Multilingual Knowledge Graphs and
Low-Resource Languages: A Review

There are, depending on the particular KG, different ways to store multilingual information [29]. A common choice is to assume that KG entities can be shared across languages, and thus for a given KG entity, one can provide relevant language-specific information such as labels in multiple languages. An example of this is illustrated in Figure 1 for the prominent Wikidata KG. An alternative is to essentially create a different KG (or sub-KG) for each language and then interlink these KGs. An example of this is illustrated in Figure 2 for the DBpedia KG.

The availability of labels can differ substantially across languages. Of course, this may simply be due to incomplete coverage, as we discuss in Section 2.3. In some cases, however, this may stem from linguistic differences in lexicalization or naming: A concept or entity may possess a name in one language but not in another, in which case it is sometimes referred to as a lexical gap. For instance, Runyankore does not have a single word for “pet” but only a description, and isiZulu’s ingcula does not have a translation into English other than a description of the object (a small bladed hunting spear). At the instance level, additional modelling efforts may be required, if Runyankore-speaking people were to be interested in representing in the KG that Lassie and Scooby Doo are pets, say, or for the development of a KG about Shaka Zulu’s armoury.

In the literature, there is a range of guidelines indicating how a KG should represent multilingual information. These include the use of standard ontologies, recommendations for the data itself as well as how the data should be modelled. Across the literature, the following guidelines have been identified, as also summarised in previous work [44].

• $\mathrm{■}$

  Stable identifiers [7]: Each entity is identified in the KG with an identifier (ID). These are crucial as the way to access information about an entity. Many authors recommend adopting Unique Resource Identifiers (URIs) or Internationalized Resource Identifiers (IRIs), as used on the Web, for better interoperability. Some KGs identify entities in the graph with identifiers incorporating natural language. For instance, DBpedia uses IRIs such as http://dbpedia.org/resource/Ada_Lovelace. Such identifiers are human-readable, and hence easy to interpret for humans, and they also take the function of a label. However, the fact that the natural language portions of such IDs carry meaning can also be a disadvantage. As identifiers are expected to remain stable, such IDs are unable to reflect potential changes in the entity label. For example, if the name of a property changes, the entire structure of the KG would be affected [85]. In contrast, opaque IDs are ones that are not easily readable by humans. For example, they could have a unique identifier for a concept in the form of letters and/or numbers, such as the Wikidata ID Q7259. As such IDs do not reflect the label of the entity, if the entity’s name changes, the ID is not affected. Moreover, such entity IDs can be more readily shared across languages.

• $\mathrm{■}$

  Label coverage [21, 90, 16]: As natural language labels are the way humans access information in the KG and interact with it, it is important that entities, classes, and properties in the KG are labelled. This ensures that information is human-readable and can be displayed to a user.

• $\mathrm{■}$

  Language tags [78]: When labelling an entity, it is important to indicate the language in which this label is provided. Even in monolingual KGs, language tags can be valuable, in order to avoid conflating different languages when fusing information from different language sources, especially when automatically merging KGs or when operating across multiple knowledge graphs in different languages through, e.g., federated queries. Language tags can help applications decide which label should be displayed to which user.

• $\mathrm{■}$

  Language coverage [31, 44]: For a multilingual KG it is crucial that entities are labelled across a large number of languages for which a relevant label exists. Only such thorough labelling can ensure access of all users to all information, independent of the languages they speak. It also enables KGs to readily make other applications multilingual.

• $\mathrm{■}$

  Monolingual islands [31]: When parts of the graph are labelled in only one language, they can form monolingual islands. This can happen because knowledge is prevalent in one cultural context and therefore is yet to be translated, such as governmental initiatives publishing structured data in only their native language. Monolingual islands can lead to worse access to a diverse set of knowledge across the graph and should be addressed when working on a multilingual KG.

• $\mathrm{■}$

  Reusing existing vocabulary: In the creation of a KG, it is essential to reuse existing vocabulary, or ontology, to describe the schema of the KG. Especially for label and language information in a KG, it is crucial for easier integration of multiple, complementary KGs. For more information on ontologies, see Section 2.2.

