The culture and the social structure of a community are reflected in the language spoken by its members. One of the most salient examples of this phenomenon is the worldwide diversity of terms used to describe family structures and relationships. While, thanks to studies such as Murdock (1970), kin terms around the globe are generally well-documented, many local variations—across dialects of a single language or across languages of a single country—have not yet been fully described or understood. For example, the term مَعزوزي maazoozi in the Algerian Arabic dialect, meaning younger brother, does not have any equivalent term in the Gulf Arabic dialect. In contrast, the Gulf word ابن العُود ibn alood meaning elder brother does not exist in Algerian, which instead uses the word سِيدي siedi.

Beyond a linguistic or anthropologic interest, the availability of digital resources on language diversity is also desirable from a computational perspective. Language processing applications need to be aware of such phenomena of diversity in order to provide high-quality results. For example, a machine translation system needs to tackle cases of lexical untranslatability where a word or expression in a source language has no equivalent in a given target language, and the choice of an approximate translation can change the meaning of an utterance. For example, for the English sentence his cousin gave birth to a twin, Google Translate provides the Arabic translation أنجب ابن عمه توأما a'njaba ibna a'mihi tawaman that means His father's brother's son gave birth to a twin. This syntactically correct yet unintended meaning of a male giving birth output is due to a lexical gap, i.e., a non-existent equivalent Arabic term for cousin. Such cases of techno-linguistic bias—where language technology provides better results by design in certain languages than in others—tend to remain hidden in monolingual resources but are revealed in multilingual settings (Bella et al., 2022a, 2023).

In recent years, there has been an increasing number of linguistic databases covering a large number of languages. These resources are usually aimed at quantitative studies for comparative linguistics, such as the classification of pain predicates (Reznikova et al., 2012), a semantic map of motion verbs (Wälchli and Cysouw, 2012), the modeling of color terminology (McCarthy et al., 2019), the CLICS database of cross-linguistic colexifications (Rzymski et al., 2020), DiACL (Diachronic Atlas of Comparative Linguistics), a database for ancient Indo-European languages spoken in Eurasia typology (Carling et al., 2018), or the Cross-Linguistic Database of Phonetic Transcription Systems (Anderson et al., 2018). Often, such databases use phonetic representations of lexical units or are limited to a few hundred or a few thousand core concepts, limiting their usability for the processing of contemporary written language. In our experience, most of the existing typology-informed NLP research is restricted to exploring language-specific morphosyntactic features and has ignored diversity within lexical resources (Batsuren et al., 2022). A notable exception is the Universal Knowledge Core, a massively multilingual lexical database that explicitly represents linguistic diversity and that we reuse in our work.

Our research is part of the LiveLanguage initiative, the overarching objective of which is to create, publish, and manage language resources that are “diversity-aware”—i.e., that reflect the viewpoints of multiple speaker communities—and that can be reused by multiple communities: linguists, cognitive scientists, AI engineers, language teachers and students (Bella et al., 2023). Contrary to mainstream exploitative practices, LiveLanguage aims to carry out its goals while empowering local speaker communities, giving them control over resources they help to produce (Helm et al., 2023). Involving human contributors and deciders from speaker communities is therefore a crucial part of our methodology.

In particular, the present paper focuses on diversity where it is less expected to appear: within dialects of the same language and within languages of the same country. Therefore, we describe a multidisciplinary study on the diversity of kin terms across seven Arabic dialects (Algerian, Egyptian, Tunisian, Gulf, Moroccan, Palestinian, and Syrian) and three languages from Indonesia (Indonesian, Javanese, and Banjarese). We consider kin terms as a domain particularly well-suited both for research on the methodology of collecting and producing diversity-aware linguistic data, and for comparative studies on diversity across languages.

Our paper aims to provide four contributions: (1) a general method for collecting multilingual lexical data from native speakers for a given domain (in our case the domain of kin terms), in a diversity-aware manner; (2) 223 kin terms and 1,619 lexical gaps collected in seven Arabic dialects and three Indonesian languages; (3) a qualitative and quantitative discussion of our results regarding the diversity observed across the dialects and languages covered; and (4) the publication of our results as an open, computer-processable dataset, as well as its integration into the Universal Knowledge Core multilingual database. Our starting point is state-of-the-art datasets on worldwide kinship terminology from ethnography (Murdock, 1970) and computational linguistics (Khishigsuren et al., 2022). Our data collection method is based on collaborative input from native speakers and language experts. Our results extend the state-of-the-art resources above with kin terms in languages and dialects not yet covered, as well as with 22 new kinship concepts not yet associated with other languages within those resources.

