Embed Size (px)
Citation preview
Mining cross-cultural relations from Wikipedia -A study of 31 European food cultures
Paul LauferGraz University of Technology
Graz, [email protected]
Claudia WagnerGESIS & U. of Koblenz
Cologne, [email protected]
Fabian FlöckGESIS
Cologne, [email protected]
Markus StrohmaierGESIS & U. of Koblenz
Cologne, [email protected]
ABSTRACTFor many people, Wikipedia represents one of the primarysources of knowledge about foreign cultures. Yet, differ-ent Wikipedia language editions offer different descriptionsof cultural practices. Unveiling diverging representations ofcultures provides an important insight, since they may fosterthe formation of cross-cultural stereotypes, misunderstand-ings and potentially even conflict. In this work, we exploreto what extent the descriptions of cultural practices in var-ious European language editions of Wikipedia differ on theexample of culinary practices and propose an approach tomine cultural relations between different language commu-nities trough their description of and interest in their ownand other communities’ food culture. We assess the validityof the extracted relations using 1) various external referencedata sources (i.e., the European Social Survey, migrationstatistics), 2) crowdsourcing methods and 3) simulations.
1. INTRODUCTIONPierre Bourdieu emphasizes the importance of taste and
related practices for analyzing culture since social groupsoften differentiate themselves from others via these practices[5]. In this work we focus on culinary practices (i.e., culturalpractices of preparing and consuming food) since previousresearch suggests that food and culture are deeply connected[7, 30].
For example, a Wikipedia article on“French cuisine”foundon the Romanian-language edition might surprise a Frenchnational when translated into her mother tongue. Unlikethe French “original”, only a brief mention of French winesexists and only a very short paragraph on croissants andpastries; but, on the other hand, it features a section onfois gras and lamb dishes so extensive that the French lan-guage pendant pales in comparison. This might suggest that
Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
the editor community of the Romanian-language Wikipediacould either have a deviant mental picture of the Frenchcuisine – or it might estimate the priorities of Romanian-speaking readers to rather be on meat-based French deli-catessen than on wine and baking goods. Further, the gen-eral interest of the Romanian-speaking readers in the Frenchcuisine (for example measured by the number of views of thearticle about French cuisine in the Romanian language edi-tion) might serve to potentially displease any Francophile,since the Romanian speaking community might show no-tably less interest in the French kitchen than in the Russianor Hungarian one. This hypothetical scenario serves as anexample for numerous similar real-world cases (which can-not only be found in the culinary domain but also, e.g., inartistic domains) where the perception of and interest in acultural practice differs widely over the language editionsof Wikipedia, arguably the world’s most culturally diverseencyclopedia when it comes to editors and readers.
Since the percentage of people who look up informationon Wikipedia is steadily increasing,1 the public is likely tobe significantly affected by relying on biased informationretrieved from articles. However, biased information andcultural misreading is almost unavoidable since people canonly understand foreign cultures through their own ethnicculture’s lenses. In this light, unveiling diverging (mutual)representations of cultures on Wikipedia is important, sincethey may explain, and even foster, the formation of cross-cultural stereotypes, misunderstandings and even conflict.On the other hand, cultural similarities and frequent expo-sure to different cultures may promote cross-cultural under-standing; which in turn may lead to positive affinities be-tween communities which may even affect the developmentof cross-cultural relations. In this paper, we are interestedin elucidating affinities, similarities and understanding be-tween cultural communities by exploring the collective de-scription and popularity of cultural practices [4] within andacross different language communities on Wikipedia.
Contributions and Findings. In this work we demon-strate how the description of and the interest in each oth-ers food culture differ widely among European languagesand propose a computational approach for mining and as-sessing relations between language communities on Wiki-
1http://www.pewinternet.org/2011/01/13/wikipedia-past-and-present/
arX
iv:1
411.
4484
v2 [
cs.S
I] 1
2 Ju
l 201
5
pedia along three dimensions (cultural understanding, cul-tural similarity and cultural affinity). We evaluate the util-ity of our approach by assessing the plausibility of the re-sults via (i) comparing them with external data that revealinformation about relations between language communitiesin Europe, more precisely shared values and beliefs accord-ing to the European Social Survey and migration statistics,(ii) crowdsourcing methods and (iii) simulations. The eval-uation highlights the general applicability of our techniquefor cultural relation mining, while uncovering potential forincluding further cultural practices (e.g., literature, movies,music) to infer cultural relations beyond the culinary do-main. Our results reveal that shared internal states suchas beliefs and values that may define a culture accordingto Hofstede and Alavi et al.[14, 2] are positively correlatedwith shared practices that can also be used to define a cul-ture as suggested by Bourdieu [5]; the advantage of sharedpractices is that they can often be observed directly andare often well documented, while shared values and beliefsmay only be observed indirectly if they manifest in sharedbehavior. Lastly, through application of our newly intro-duced methodical toolset, we gain insights into patterns ofcultural relations and factors that may be related to them:for instance, as suggested by Tobler’s first law of geography[31], we find that neighbouring countries tend to have moresimilar cultural practices than more distant countries andthat cultural understanding can in part be explained by theglobal importance of a food culture and by migration.