• $\mathrm{■}$

  Unambiguity [21]: As there are multiple complementary as well as overlapping ontologies, it is crucial to make sure to avoid ambiguity. According to the Semantic Web standards, it is recommended to use the labelling property rdfs:label to provide natural language labels of entities. Furthermore, it is often recommended to provide a single preferred label in a given language per entity, while using other properties such as alias to describe alternative names for a concept.

Conceptual differences between different languages may lead to modelling challenges. An important example is the subclass/superclass relationship (which corresponds to hypo-/hypernymy as a lexical relation). For instance, consider the single concept and word for river in English, whereas in the French language and presumably also the corresponding conceptualisation, one distinguishes rivers that flow into other rivers and those that flow into the sea (fleuve versus rivière) [62]. This may occur similarly for relational properties [54]. A common solution is to treat incompatible concepts across different languages as distinct entries in the KG. Thus, one can avoid conflating the English concept of river with the two French ones and optionally also explicitly describe how the different entries relate to one another cross-lingually. It may also be the case that in one language only a verbal form of a concept exists, i.e., it is assumed to be only a relational property, and in another language it is nominalised, i.e., exists as a unary object or object type only, such that heterogeneous mappings may be needed [26].

Ontologies define the structure of a KG by setting standards for the different properties or relationships to be used, and, for example, the classes used in a KG and across the web of data [36]. The W3C Web Ontology Language (OWL) is a particularly prominent formalism to define ontologies across different KGs to ensure the interoperability of different ontologies [37].

To be able to create multilingual KGs, it is crucial to understand how multilingualism is addressed in the ontologies, be it in the formalism or in a declarative model associated with it, how natural languages are incorporated, and how translation of entities may be recorded.

Gillis-Webber and Keet [29] survey multilinguality in and for ontologies. They grouped the literature and practices in multilingual ontologies into three types of modelling multilingual information: (1) using multilingual labels, i.e., labelling one entity across languages using a labelling property and mentioning the language of the label in the label string (such as in Wikidata in the example in Figure 1), (2) linguistic models (covered in Section 2.4), and (3) mapping-based approach, i.e., creating one entity per language and linking them across languages to each other with an appropriate property (such as in DBpedia in the example in Figure 2). They observed that there is currently a limited uptake of multilingual ontologies at large. Further, limitations of each of the three described ways of modelling multilingual data, such as “accurate representation of languages that require grammatical features such as inflected forms and gender” [30], are yet to be addressed, which apply at least in part also to KGs (discussed in Section 3.2 below).

As ontologies can be defined with the creation of each new KG anew, a number of monolingual ontologies in different languages exist that could be interlinked so as to be more interoperable. Efforts in this direction align ontologies across languages, making it possible to link different KGs across languages [41] or translate labels across languages to make ontologies interoperable and reusable [22].

Covering a large number of languages in the data in any KG is crucial to be able to support a wide range of languages in downstream applications. For example, in the domain of Question Answering (QA), using a multilingual KG facilitates easy switching between languages and finding the KG best suited for the language requirements of an application and its user base [46].

A range of studies describes the language coverage of KGs to understand how well they currently serve multilingual users. Studies have found that, across numerous language graphs, there is a lack of non-English information in the form of labels. Further, the most used labelling property is rdfs:label, and across the web of data at large there is a widespread lack of labels, i.e., a large share of entities are not labelled at all. Additionally, most entities are labelled in only one language [49, 21, 44].

Ell et al. [21] and Kaffee and Simperl [49] developed frameworks to analyse language and label coverage in the web of data, in their case a collection of KGs available online. Zaveri et al. [90] survey data quality metrics and describe a metric for human readable labelling, which characterises the coverage of entities by labels. This metric was picked up and made actionable by Debattista et al. [16].

Wikidata specifically shows a slightly different distribution than the web of data at large. While there is still an English bias in the KG, there is a higher degree of language diversity overall [47, 44]. This more varied representation of languages could have different potential causes; for one, there is a multilingual community editing the KG, which leads to a larger number of perspectives added to the graph, compared to KGs that are automatically assembled or mostly contributed to by a single community situated in one part of the world [48]. This international community has been recruited to a large extent from the famous sister project of Wikidata, Wikipedia, where an international community already edits knowledge in their respective language [79]. Further, the community of humans is supported by so-called bots, which are automated tools that import and edit knowledge on Wikidata across languages [45]. These approaches can be instructive for future projects seeking to create more multilingual KGs.

