Discussion

• In this repository we have provided counts of translations from source languages to specific target languages.

• We also have reported counts on the language of glosses. While these may not be translations, it still may be useful interlingual information for the user. For some languages, the gloss in a second language (usually English) is often present, even when the other properties we use to count translations are absent.

• 'Synonym' is a common property on senses. It is currently the fifth most common property. However, finding translations from the synonyms of a sense in addition to from the sense itself does not appreciably increase the number of translations found.

  • Note that finding translations from synonyms of senses is similar to what is done on English-language Wiktionary (with the caveat that the synonym link on Wiktionary is usually from a sense to a lexical entry and not a sense to a sense). That is, for exact synonyms, the community guidelines on Wiktionary specify having a translation table on only one of the synonyms.

• Considering translations as paths using the 'translation' property from the target language to the source language does not appreciably increase the number of translations found. This is not unexpected, since there are constraint warnings generated when a translation is only in one direction.

• Despite large increases in the amount of data in Wikidata, the number of translations in Wikidata is low. Using the more restrictive criteria from the main analysis found only six source language by target language pairs with translation counts over 10,000. (TBD: compare to Wiktionary.) The number of pairs with translation counts over this threshold was only expanded by also couting sense glosses as translations which gives 21 pairs with counts over 10,000.

• We cast a broad net when counting translations. In our main analysis, in addition to the 'translation' property we considered the 'item for this sense', 'predicate for', and 'demonym of' properties, consistent with the Wikidata documentation for senses. However, that documentation page does not mention whether translation paths of length more than 1 should be considered. On the Wikidata lexicographical data 'Best practices' page, the current instructions in the 'Translation' section imply paths of length more than 1 should be considered:

  • "It is generally good practice to avoid adding translation (P5972) statements between every pair of possible translations."
    • "If there is a path of translation (P5972) statements between two senses, that is enough to establish a link between them."
    • "If every such word is connected to the same item via item for this sense (P5137) statements, these connections are also enough to establish links between them."

• Thus, our criteria for translation property path lengtsh are more restrictive than that suggested by the Wikidata documentation, as we considered only paths of length 1 or 2 from the source to target language and (as a sensitivity analysis) a path of length 1 from the target language to the source. A reasonable compromise might be to consider paths of length 3, as it allows for the possibility that (1) a local language with a small number of speakers (or low capture in Wikidata) has a translation to a geographically nearby language with many speakers which has a translation to another world language with many speakers which has a translation to another language with few speakers. However, we would expect the effect of implementation of this modification to be minimal, as there are currently (7 October 2026) only 7 languages that have more than 1000 senses using the 'translation' property reported in the Wikidata statistics. Paths in the reverse direction could also be considered.

• The reverse direction of the paths mentioned in the previous bullet could also be considered. Currently we only consider 'translation' property paths of length 1 from the target language to the source. A minimal extension that might be useful would be to allow one translation property path from a sense to a second sense and also allow a path from the target language sense to the second sense (i.e. P5972 / ^P5972). However, the bounds mention in the first bullet also apply with the same caveat and the effect will be even smaller since a property constraint generates warnings if the translation property is not used in a symmetric manner.

• We have not considered various qualifiers of the properties we used or other properties of the senses linked that might influence whether the translation defined by the link is qualified. For example, the German sense 'Berlinerin' (female resident of Berlin) has the property 'semantic gender' with value 'female'. The English sense 'Berliner' has no such property, as it refers to both male and female residents. Nevertheless, 'Berlinerin' (German) is counted as a translation of 'Berliner' (English) and vice-versa.

• There are at least a small number of language pairs with more interlingual data in Wikidata than in Wiktionary. For example, when considering gloss senses with more than 10,000 translations on Wikidata, we found more translations on Wikidata for Bokmål, Nynorsk, Egyptian, Sumerian, Akkadian and Northern Sami. However, there are far more languages (69 others) with more than 10,000 translations to English on English Wiktionary and less than 10,000 translations on Wikidata. We have not compared counts for languages with less than 10,000 translations on both platforms, nor have we compared translation counts for Wiktionaries other than the English Wiktionary (TBD?).

• Comments on setting up Apache Jena as our local RDF store are in our repository discussing DBnary.