2. RELATED WORKPrevious research acknowledged the fact that interesting
differences exist in different language editions of Wikipedia[12, 8, 24, 13, 23]. Systems like Omnipedia [3] or Manypedia[18] aim to allow users to compare and browse the differentlanguage-specific views which are inherent to Wikipedia. Forexample, Hecht and Gergle [13] showed that the diversityof information across Wikipedia language editions is muchgreater than initially estimated by literature and that onlyone-tenth of one percent is compromised of common con-cepts (i.e., articles about the same topic). In [8] the authorsanalyze the distribution of information describing a singleconcept across multiple language editions of Wikipedia andfind that no facts describing a concept in two different lan-guages were contradictory, but that different language edi-tion focused on different information. Previous research alsofound that the strongest predictor of similarity between twolanguage editions of Wikipedia is size [32], which suggeststhat one needs to be careful when using the concept sim-ilarity across different language editions as proxy for cul-tural similarity. However, our results show that inferringcultural similarity across language editions is possible if werestrict our analysis to a meaningful sub-sample of articlesthat belong to a domain that is related to culture. Our workgoes beyond previous work by exploring how a meaningfulsub-sample of Wikipedia articles about cultural practices arepresented and consumed (i.e., viewing patterns) by the widerpublic and to what extent different factors may explain thecross-cultural relations which we observe on Wikipedia.
Callahan and Herring [6] found culturally biased differ-ences for both, the extent and the concepts, with which thepersons were described on the English and Polish Wikipedia.Also the editing behavior of the Wikipedia community re-veals cultural differences. For example, countries with a
lower Human Development Index (HDI) such as Russia orPoland show less interest in editing and maintaining Wikipediathan more developed countries such as Denmark or Germany[25].
Recent research in the field of computational social sciencehas revealed the potential of non-reactive methods that arebased on the analysis of large amounts of observational datafor studying cultures. Examples include studies of culturaltrends in literature [22], activities of Twitter users [11], in-stant messaging [16], Flickr image tagging [34], Foursquarecheckins [28] or Eurovision songcontest voting data [9]. Thisresearch nicely shows that values and preferences of differentcultures manifest to a certain extent in the online behaviorof people and its observable outcome. Unlike previous workwe exploit the collectively generated descriptions of culinarypractices on different language versions of Wikipedia sinceeach language community may perceive and document theirown and other cultural practices through their particularcultural lenses. In addition to the content we exploit theview statistics to approximate the interest of different lan-guage communities in their own and foreign cultural prac-tices.
3. METHODS & DATASETSThough multiple divergent definitions and measures of
culture have been proposed in previous work (see [17] foran excellent overview), many scholars agree that more ob-servable aspects of culture (e.g., norms, practices or myths)and less observable aspects of culture (e.g., beliefs or values)exist (see e.g., [15, 27]). In “Theory of Practice” [4] and “Ladistinction” [5] Pierre Bourdieu emphasizes the importanceof taste and related practices that people use to differenti-ate themselves from others. He argues, e.g., that the tastesin food, music or art and the related practices which onecan observe (i.e., what people tend to eat or listen to) areindicators of a social class.
In this work we adapt his view on culture and present acomputational approach for analyzing cultural practices ofdistinct communities. We use language as a proxy for cul-tural communities since language is closely linked to bothnational and cultural boundaries; it facilitates the develop-ment of a common identity as is illustrated by the fact thatthe most obvious subnational divisions of cultural groupsare found between language groups in multi-lingual societiessuch as Belgium or Canada [33].
3.1 Cross-Cultural Relation MiningIn the following, we present our approach for Cross-Cultural
Relation Mining (CCRM) that looks at three distinct butinterdependent dimensions of cultural relations:
Cultural Similarity. We assess cultural similarity be-tween two communities A and B by comparing the Jac-
card Similarity2 FA∩FB ||FA∪FB | between the sets of concepts A
and B that are used to describe the cultural practice of theircorresponding community. Concepts on Wikipedia are lan-guage independent and are identified via the links (to otherWikipedia pages) of articles describing cultural practices indifferent language editions. For instance, if the article aboutthe French cuisine links to concepts like “Wine”, “Cheese”and “Baguette”, the Italian cuisine references concepts like
2Other similarity measures might be suited as well and weplan to explore those in future work.
“Wine”, “Pasta” and “Cheese” and the German cuisine isdescribed by concepts like “Beer”, “Sausage” and “Cheese”,then we can conclude that the Italian cuisine is more similarto French than to the German one.
Cultural Understanding. To assess the cultural under-standing of a community B for the culture of another com-munity A, we measure the Jaccard similarity between thedescriptions of A’s cultural practices in the language of com-munity B. If community B uses the same concepts to de-scribes A’s cultural practice as A, we can conclude that com-munity B has a “good” understanding of A’s culture, where“good” means “close-to-native”. If, for example, the Greekcuisine is described using the same culinary concepts (e.g.,ingredients and dishes) by the Italian and Greek languagecommunities, we can conclude that the Italian communityhas a good understanding of the Greek food culture.