This opens promising directions for the future development of multilingual KGs. With a dedicated community and specialised tools, better language coverage can be achieved in the future. In Section 4.1, we discuss automated solutions to improve the language coverage of KGs.

Given that KGs can be used to store diverse kinds of knowledge, their versatile graph structure also facilitates storing information about language. Some KGs, such as Lexvo [14] and Glottolog [33], focus specifically on linguistic metadata about languages and dialects as well as scripts and characters.

Lexicographical data in KGs describes a subset of data that expresses information about the lexicon of languages, i.e., information about words, phrases, and other linguistic expressions. For example, this linguistic information could describe how to conjugate verbs across different languages. In this example, the verb may be modelled as an entity in the graph that is connected to its different grammatical forms through its properties (edges in the graph).

To express and standardise this type of linguistic information, OntoLex Lemon⁸⁸8https://www.w3.org/2016/05/ontolex/ was introduced based on the RDF standard, focusing on expressing linguistic information as Linguistic Linked Data [64].

The concept of lexemes, a form of describing lexical knowledge, has also been introduced to the KG Wikidata. Wikidata uses a custom community-created schema that only loosely resembles Lemon. However, considering the actual data provided by Wikidata, statistics reveal that there are only a few languages with significant lexeme coverage, while the vast majority of languages have little to no representation [73]. This limited language coverage shows that there is still a long way to go to reuse such linguistic information for applications focusing on low-resource languages.

One of the most successful, widely used knowledge resources is WordNet [65], as it has over the years seen extensive use in a number of NLP applications, such as text summarisation [5] and text categorisation [20]. A large number of similar resources have been created for a multitude of different languages, including low-resource languages, and many of these resources have been interlinked. The Universal WordNet [15] and BabelNet [72] were among the first massively multilingual knowledge graphs, both drawing on WordNet as their backbone. WordNet is based on the notion of synonym sets as linguistic concepts, which are connected by various linguistically inspired semantic relationships.

Despite the widespread use of multilingual lexicographic KGs in NLP, it is important to acknowledge that the coverage is uneven. While high-resource languages are well-represented, many widely-spoken languages of socio-economic importance are covered inadequately. For some low-resource languages, only very basic terminology is covered. Many others are missing entirely. While information about grammar can be useful for languages with sparse training data, the lack of representation in KGs cannot yet fill this gap. However, as argued previously, having a central general-domain storage of this information for low-resource languages can support future applications in broadening language coverage. Therefore, we argue, it would be beneficial to build these resources and maintain and widen existing ones.

Within the narrow computational resource-oriented view, a consequence is the existence of blocking interdependencies. As a concrete example related to Figure 1, within Abstract Wikipedia [89], it has been proposed to automatically induce templates for a template-based approach [32] or grammars for the Grammatical Framework-based approach [81], ultimately to facilitate rule-based natural language generation to generate Wikipedia articles from the KG. This, however, presupposes the existence and usability of corpora, of good quality part-of-speech taggers and of morphological analysers, which are rarely available for LRLs. These resources are likewise needed for automated KG construction and use in KG-driven information systems, such as educational question generation and document navigation. Consider KG verbalisation for, e.g., isiZulu, a LRL in South Africa spoken by around 23 million people [53]: nouns had to be pluralised, but there was no pluraliser, and verbs of the object properties needed to be conjugated, yet grammar books were outdated and recent linguistics research is scattered, so all that had to be investigated alongside the actual KG task [8, 52]. Also, popular multilingual realisers, such as SimpleNLG [28], are easily adapted among a selection of well-resourced languages in the Indo-European languages family, but they may not be suitable for a language that needs subword-level management, and so a new modular realisation engine may be needed [60]. Conversely, any prospective KG task helps focus language resource development on a specific, measurable, and achievable segment, which is confidence-building and a way to gradually expand the resources.

These anecdotes, and similar observations, are illustrative of several general issues when creating KGs or using them in applications with LRLs, be this for data-oriented techniques and applications or other KG tasks, being:

• $\mathrm{■}$

  There will be linguistic hurdles (gaps in the linguistic knowledge and sociolinguistics) to overcome, which are on top of the intended KG task or KG as a solution. In addition, computerising the language information takes time and scarce human resources, delaying the KG task.