The structure of the paper is organized as follows. In Section 2, we give an overview of lexical typology and the phenomena of lexical untranslatability and lexical gaps with respect to the domain of kinship in particular. The Universal Knowledge Core resource is presented in Section 3. In Section 4, we describe our data collection method. Sections 5 and 6 introduce two case studies on Arabic dialects and Indonesian languages, respectively. Section 7 discusses previous studies related to our work. Finally, we provide conclusions in Section 8.

Linguists understand translation from one language to another as a complex and multidimensional problem, ranging from multiple coexisting forms of meaning equivalence to untranslatability (Catford, 1965; Bella et al., 2022a). The diversity between cultures is a major cause for this problem appearing on several lexical-semantic levels. Some examples of the linguistic diversity are the richness of Toaripi vocabulary on the various forms of motion verbs describing walking around the beach like (isai) meaning “go beachward” and (kavai) meaning “go inland with respect to the beach”, the language of the coastal Papua New Guinea country, the lack of vocabulary for the word meaning “sailing” in Mongolian, which is the language of a landlocked country, or the Arabic word تَسنَّم meaning “to ascend a camel's hump”.

The domain of kinship terms, which is the subject of our paper, is known to be extremely varied across languages, due to the different ways family structures are organized around the world. Matriarchal societies may describe certain female relatives with more detail, while strongly patriarchal ones are more descriptive with respect to male relatives. Arabic dialects, for instance, distinguish paternal and maternal brothers but also blood brothers, full brothers, and breastfeeding brothers. Thus, not only are kinship-related vocabularies “richer” or “poorer” across languages, they are also structured in different manners.

In this research, we focus on lexical untranslatability, which manifests most clearly through the lexical gap phenomenon when a word in a source language does not have a concise and precise translation in a given target language. Lexical gaps are often the linguistic manifestation of culturally or spatially defined specificities of a community of language speakers that cannot entirely be predicted or explained through systematic principles or recurrent patterns (Lehrer, 1970). Table 1 below presents this phenomenon for nine concepts representing sibling relationships from the kinship domain in eight languages.¹ One can observe that none of the eight languages has concise lexicalizations for all nine concepts, yet each concept is lexicalized in at least one language. Such variations in lexicalization pose a problem for both machine and human translation: for instance, substituting a specific term instead of a broader one may result in injecting unintended meaning. In Javanese, at least four specific terms—(sedulur/sibling), (adhi/younger sibling), (kangmas/elder brother), and (Mbakyu/elder sister)—are used for expressing the sibling relationship, and accordingly, translating this sentence through Google Translate (my sister is ten years older than me) to Javanese gives this non-sensical sentence (adhiku luwih tuwa sepuluh taun tinimbang aku) meaning (my younger sibling is ten years older than me). This result is due to the lack of Javanese vocabulary for the word meaning (sister), and also lacks the term meaning “younger sister”, so the machine translator uses (adhi) meaning “younger sibling,” which finally produces the semantically absurd output.

Lexical typology is a field of linguistics that studies the diversity across languages according to the structural features of languages with respect to specific semantic fields (Plungyan, 2011). Different classical studies are conducted in this field on grammar and phonology, such as VoxClamantis V1.0–a large-scale corpus for phonetic typology (Salesky et al., 2020) and the structure of the space semantic field by identifying a set of semantic parameters and notions depending on the grammatical information of the field's constituents (Levinson and Wilkins, 2006). Other examples of such studies have been conducted on lexical-typological issues that appear across languages during translation, like the presence or absence of lexicalizations in languages. In these articles, authors focused on semantic fields that offer the richness of cross-lingual diversity: family relationships (Kemp and Regier, 2012), colors (Roberson et al., 2005), food (Bella et al., 2022b), body parts (Wierzbicka, 2007), putting and taking events (Kopecka and Narasimhan, 2012), cutting and breaking events (Majid et al., 2007), or cardinal direction terms (Arora et al., 2021). However, as mentioned in the introduction, only a few open datasets have been published in the scientific research area. These include the classification of kinship by Murdock (1970), which has been published in D-PLACE (Kirby et al., 2016). Part of Kay and Cook (2016)'s work on colors is published under the lexicon chapter of the World Atlas of Language Structures (WALS) (Dryer and Haspelmath, 2013). Additionally, a color categorization dataset by McCarthy et al. (2019) is available on GitHub².

