Wikidata:Recoin/fr
Recoin (Relative Completeness Indicator) est un script qui ajoute des informations sur la « exhaustivité relative » de l’information sur les pages d’entités de Wikidata. « Exhaustivité relative » fait référence à l’étendue des informations trouvées sur un élément par rapport à d’autres éléments similaires.
L’indicateur regroupe l'étendue des informations dans une barre de progression colorée, montrant 5 niveaux possibles de complétude, allant de « informations très détaillées » à « informations très basiques ».
Recoin veut aider à la fois les contributeurs à savoir où concentrer leur attention et à sensibiliser les consommateurs de données au degré d’information d’un article spécifique.
Motivation
Recoin est destiné à aider les auteurs et les consommateurs de Wikidata.
Pour les « utilisateurs » (consommateurs), il fournit un résumé pratique du degré d’exhaustivité des informations dans Wikidata, qui peut les aider à décider de s’appuyer ou non sur Wikidata afin de satisfaire leurs besoins en informations. Cela est dû au fait que juger purement par la longueur de l’article peut ne pas toujours être une bonne idée, comme par exemple le joueur d'échecs Jeff Sarwer (Q3494327) a un long article en raison de nombreuses déclarations sur sa note Elo, mais jusqu’à récemment manquait même des informations très basiques telles que la citoyenneté ou le nom de famille.
Pour les "auteurs", il fournit également des informations sur les personnes dont les informations sont plus complètes que celles des autres, ce qui leur permet de concentrer leur attention sur les personnes les plus incomplètes. Pour une personne individuelle, cela lui permet de voir les propriétés les plus importantes qui manquent, les auteurs qu'elle pourrait s'efforcer de compléter ou, s'il n'existe aucune valeur pour ces propriétés, de le signaler par une assertion "sans valeur".
Ce qu'il montre
Recoin peut ajouter deux types d'informations aux pages Wikidata :
- Une icône d'indicateur de statut à 5 niveaux, allant de "très détaillé" à "très basique", résumant l'étendue de l'information par rapport à d'autres entités similaires ;
- Deux listes extensibles des propriétés absentes les plus pertinentes et des identifiants externes sont ajoutées en haut des pages d'entités.
Fonctionnement
Architecture
The architecture depicted in the figure to the right shows both JavaScript modules recoin-core.js and recoin-explanations.js that send request to the getmissingattributes.php located on Toolforge. In turn this PHP script does the computation by making requests, first to the Wikidata SPARQL endpoint to get occupations for the given entity, and then by queries to databases on ToolsDB, to retrieve the attribute frequencies for the (previously computed) occupations (humans) or class (all non-humans). The results (completeness and the missing properties) are returned in JSON serialisation and are used by the JavaScript modules to render the page.
Calcul
The script so far does computation for all classes contained in the table wikidatawiki_p.wbs_propertypairs [1]. Furthermore, it gives more refined results based on the 1000 most frequent professions of humans, by treating professions like classes.
Détermination des propriétés et des identifiants absents
We first describe the case of an entity belonging to a single class/profession, and discuss multi-class-membership later below.
Given an entity that belongs to a certain class, we compute the properties most frequently occurring in that class, and check how many of those are absent for the entity. The top-10 missing properties are shown by the core script (a second script shows also external IDs). For classes contained in wikidatawiki_p.wbs_propertypairs, we use all properties available there. For professions of humans, we use the 100 most frequent properties per profession.
For instance, Jimmy Wales (Q181) misses, among other things, the properties languages spoken, written or signed (P1412), member of political party (P102) and position held (P39), which are specified for 13.435%, 9.347% and 8.376% of people of same occupation.
Calcul de l'indicateur d'état
To determine the relative completeness on the 5-level scale, we compute the average frequency of the top 5 missing properties (if there are less than 5 missing properties, we assume their frequency to be zero). We then set the level as follows:
- Level 5 (most complete) 0%-5% average frequency @ top 5 missing properties
- Level 4 (quite complete) 5%-10% average frequency @ top 5 missing properties
- Level 3 (medium complete) 10%-25% average frequency @ top 5 missing properties
- Level 2 (low completeness) 25%-50% average frequency @ top 5 missing properties
- Level 1 (least complete) 50%+ average frequency @ top 5 missing properties
For example, Arno Kompatscher (Q15074414) is missing
- P39 (position held) - 54.33%
- P1412 (languages spoken, written or signed) - 49.93%
- P102 (member of political party) - 46.62%
- P1559 (name in native language) - 31.14%
- P937 (work location) - 30.67%
Thus, the average frequency of the top 5 missing properties is 42.53%, and thus his level of completeness is 2 (low).