• $\mathrm{■}$

  Freely available pre-existing data is often imprecise, incorrect, or outdated, and thus not a good basis to rely on [solely/at all], requiring an additional data collection stage in a data-driven KG task; therefore, an expert-driven rules-based approach may be more viable for some tasks.

• $\mathrm{■}$

  A resource shortage also tends to imply a human shortage, both in numbers and capacity/knowledge/skills of the language, limiting the scale of human evaluation and quality of survey or crowdsourced responses.

• $\mathrm{■}$

  The notion of “good” quality is relative and a lower overall size or quality may still mean that the KG task itself performs well (but lower in context due to compounding of less well performing preceding steps).

• $\mathrm{■}$

  Each LRL has its own set of hurdles, and its own history, and how the low resourcedness came about and therewith may need a context-specific incentive to realise the KG task for the LRL.

Also, and separate to KG building, there are power dynamics. Those who build the KG wield power over those who use it and, as Vang argues, “to some degree contests the autonomy of the user” [43]. A multilingual KG would ideally be built with the community of prospective users who speak that LRL, not just for them. In addition, it is not clear how the notion of KG co-ownership or KG benefit-sharing of extracted and systematised community knowledge, as alluded to in, among others, the groundbreaking San Code of Research Ethics¹⁰¹⁰10https://www.globalcodeofconduct.org/affiliated-codes/, can be realised. Finally, the LRL may have features that do not fit well with the KG language. Since this may be applicable to a subset of LRLs only, we elaborate on this in the next section.

In this paper, we introduce KGs as a well-suited technology to address limitations in language coverage in downstream applications. There are obvious limitations to this claim. Popular KGs tend to be Eurocentric, in terms of their editors (see for example Wikidata [48]) and data covered.¹¹¹¹11Visualisation of Wikidata entities with geolocation: https://wmde.github.io/wikidata-map/dist/index.html Moreover, they have been developed to a large extent by researchers working in what is known as the global north, emphasising the bias in the type of content and how the content is captured and displayed.

The subject-predicate-object paradigm of capturing facts in a knowledge graph, OWL’s assumptions about axiomatisations, and the likes of frame-based approaches suit English well, as it typically takes an object-based approach, has limited verb inflections and has a disjunctive orthography, and an SVO ordering of sentences that are statements.

The practices of modelling assume nouns in the singular and predicates in the 3rd person singular, whose string remains fixed irrespective of the domain or range of the property. This is not the case for several LRLs, however, such as at least Niger-Congo B languages [53], nor for several Indo-European languages that have an extensive noun class system where the verb changes contextually depending on the actor. For instance, the “eats” is all the same in $⟨$Human, eats, apple$⟩$, $⟨$Elephant, eats, apple$⟩$, and $⟨$Microbe, eats, apple$⟩$, regardless who or what (in sg.) is doing the eating, where the “eats” property is being reused as it should be. Consider now a direct translation to isiZulu, respectively: $⟨$Umuntu, udla, i-apula$⟩$, $⟨$Indlovu, idla, i-apula$⟩$, and $⟨$Igciwane, lidla, i-apula$⟩$. The 3rd person singular differs. It is not the case that the predicate is different, just the natural language rendering of it is. However, most KG languages by design typically conflate elements with surface rendering [25]. One could use an identifier, as OBO did, and carry over to OBO Foundry ontologies and in Wikidata, but that still requires additional machinery somewhere to complete the -dla verb stem in accordance with the noun class of the noun. Or: to start properly with KG development, one first needs to figure out some sort of extension of, or addition to, RDF. Depending on the LRL, there may well be up to 20 variants, with one for each noun class.