Digital lexicons have been increasingly used in lexical typology, enabling typologists to explore a broader range of languages and semantic domains. One noteworthy example is the KinDiv³ lexicon (Khishigsuren et al., 2022), which encompasses 1,911 words and identifies 37,370 gaps within the domain of kinship, spanning 699 languages. In our current research, we extend our investigation into the kinship domain, specifically focusing on exploring linguistic diversity among Arabic dialects and Indonesian languages. Other examples include Viberg (1983)'s seminal study, which was conducted on perceptual terminology in 50 languages and has been expanded upon by Georgakopoulos et al. (2022) to cover 1,220 languages. Furthermore, the Kinbank database, recently introduced by Passmore et al. (2023), serves as a comprehensive repository of kinship terminology, encompassing more than 1,173 languages and offering a broad coverage of various kinship subdomains.

This section describes the Universal Knowledge Core (UKC)⁴, a large multilingual lexical database that we adopt for the production of diversity-aware datasets in this research (Giunchiglia et al., 2017). The use of the UKC is motivated by its ability to represent linguistic unity and diversity explicitly: conceptualizations shared across languages, word senses appearing only in certain languages, shared lexicalizations (e.g., cognates), as well as lexical gaps. The theoretical underpinnings of the lexical model of the UKC have been described in Giunchiglia et al. (2018) and in Bella et al. (2022b), and are illustrated in Figure 1.

The UKC is divided into a supra-lingual concept layer (as shown at the top of Figure 1) and the layer of individual lexicons (at the bottom of Figure 1). The concept layer includes hierarchies of concepts that represent lexical meaning shared across languages. Concepts are language-independent units and act as bridges across languages, and each one should be lexicalized by at least one language to be present in the concept layer. Supra-lingual concepts and their relations (e.g., hypernymy, meronymy) are in part derived from third-party resources such as Princeton WordNet (PWN) (Miller, 1995), and are in part proper to the UKC. In particular, the UKC contains an extensive formal conceptualization of kinship domain terms computed from the KinDiv database, spanning about 200 distinct concepts.⁵ KinDiv itself is based on ethnographic evidence from 699 languages (Khishigsuren et al., 2022). While this existing hierarchy of kinship concepts does not fully cover all terms that appear in our study, it is the most complete one we are aware of, motivating our choice of the UKC as a platform for our research.

The lexicon layer consists of language-specific lexicons that provide lexicalizations for the concepts from the supra-lingual concept layer, while also asserting lexical gaps whenever lexicalizations are known not to exist. Lexicons also provide term definitions as well as lexical relationships specific to the language, such as derivations, metonymy, or antonymy relations. Lexicons can also contain language-specific concepts that do not appear in the supra-lingual concept layer. For example, in Figure 1, the Arabic شَقِيقة, meaning “a female person who has the same father, mother, or both parents as another person”, is represented as a language-specific concept. The dual mechanism of defining lexical concepts either on the supra-lingual or on the language-specific level allows for the representation of differing worldviews that would be hard or impossible to reconcile into a single global concept graph. The richness of its lexicon-level linguistic knowledge makes the UKC unique among multilingual lexical databases and particularly suitable for our study.

As mentioned in Section 2, a lexical gap for a specific concept is present in a language if there is no concise equivalent word meaning for the concept in that language. For example, neither English nor Arabic has a word meaning elder sibling; for such cases, the UKC provides evidence of meaning non-existence and untranslatability by representing lexical gaps inside lexicons, as shown in Figure 1. This information can be used by the NLP community to indicate the absence of equivalent words to downstream cross-lingual applications.

Beyond providing lexical relations between shared word meanings as other multilingual lexical databases do, the UKC also represents a richer set of lexical-semantic connections between language units in a lexicon. For example, the antonym lexical relation expresses that two senses are opposite in meaning. While the lexical-semantic relation, similar-to, is used to connect two concepts with similar meanings, and the hypernym-of connects parent meaning with its child. For instance, in Figure 1, the English (little brother) and (brother) are connected through a hypernym-of relationship. Such information can be used by the NLP community to indicate the concise equivalent language-specific word meaning to downstream cross-lingual applications, e.g., as the position of a language-specific meaning in a language hierarchy in a lexicon.