Traitement de l'appartenance à plusieurs classes
For entities belonging to multiple classes (see e.g. Dresden (Q1731)) or persons with multiple occupations (e.g. Arno Kompatscher (Q15074414)), Recoin does the computation based on the weighted frequency of each class/profession.
For instance, Arno Kompatscher (Q15074414) is both a politician and jurist. There are 297,370 politicians and 12,635 jurists in Wikidata. If among politicians, 40% do have the property position held (P39) set, while among jurists 20% do have, the final computed frequency is the weighted average of 39%.[2]
Cas particuliers
- For humans, the properties place of death (P20) and date of death (P570) are strictly filtered out, as they are frequent yet frequently undesired for living humans;
- In the case of an entity belonging to a single class that does not have data in wikidatawiki_p.wbs_propertypairs, nothing is shown;
- In the case of an entity belonging to multiple classes or professions, with one having no data, the frequency of properties in that class is assumed to be zero
- Properties having a frequency of less than 0.01% in a class are assumed to have frequency zero
- For entities that have a profession that is not among the 1000 most frequent ones, missing properties are computed based on general humans
Multilinguisme
By default, Recoin shows the property labels in the language defined in the user settings, or where no label is available, in English. The same holds for the Strings of the tool (caption at the top of the page, altLabels of the status indicator icon). Translations can be added here.
Installation
Gadget principal
Recoin can be enabled at Special:Preferences under the section "Gadgets/Wikidata-centric".
Version spéciale : uniquement les identifiants
A special version only showing ID properties can be enabled by adding the following line to Special:MyPage/common.js:
importScript('User:Vvekbv/recoin_id.js');
Where you maintain a global common file, the code to use in m:Special:MyPage/global.js:
mw.loader.load('//www.wikidata.org/w/index.php?title=User:Vvekbv/recoin_id.js&action=raw&ctype=text/javascript');
Les API
Accès par entité
Recoin can also be accessed via an API available at
https://tools.wmflabs.org/recoin/getmissingattributes.php?lang=en&subject=Q15074414&n=10
and
https://tools.wmflabs.org/recoin/getmissingattributes_id.php?lang=en&subject=Q15074414&n=10
(substituting the desired entity Q-code, the language(default language is English) and n required properties(default is 10)).
Accès par classe
To obtain a list of most frequent properties for a specific class, the following API can be used
https://tools.wmflabs.org/recoin/getbyclassid.php?subject=Q185351&n=200
(substituting the desired class Q-code, "n" is the number of results returned(default is 200))
Vidage des données
An August 22, 2019 dump of property frequencies for classes and occupations is available here.
Besides the API above, a way to get fresh data on property frequencies for classes is quarry (example: most frequent properties for films: query).
Informations supplémentaires
Contact:
- Vevake Balaraman - vevake.balaraman@gmail.com
- Simon Razniewski - srazniew@mpi-inf.mpg.de
- Werner Nutt - nutt@inf.unibz.it
Further reading:
- Scientific paper "Recoin: Relative Completeness in Wikidata" by Vevake Balaraman, Simon Razniewski, Werner Nutt, Wiki Workshop at The Web Conference 2018 (link)
- Talk at WikidataCon 2017 "How to know what Wikidata knows"
- Scientific paper "Assessing the Completeness of Entities in Knowledge Bases" by Albin Ahmeti, Simon Razniewski, Axel Polleres, ESWC P&D 2017 (link)
Related projects:
- Wikipedia article quality assessment using ORES
- Wikidata property suggester, a tool that uses aggregated association rules for the suggestion of properties to add
- COOL-WD, a tool that allows to assert the completeness of individual properties directly inside Wikidata.
Acknowledgment: This work is partially supported by the project TaDaQua, funded by the Free University of Bozen-Bolzano.
- ↑ 42078 as of November 15, 2017; query
- ↑ This is not the most precise way, as entities that are both politicians and jurists this way have twice the weight of other entities, but a precomputation of all combinations of professions/classes is infeasible both on the fly or a priori, and this weighting is a reasonable approximation.