It has also become common practice in “English KGs” to insert prepositions into the property name or label, such as “works for” and “part of”. They may be realised differently in many other languages, such as being affixed to the noun of the class in the range (object) position or infixed in the verb. The affixation to the noun is also specific to the context where it is used, i.e., in which axiom, not the name of the class or individual, for which there is no established KG language yet. An extreme case is the containment relation, typically used in KGs as contained in. The notion of containment in isiZulu is realised through indicating the container, by means of noun affixation for locatives and determined by both the container and the containee such that there is no “contained in” verb or name to put in the predicate position [51]; e.g., a bolus of food (indilinga yokudla) that is contained in the stomach (isisu) becomes indilinga yokudla isesiswini. What would the triple be? $⟨$bolus of food, contained in, stomach$⟩$ maps neatly with the natural language in English and many other well-resourced languages, but neither a $⟨$indilinga yokudla, blank_prop, isisu$⟩$ nor a $⟨$indilinga yokudla, inverse(L3951-S4), isisu$⟩$ are satisfactory solutions.

Therefore, a KG may need to be accompanied by a declarative language model and a set of grammar rules, or a different way of usage to represent all the required inputs, or the KG representation language may need to be revised. This is regardless of the usage scenario, from rendering the content of the KG correctly and understandable to the user to automated KG creation. The W3C community standard OntoLex-Lemon [13] as well as the more expressive lemon model [61] that aim to provide such a declarative model address this only in part [29, 10].

These language differences lay bare certain English-oriented assumptions baked into the representation language and in naming conventions. This is complicated further if the conceptualisation or terminology diverges not only for domain knowledge but also for a foundational relation such as parthood (e.g., [23, 54]). This need not impede KGs as a possible contributing solution to LRL applications but is to be borne in mind in both KG construction and in their use.

These limitations need to be addressed to make KGs truly inclusive of a wide range of languages.

Recent research has seen an enormous development in the use of KGs to improve Natural Language Processing tasks such as Natural Language Inference (NLI), Question Answering (QA), and Recommender Systems. Even well-known KGs like DBpedia and Wikidata, which are widely used, are the largest in their English versions despite major human efforts to make them available across languages. Furthermore, geographic area-specific facts are frequently restricted to the KG unique to the region or the native language. The incorporation of Machine Learning (ML) models into several languages is constrained by the scarcity of multilingual knowledge.

Machine Translation (MT) systems have been used [2, 63] to improve the language coverage in multilingual KGs, but these efforts only aim to translate domain-specific KGs from English into a target language. These methods ignore the graph structure of KGs, which is critical for determining the domain in which the word must be translated in the target language. Taking into account the graph structure of KGs can help an MT system identify the proper translation for ambiguous labels. As referred to in the survey [70], Rule-based (RBMT), Example-based (EBMT), and Statistical Machine Translation (SMT) based models have been used in the past to translate Semantic Web Technologies (SWT). However, MT for SWT still remains open research due to: (1) lack of clearly specified object attributes, such as cardinality or reflexiveness, (2) concept blending across thesaurus, vocabulary, and ontology, (3) inaccurate definitions of the domain and range, and (4) using ambiguous annotations [70]. However, domain-specific terms from the medical and financial areas have been translated using a Neural Machine Translation (NMT) architecture, outperforming SMT results [3]. Feng et al. [24] introduced a gated NN strategy for translating English KGs into Chinese, learning continuous triple representations. Source and target triples were mapped in the same semantic vector space using their method, which was extracted from Freebase. Their modified NN strategy increased translation accuracy compared to a strong NMT baseline, highlighting the significance of taking into account KG structure for KG translation and enhancing the quality of disambiguation for ambiguous phrases. Another NMT model, THOTH [69] trains to translate the facts from one language into another, treating the facts (i.e., triples) in a KG as sentences with URIs acting as tokens. It uses KG embeddings and two separate recurrent neural network models to extract bilingual alignments between a source and target KG and then learns the translation.

KG embedding based models have been proposed to enhance multilingual KGs. A more realistic method would draw on the information in several language-specific KGs, keeping in mind that individual KGs have their own strengths and limitations on data quality and coverage. This is a significant challenge since inconsistently expressed facts and a lack of sufficient alignment information can make it difficult for knowledge to be transferred between many independently maintained KGs. KEnS [11] is one such approach that embeds all multilingual KGs in a shared embedding space where the association of entities is captured through self-learning.