The UKC currently does not explicitly distinguish between languages and dialects: each vocabulary is a separate entity labeled with a standard three-letter ISO 639-3 code. When such a code is not available, the UKC uses a standard extension mechanism where three additional (not standardized) letters are added to the ISO code: e.g., for Syrian Arabic, the code arb-syr is used.

This section presents the general method by which we collected and produced lexicalizations and gaps from native speakers and language experts. The same method presented below was employed in an independent manner for each Arabic dialect and Indonesian language covered by our study. The contents of this section aim to serve as a tried and tested recipe for gathering lexical data in a diversity-aware manner, that we intend to reuse in future lexicon development projects.

We exploit the UKC to import language-independent concepts (e.g., kinship concepts) to be used as an input dataset to our method and use its data representation model to formalize our data. We reuse an already broad and well-formalized hierarchy of 184 kinship concepts from the KinDiv database, which includes kinship terms and gaps in 699 languages. Data in KinDiv is based on the well-known results of Murdock (1970), as well as on lexicalizations retrieved from Wiktionary that we consider as an overall good-quality resource. In Khishigsuren et al. (2022), the accuracy of KinDiv was evaluated to be above 96%. One language expert per language provided this percentage, which represents the proportion of the number of words (or gaps) validated as correct to the total number of collected words (or gaps).

Our work extends KinDiv data by new concepts, lexicalizations, and lexical gaps in languages and dialects that are either not present in KinDiv or are incompletely covered. A lexical-semantic expert generates a contribution (kinship terms and gaps) task, then a group of native speakers collects contributions from a dialect (and a local language). After that, two steps for validating collected contributions: language experts evaluate collected lexical units and gaps of a dialect, and a lexical-semantic expert evaluates explored kinship concepts (not existing in UKC). Additionally, resulting data (including gaps, words, and new concepts) is used to update and enrich UKC. So, gaps and words are merged into the lexicons of the UKC while new concepts are integrated with the (top) concept layer. A general view of the method is depicted in Figure 2.

Accordingly, the macro-steps of our methodology are as follows:

This section demonstrates the use of the methodology described in Section 4 on kinship terminology from seven dialects of the Arabic language. Arabic is the official language of more than four hundred million native speakers in twenty-two countries in the Middle East and northern Africa. Classical Arabic or Modern Standard Arabic (MSA) refers to the standard form of the language used in academic writing, formal communication, classical poetry, and religious sermons (Elkateb et al., 2006). Surprisingly lexical diversity is manifested between Arabic dialects, evident in our study between seven of the twenty dialects spoken worldwide. The selected dialects are Egyptian, Moroccan, Tunisian, Algerian, Gulf, and South Levantine (two examples: Palestinian and Syrian). Let us take the example of the Gulf word الخَال العودْ meaning “mother's elder brother,” which has no equivalent in South Levantine or Moroccan; instead, they use the more general word الخَال meaning “mother's brother,” which can be used for both meanings “mother's younger brother” or “mother's elder brother”. In this paper, we perform an experiment on the Arabic dialects to capture their diversity in the kinship domain. The resulting dataset with dialect-specific kinship terms will be integrated with an instance of the Universal Knowledge Core for Arabic (Arabic UKC)⁶ ongoing project, which is the first diversity-aware lexical resource for Arabic dialects so far.

This section demonstrates the use of the methodology described in Section 4 on kinship terminology from three Austronesian languages from Indonesia: Indonesian, Javanese, and Banjarese. Contrary to the Arabic dialects in Section 5, these three languages are not mutually intelligible.

Indonesia is the fourth most populous country in the world, and it has more than 700 living languages (Eberhard et al., 2022). The national language spoken in Indonesia is Bahasa Indonesia/Indonesian language, which was decided in the historic moment of Youth's Pledge, October 28th, 1928. However, many Indonesians speak more than one language. For example, out of 198 million people that speak Indonesian, 84 million of them speak Javanese (Aji et al., 2022).

