers of or fluent in the target local language and (b) finished at least high school education (id: SMA/sederajat) are selected.


There are 3 translators for Acehnese, but only 2 of them responded when asked for demographic information. Thus, what follows is the demographic information of only those 2 translators. One has some experience in translation work, while the other does not. One identifies as male, and the other as female. Both are in their 20s. Lastly, one works as a freelancer, while the other is a farmer.


Three people translate into Balinese. Two of them have previous experience in translation work, and both identify as female. The other one, who identifies as male, does not have such experience. Two of them are aged 20-29 years old, while the other is in their 30s. Their occupations are university lecturer, school teacher, and civil employee respectively.


Two translators are employed for Banjarese, but only one responded when asked for demographic information. The translator has prior experience in translation work, identifies as male, is in his 40s, and works as a university lecturer.


Buginese is translated by 2 people, but only one responded when asked for demographic information. The person has prior translation experience, identifies as male, is aged 30-39 years old, and runs an Islamic boarding school as a living.


Four translators are employed for Javanese, but one did not respond when asked for demographic information. The other three have prior experience in translation work. Among them, two identify as female, and one as male. All of them are in their 20s. Two of them are university students, and the other one works as a freelance assistant editor.


There are 3 translators for Madurese. Only one of them has previous experience in translation work. Two of them identify as female, while the other as male. One person is aged under 20 years old and is a university student. The others are 20-29 years old and work as a school teacher and an employee in a private company respectively.


Three people translate into Minangkabau. Two of them have previous translation experience. All three identify as female and are aged 20-29 years old. They work as a civil employee, a university student, and a senior data annotator respectively.


Two translators work on Ngaju, but only one responded when asked for demographic information. The translator has prior experience, identifies as female, is aged no less than 50 years old, and is a stay-at-home mother.


There are 5 translators for Sundanese, four of which identify as female, and the other one as male. Three translators are in their 20s, one is younger than 20 years old, and the remaining one is in their 30s. The translators work as a school teacher, a university student, a university lecturer, and the remaining two as employees in a private company.

Toba Batak

Three translators are employed for Toba Batak. One has prior translation experience. Two translators identify as male while the other as female. All three are in their 20s. One works for a private company, and the others are university students.

Appendix B Hyperparameters

b.1 Sentiment Analysis

Hyperparams NB SVM LR
feature {BoW, tfidf} {BoW, tfidf} {BoW, tfidf}
alpha (0.001 - 1)
C (0.01 - 100) (0.001 - 100)
kernel {rbf, linear}
Table 7: Hyperparameters of statistical models on sentiment analysis.
Hyperparams Values
learning rate [1e-4, 5e-5, 1e-5, 5e-6, 1e-6]
batch size [4, 8, 16, 32]

num epochs

early stop 3
max norm 10
optimizer Adam
    Adam (0.9, 0.999)
    Adam 0.9
    Adam 1e-8
Table 8: Hyperparameters of pre-trained language models on sentiment analysis. Bold denotes the best hyperparameter setting.

For statistical models, we use a spaCy as our toolkit, and we perform grid-search over the parameter ranges shown in Table 7 and select the best performing model over the devset. For all pre-trained language models, we perform grid-search over batch size and learning rate while keeping the other hyperparameters fixed. The list of hyperparameters is shown in Table 8.

b.2 Machine Translation

Table 9 shows the hyperparameters of deep learning models on machine translation.

Hyperparams IndoGPT IndoBARTv2 mBART-50 mT5
learning rate 1e-4 1e-4 2e-5 1e-3
batch size
max epochs
early stop
Table 9: Hyperparameters of pretrained language models on machine translation.

Appendix C Examples

In Table 10, we provide several examples of the translated parallel data of NusaX corpus.

Language Text
Indonesian Ratusan rumah di medan terendam banjir
English Hundreds of houses in Medan were submerged
by the flood
Acehnese Meureutoh rumoh di Medan keunong ie raya
Balinese Satusan umah ring medan merendem banjir
Banjarese Ratusan rumah di medan tarandam banjir
Buginese Maddatu bola okko medan nala lempe
Javanese Atusan omah ing medan kebanjiran
Madurese Ratosan bangko e medan tarendem banjir
Minangkabau Ratuihan rumah di medan tarandam banjir
Ngaju Ratusan huma hong medan lelep awi banjir
Sundanese Ratusan bumi di medan karendem banjir
Toba Batak Marratus jabu di medan na hona banji
Table 10: Example of the translation data.