This paper proposes a methodology for creating an Internationalization Contextualized Knowledge Graph (i18n-CKG) that is designed to capture the wide range of international concepts and concerns. This methodology rst surveys how well known entities are represented through di erent properties in various language locales. Next, the i18nCKG is constructed as a combination of the i18n properties with interlanguage translations and a scoring system to measure the universality of the properties. As a brief exercise, an i18n-CKG that models the i18n di erences for a popular entertainer in English, Korean, Russian, and Chinese is presented.
Internationalization, often abbreviated as \i18n" (18 represents the number of letters between the rst and last), refers to the adaptation of a product or service based in a certain language or locale for another language or locale. As di erent countries, cultures, and communities have di erent priorities and ways of viewing the world, it is important for knowledge graphs to be able to exibly represent these i18n di erences. In this paper, a methodology for creating a contextualized knowledge graph that is sensitive to i18n issues, an i18n-CKG, is presented.
First, a brief literature review will show how i18n considerations have been
previously addressed in existing knowledge graphs, such as DBpedia. Next, the
methodology for creating an i18n-CKG will be described with a case study
using representative knowledge graph applications from di erent language locales,
English (en), Korean (ko), Russian (ru), and Chinese (zh). The paper ends with
a discussion of future work.
A number of prior works have expanded the i18n capabilities of Linked Data.
There have been case studies of i18n for the Greek DBpedia [
It is also important to recognize that people may underestimate the di
culties of representing international knowledge in Linked Data. Often, it is not
as simple as mapping ontologies from di erent languages together. In software
development, a number of notable resources have pointed out these false
assumptions with names [
The intuition for this methodology can be expressed as follows. Although the English Wikipedia1 is a widely used website, people from di erent languages, communities, and cultures maintain their own Wikipedia or equivalent repository of open knowledge. Similarly, although the Google search engine holds a signi cant market share globally, other search engines are prominent in certain locales, such as Naver in South Korea and Yandex in Russia. By analyzing these i18n data sources and search services as a \ground truth", it is possible to compare how di erent locales represent relationships between entities.
First, the i18n data sources and popular local services for delivering
structured information should be identi ed. Wikipedia2 consists of many language
versions. The infoboxes in certain articles also display the structures (or
ontologies) in the respective language Wikipedia. These multilingual infoboxes act as
up-to-date sources of i18n data that may not be captured in older releases of
linked data such as Wikidata and DBpedia. In the case of search services, Google
is widely popular and holds a signi cant market share worldwide [
For an implementation of this methodology, a notable entity that would be displayed in the search services of all four language locales (English (en), Korean (ko), Russian (ru), Chinese (zh)) was selected. As a popular Korean pop music entertainer with appearances in international music, television, and lms, South Korean singer \Rain" was compared across the aforementioned services to determine which relationships (properties) were most prominently displayed in the structured infoboxes and search engine results. With the structured data from the various sources collected computationally, machine translation of nonEnglish texts and human-supervised reconciliation of complex properties were used to create the nal set of triples from each i18n source.
As part of the i18n-CKG, an additional measure is proposed to score how frequently the i18n properties are present across language locales. This expresses how \local" a speci c property is for a language. Such scores would also be 1 https://en.wikipedia.org/wiki/Main Page 2 https://www.wikipedia.org/ represented as triples: (p, i18nPropScore en, value). A simple implementation is to use sl = rt where r is the presence of the property in a language (0 or 1) and t is the total times the property is present across l locales. Higher scores indicate more \local" or locale-speci c properties, while lower scores mean more \universal" properties, unless the value is zero. The code to generate the proposed i18n-CKG is maintained on GitHub3.
For a complete translation of the ontology across languages (ontology
localization), this can be accomplished through methods outlined in existing guides
[
Table 1 shows how the four language locales represent the properties of the entity \Rain". For instance, the \date of birth" and \place of birth" properties are the most consistently displayed across all data sources, suggesting that these properties are universally interesting across locales. A number of properties represent less well known attributes, such as weight, blood type, and zodiac sign. These properties are likely to be of interest in the language locales where they are represented, but not as relevant or interesting for locales that omit them.
A selection of the preliminary calculations of the i18n property scores are noted below. The higher scores indicate that a property is more relevant in a particular language locale, such as \Blood type" for zh, while lower scores are attributed to properties that are widely found in many language locales. Once all additional triples have been created, the resulting i18n-CKG data may be loaded into an existing triplestore or knowledge graph to provide i18n context. { (Date of birth, i18nPropScore en, 0.25 (= 1/4)) { (Date of birth, i18nPropScore zh, 0.25 (= 1/4)) { (Blood type, i18nPropScore en, 0.0 (= 0/1)) { (Blood type, i18nPropScore zh, 1.0 (= 1/1)) 5
A number of extensions to the methodology proposed in this paper are possible. Qualitatively, the di erences between the properties in the di erent language ontologies can be compared more deeply. For example, for the property label \Born", there may be a combination of \place of birth", \date of birth", and/or \birth name" that are actually found in the data. In addition, there are also
opportunities to expand the computational methods in future iterations of this research to scale with the amount of entities (on the order of millions) and languages (on the order of hundreds).
At an implementation level, the creation of i18n labeled properties requires international language translation capabilities that can also scale. A combination of reliable machine translation tools and linguistic veri cation by bilingual experts may need to be implemented to maintain accuracy. While this is straightforward for some properties (e.g. \date of birth"), di culties may become apparent as more obscure or i18n-speci c properties are examined. Improvements in these processes should be undertaken while also building on similar ontology localization and mapping e orts in such cases as DBpedia4.
Limitations to the i18n-CKG approach are also important to consider. For instance, biases in the data may a ect the coverage of i18n-CKG. Certain languages and communities that are not well represented online would not be in
cluded to the same degree as more widely used languages such as English and French. There may also be divergent representations of the same entity or property within the same language or locale, which require additional work to resolve.
As more and more locales are covered, it will also be necessary to develop processes for dealing with data that con ict between i18n locales. Con icting data between i18n locales may arise due to political issues (e.g. territorial disputes), cultural di erences, or unreliable i18n sources, among other reasons. Further work is needed to handle such potentially sensitive matters.
In this paper, a preliminary methodology for creating a contextualized knowledge graph that is sensitive to i18n issues, an i18n-CKG, was described. The two stage approach included a survey of i18n di erences using various i18n sources of \ground truth" and the construction of a knowledge graph that accounted for the i18n di erences. Future research in building contextualized knowledge graphs with i18n considerations in mind should prove fruitful as new methods and improved processes are developed.