LLM-based KG creation has been used recently in an attempt to create monolingual knowledge facts for KGs [77], but multilingual KG construction and enrichment have not yet been attempted at their full strength. Due to the potential for complementary and unequally dispersed knowledge stated in many languages, multilingual LLMs can offer richer combined knowledge for multilingual KGs. Prix-LM [91] is a pioneering model for multilingual KG construction and completion that employs both monolingual triples and cross-lingual linkages retrieved from existing multilingual KGs, followed by fine-tuning a multilingual language encoder XLM-R via a causal language modelling objective. Hou et al. [38] proposes another model that leverages adapters to fine-tune LLMs for multilingual KGs, including low-resource languages, enhancing the corresponding downstream tasks. Therefore, leveraging LLMs for the construction and completion of multilingual KGs emerges to be one of the promising future directions of research, considering that their information content has shown promise for knowledge extraction [77].

In this section, we survey a set of use cases for multilingual KGs. When developing new KGs, new ontologies, or improving the language coverage of existing KGs, it is crucial to keep the use-cases in mind to ensure that the KG can indeed serve knowledge that is useful in the envisioned tasks. Several applications already make use of multilingual KGs, in the following we present a selected few tasks.

Multilingual Knowledge Graph Question Answering (mKGQA) is a task that involves answering a user’s questions in a set of languages based on facts stored in a KG. This is a research topic of particular importance, as these systems can bridge the gap between users and pertinent information on the Web. The task should ideally address all languages, making it possible for users to ask questions independent of the languages of the information in the KG. Currently, however, these systems typically define a set of languages in which questions can be answered, relying on information in the corresponding language stored in the target KG. mKGQA is strongly dependent on data, and there is currently a lack of multilingual data on the web, making it accessible to only a fraction of people¹²¹²1225.9% according to https://www.statista.com/statistics/262946/share-of-the-most-common-languages-on-the-internet/, retrieved 2023-02-15, therefore causing a “cultural gap” on the Web [66]. Further, the task of mKGQA suffers from a lack of multilingual benchmarks.

Perevalov et al. [75, 76] identified 17 mKGQA systems. We have provided in the table below a list of 4 mKGQA systems filtered based on the coverage that they provide for languages besides English (with a focus on lower-resourced languages), along with the KG used, data sets, and the specific languages covered.

Knowledge Graph Completion aims to add new, missing facts to a KG. This remains a challenging problem that is particularly pronounced in multilingual low-resource languages, given that human annotations are rare and difficult to procure [39]. A number of solutions have been devised to mitigate this problem, such as using KG embeddings [40] or jointly predicting entity alignment across languages and new facts for KG completion [9, 86, 11]. A promising approach is the self-supervised adaptive graph alignment (SS-AGA) method [39], which regards alignment as a new edge type between parallel entities instead of a loss constraint, and fuses KGs from different languages in a single graph. This approach has been evaluated on both the public multilingual DBpedia KG and a newly created industrial multilingual E-commerce KG.

Cross Lingual Fact Extraction (CLFE) is the task of extracting facts from a text to, e.g., store the facts in a structured data format. Extracting facts from source text of different languages has not received as much attention as monolingual fact extraction [83]. KGs can support multilingual fact linking and extraction, which is important in many downstream tasks such as QA. The REFCOGLink model [57] is based on linking facts expressed in a sentence to the corresponding fact labels (i.e., language-specific representation of the fact) in the KG and outperforms standard retrieval + re-ranking. The CLFE (Cross-Lingual Fact Extraction) [1] model demonstrates strong performance in multilingual and cross-lingual fact extraction tasks, specifically in English and seven other LR Indic languages. It achieves an F1 score of 77.46% using two different approaches. The first approach is a classification-based method, where the model first extracts the object or “tail” of the fact and then predicts the relationship between the extracted tail and the subject. The second approach is a generative one, which combines both tasks into a single step. CLFE makes use of the XAlign dataset, which contains 0.45M pairs across 8 languages, of which 5,402 pairs have been manually annotated.