Even with the high number of speakers, the count of natural language processing research on Indonesian languages is very low compared to other languages around the world. As of 2020, the count of published papers on the Indonesian language is lower than other languages with less speaker count, such as Polish and Dutch (Aji et al., 2022). Not surprisingly, the amount of research on other languages (i.e., Banjarese and Javanese) in Indonesia is much lower than that. It is therefore motivating to conduct this study that discovers the richness of linguistic diversity across three Indonesian languages: standard Indonesian, Banjarese, and Javanese. In one semantic field, kinship, we have found that diversity is manifested in these languages; for example, in Javanese, the word ponakan jaler meaning “nephew”, is a lexical gap in Banjarese, and in the opposite direction, the Banjarese gulu meaning “parent's second eldest sibling” is also a gap in Javanese.

Ethnologists and linguists have for a long time studied how family structures map to kinship terminology across languages and social groups. The most famous and comprehensive ethnographic study on kin term patterns is that of Murdock (1970), upon which our work also indirectly relied: our cross-lingual formalization of kin terms is based on the one provided by the KinDiv resource, itself in part derived from Murdock's data. KinDiv covers 699 languages and is a computer-processable database that can also be exploited for applications in computational linguistics. Our results provide linguistic evidence in seven Arabic dialects and three Indonesian languages that do not figure in these resources.

The exploration of kin terminology and the building of large-scale databases on the topic has also been the subject of more recent efforts—we only cite two examples here. The AustKin project¹³ has produced a large-scale database on kin terms in hundreds of indigenous Australian languages. The recent Kinbank database (Passmore et al., 2023) is a comprehensive resource on kinship terminology, covering over 1,173 languages, with a broad coverage of kinship subdomains. As Kinbank was released after the initial submission of our paper, we did not rely on it for our work. We consider our research as complementary to Kinbank: concentrating on a relatively low number of dialects and languages, our results could, in principle, be integrated into Kinbank in order to extend its coverage. And vice versa, we see potential in using Kinbank data in order to cross-validate and possibly to extend the Indonesian terms we collected (as the three Indonesian languages of our study are also covered by Kinbank). There is, however, an important methodological difference between the our and Kinbank's way of representing terms: Kinbank does not explicitly indicate lexical gaps. For example, our work considers the concept of son of father's brother as pronounced by a male speaker to be a lexical gap in Javanese, while Kinbank maps the Javanese term sedulur misan, simply meaning cousin, to this and 95 other meanings. Our work, instead, identifies the Javanese term as the general meaning of cousin and considers all other (more specific) cousin terms as lexical gaps. This distinction is useful in comparative linguistics and cross-lingual applications where the explicit indication of the lack of precise meaning equivalence can be exploited.

Concepticon (List et al., 2016) is ‘a resource for linking concept lists' frequently used in comparative linguistics. The concept sets of Concepticon serve the same purpose as the supra-lingual concepts of the UKC in our study, namely to provide meaning-based mappings among lists of terms (aka concept lists in Concepticon) across languages. As of mid-2023, Concepticon consists of nearly 4,000 concept sets, principally targeting core vocabularies (basic-level categories) that are the main subject of study of historical and comparative linguistics. Concepticon is under continuous development and has more recently evolved from a flat list of meanings to a hierarchy with broader–narrower relations. At the time of writing, the kinship domain seems to be partially represented in Concepticon: while sibling or grandparent relations are widely covered, fine-grained cousin relationships are mostly missing from it. The UKC, which contains over 100,000 supra-lingual concepts and a wide range of lexical and lexico-semantic relations, was a more suitable resource for our study due to its more complete coverage of the kinship domain and its explicit support for representing term untranslatability via lexical gaps.

Multilingual computational applications being in the core of our focus, we also review relevant resources from computational linguistics. For NLP applications, the most popular and widely-known representation of lexico-semantic knowledge is that of wordnets that follow the general structure of the original English Princeton WordNet (Miller, 1995). The wordnet expansion approach by Fellbaum and Vossen (2012)—an expert-driven lexicon translation effort—is frequently used to produce new wordnets for lower-resourced languages: this approach consists of ‘translating' (i.e., finding lexicalizations for) English WordNet concepts (‘synsets' in wordnet terminology) into the target language. While this is a straightforward approach that produces resources that remain cross-lingually linked, its downside is that the translation approach cannot involve concepts and words specific to the target language and not present in the source language (which in most cases is English). In cases of diverse conceptualizations of the world, the translation approach often results in incorrect approximations. To take the example of Arabic, both versions of the Arabic Wordnet (Elkateb et al., 2006; Abouenour et al., 2013) map the English synset of uncle (“the brother of your father or mother; the husband of your aunt”) to the Arabic synset of عَمْ, which means “the brother of your father.”