Multilingual Relation Classification is the task of extracting relations (i.e., triples in the context of KGs) from natural language text in various languages. Multilingual relation classification has been explored through the method of prompting, which can receive promising results even for lower-resourced languages [12]. One of the challenges for the task of multilingual relation classification is the lack of multilingual datasets and benchmarks. The dataset RELX [56] aims to close this gap by providing a baseline model and benchmark for English, French, German, Spanish, and Turkish. IndoRE [71] is a comprehensive dataset comprising 21,000 gold-tagged sentences for named entity recognition (NER) and relation extraction (RE) in three Indian languages (English, Bengali, Hindi) as well as English. The dataset provides valuable resources for advancing research in relation classification of Indian languages, which is crucial for KG augmentation and Question Answering systems. The authors employ multilingual BERT and transfer learning techniques, and propose TransRel, a multilingual system for joint named entity recognition (NER) and relation extraction (RE) with interlingual transfer.

Neural Machine Translation translates text from a source to a target language using deep learning architectures. The performance of Neural Machine Translation (NMT) systems can be enhanced through the integration of KGs, particularly when translating domain-specific expressions and named entities. For example, Moussallem et al. [68] introduce an approach incorporating the KG DBpedia into NMT models, resulting in considerable improvements in performance of these NMT models.

Automatic KG creation for low-resource languages has been leveraged as another possible route to tackle the lack of low-resource language KGs. HKC has been recently introduced as a framework that constructs a knowledge graph for the Hindi language [88] using various NLP techniques. FarsBase, a Persian multi-source knowledge graph [4] is another example of such an application. To construct such a KG, the authors apply a number of techniques to integrate data from Wikipedia and both structured and unstructured data from the web.

Overall, KGs offer immense potential in supporting multilingualism, which in the future will likely also increasingly benefit low-resource languages. By constructing multilingual KGs, aligning entities across languages, and employing techniques such as machine translation and cross-lingual extraction, KGs can effectively bridge the knowledge gap and facilitate access to information in low-resource languages. These advancements not only empower speakers of all languages but also contribute to a more inclusive and diverse knowledge ecosystem.

The lack of multilingual information in KGs should be addressed in the future, and while there are a number of approaches seeking to mitigate this problem, none have thus far succeeded in addressing this important challenge. One of the notable issues with current KGs is that they are typically not aligned across languages, given the usage of different language ontologies. This misalignment poses a large challenge when merging knowledge across languages. Addressing the cross-lingual alignment of ontologies is one of the crucial challenges that will build the foundation for the wide availability of cross-lingual information.

Merging KGs across languages, even with aligned ontologies, remains a non-trivial task. Not only is there a need to identify alignments at the entity and class level. Future studies are needed to understand which languages are covered to which extent in the different KGs and how these different KGs with potentially different topic-focus could be merged harmonically without centering one language in the approach.

Another open problem, especially for community-edited KGs, is the question of how to interest a larger, more diverse community in the contribution to the knowledge stored in the KG. If a diverse community is contributing, naturally more languages will be covered. Creating incentives and showing how the information could be used, such as building tools based on KG information, could be a way to address this challenge. However, future work will have to better understand existing incentives and broaden them.

While there have been approaches suggested to address the lack of languages covered in KGs (see Section 4), there is currently a glaring gap when it comes to low-resource languages. Having very little available information in any language is a challenge for all applications, especially neural- or deep-learning-driven ones, such as large language models. This then also raises the question of optimal strategies for KG construction, such as human-in-the-loop procedures due to limited and overly noisy data and appropriate incentive strategies for manual modelling. KGs could contribute to closing this gap to some extent by providing central storage of multilingual, linked, reusable data. One of the advantages of using KGs may be that while a fact is not yet translated to a lower-resourced language, it already exists in another language, and by linking the data across languages, these facts can be reused even before translation. However, the challenges described in Section 3.2 in the form of assumptions of language modelling need to be urgently addressed for under-represented languages to be able to catch up in the content representation. From an ontology engineering perspective, it is important to focus on representing the diversity of language structures accurately, lest the technique limits its use to applications in a few languages only rather than the breadth of opportunities across the world.

For the reuse of multilingual KGs in downstream tasks, there are still many opportunities to better integrate multilingual KG information with approaches from the field of NLP and with multi-modal approaches. Reusing the knowledge graph information is not only beneficial to the applications that use them but can also create an incentive to create better, more diverse KGs. To ensure this, the current information needs to be accurately curated, and multilingual information should be put into focus for future approaches in increasing KG coverage. Understanding coverage in different low-resource languages as well as reusing this information is a promising path forward for future work that can have a real-world impact.