A similar situation is observed for Indonesian. As far as we know, the only Indonesian Wordnet currently accessible is Bahasa Wordnet—a bilingual Wordnet for standard Indonesian and Malay languages (Noor et al., 2011). It was formed by merging three different wordnets (one in Indonesian and two in Malay) developed mainly by the same expansion approach from PWN. Due to this approach, many English words that have no equivalents in Indonesian are incorrectly mapped, resulting in meaning loss. For example, in Bahasa Wordnet, the English word sister, which means “a female person who has the same parents as another person,” was mapped to the Indonesian word kakak which means “elder sibling.”

Finally, we mention MultiWordNet as an early effort at improving the representation of linguistic diversity in multilingual lexical databases (Pianta et al., 2002). It is a multilingual lexicon that was built using the merge method that, contrary to the translation-based expand approach presented above, maps together existing high-quality bilingual dictionaries. MultiWordNet explicitly represents lexical gaps in its Italian and Hebrew wordnets: about 1,000 in Italian and about 300 in Hebrew (Bentivogli and Pianta, 2000; Ordan and Wintner, 2007). MultiWordNet, however, is a discontinued effort that does not cover the kinship domain and is thus was not suitable for our purposes.

The methodology we present in Section 4 follows neither the expansion nor the merge approach but a third one, more adapted to diversity-aware lexicography: our starting point is a supra-lingual, diversity-aware conceptualization of the domain of study (kinship in our case). The task of contribution collection is performed by native speakers with respect to the supra-lingual concept hierarchy based on evidence from comparative linguistics and covering a wide range of languages. While there is no guarantee that our initial conceptualization is complete—indeed, it was not the case in our study—it is less biased toward the concepts of a single language and speaker community than the expansion approach.

Our paper formally captures lexical diversity across languages and dialects by representing language- or dialect-specific concepts and linguistic gaps. It introduces a systematic method to produce such data in a human-based manner from one semantic domain rather than from general domains, as the efforts of covering the WordNet domains (Magnini and Cavaglià, 2000) that have been conducted in building these wordnets, Mongolian (Batsuren et al., 2019), Unified Scottish Gaelic (Bella et al., 2020), and MultiWordNet (Pianta et al., 2002).

The method is verified through two large-scale case studies on kinship terminology, a domain known to be diverse across languages and cultures: one case study deals with seven Arabic dialects, while the other one with three Indonesian languages. The experiments show that our method outperforms the existing methods in terms of the quantity of explored gaps and words and the quality of results. Overall efforts resulted in 1619 gaps, and 223 words were identified in 10 languages and dialects. Moreover, 22 new word meanings with respect to the imported list of independent-language concepts from the UKC are explored in this research.

In future work, we plan to automate the method presented in this paper and apply it to new languages, such as the rest of the Arabic dialects and Indonesian language, as well as to new domains that are known to be diverse, such as body parts, food, color, or visual objects (Giunchiglia and Bagchi, 2021; Giunchiglia et al., 2023).

Finally, diversity-aware lexicons such as the UKC (which includes our produced datasets) provide essential information to cross-lingual applications, such as multilingual NLP tasks or cross-lingual language models. In the future, we plan to use this resource in implementing one such application, i.e., machine translation.

The original contributions presented in the study are publicly available. This data can be found here: https://github.com/HadiPTUK/kinship_dialect.

FG and GB conceptualized and supervised the study. GB and HK imported and formatted the dataset of inputs. HK wrote the original manuscript draft and performed the Arabic experiments. AF and HK validated the collected Arabic data at the lexicon level. SD performed the Indonesian experiments and validated the results at the lexicon level. GB validated the identified diverse data at the concept level. FG, GB, AF, and HK analyzed the Arabic and Indonesian data. FG, GB, AF, SD, and HK reviewed and edited the manuscript. All authors contributed to the research and approved the submitted version.