Cultural Affinity and Bias. We assess the cultural affin-ity between a community A and B (or A’s bias for B) bymeasuring how much attention community A pays to cul-tural practices of community B. Both, the content of thearticles which describe the cultural practices of a communityand their access volume are used as a measure of attention.For instance, if the Finnish Wikipedia describes the Irishcuisine in much more detail than we would expect and/orthe Irish cuisine page in the Finnish Wikipedia is viewedmuch more often than we would expect, then this clearlyindicates that the Irish cuisine is important for the FinnishWikipedia users. Therefore, we can conclude that they havea positive affinity towards the Irish food culture.
Since some resources might be more popular than oth-ers, we normalize the affinities between communities by theresource popularity as follows:
bias(l, o) =
attention twds. res. o︷ ︸︸ ︷f(l, o)∑
o∈O
f(l, o)︸ ︷︷ ︸attention twds. all res.
− 1
|L| − 1
∑l∈L\l
f(l, o)∑o∈O f(l, o)︸ ︷︷ ︸
normalized rel. attention of others
(1)L is the set of all languages, O is the set of all country-
specific cultural practices which we analyze. bias(l, o) de-fines the bias of language l towards the cultural practices oof one country (in our case the cuisine of a country). Thefirst term of the equation normalizes for the size of the lan-guage community - i.e., how important they consider thecuisine of that country compared to all other cuisines. Thesecond term normalizes for the global importance of the cui-sine. This yields to an affinity value between −1 and +1. Apositive (negative) affinity value indicates that the commu-nity pays more (less) attention to a country-specific culturalpractice than we would expect.
Beside the cultural affinity between communities, we arealso interested in the affinity of each community towardsitself (i.e., its self-focus bias[12]) and towards geographicallyclose communities (i.e., its regional bias). The self-focusbias reveals the tendency of a language community (e.g., theFrench community) to focus on their own cultural practice(e.g., the French cuisine). The self-focus bias is based onthe previously introduced affinity measure and is defined asfollows:
sfb(l) =1
|Oown|∑
o∈Oown
bias(l, o)︸ ︷︷ ︸avg. bias towards own res.
− 1
|Oother|∑
o∈Oother
bias(l, o)
︸ ︷︷ ︸avg. bias towards others
(2)That means, we calculate the difference between the averageattention towards the resources relevant to a culture l andthe average attention towards resources not relevant to theculture (Oother).
The regional bias of a community is defined in the sameway and measures how much more attention a communityattributes to their neighbours’ cultural practices comparedto the practices of non-neighbouring communities3.
3.2 DatasetsAltogether, 27 different language editions of Wikipedia
and 31 different cuisines from across Europe were analyzed,as listed in Table 1. Using this manually created list ofcuisine articles of different European communities as a seeddataset4, we created the following datasets:
Outlink dataset The first dataset consists of the outgoinglinks of all seed articles and the language independent con-cept of each article to which an outlink points. Note thatwe also experimented with a two-hop dataset where we ex-tended the set of seed articles with articles to which theypoint and got compareable results.
View counts dataset Additionally to the content of the ar-ticles, which could in theory be heavily influenced by singlecontributors, the view counts represent the attention differ-ent cuisines receive. We use the view counts for each seedarticle in each language edition between May 2013 and June2014.5
Figure 1 shows the relation between the size of the differ-ent Wikipedia editions (i.e., their total number of pages) andthe number of (European) cuisine pages that they contain.The strong correlation (spearman correlation coefficient is0.82, p� 0.001) shows that larger language communities onWikipedia indeed describe more foreign food cultures. How-ever, some language editions tend to describe many cuisines,indicating a greater interest of their language communityin food cultures, such as the Italian, Ukrainian or Finnishone, whereas others of comparable size only contain articlesabout very few cuisines, such as the Dutch or NorwegianWikipedia. This suggests that the Dutch and the Norwe-gian language community seem to have less interest in foodcultures than we would expect.
4. RESULTSIn the following section we present our results on describ-
ing cultural relations on Wikipedia using the approach whichwe described in the previous section.
3The information about the country adjacency was retrievedfrom https://github.com/P1sec/country_adjacency,which builds upon the Correlates of War Project(http://correlatesofwar.org/) considering two countriesto be adjacent if they are separated by a border or amaximum of 24 miles of water.4https://www.dropbox.com/s/9rq9eqxsgnggtz3/urls.txt?dl=05http://stats.grok.se/
0 500 1000 1500
Wikipedia size in thousands
0
5
10
15
20
25
30
35num
ber
of
cuis
ine p
ages
Ukrainian
Dutch
Italian
Finnish
Norwegian
Swedish
HungarianEstonian
Catalan
Figure 1: Language community’s interests in differentfood cultures: The plot shows the relationship between thesize (in articles) of the respective language edition and thenumber of European cuisine articles it contains. Data pointsin red are furthest from the best fitting linear approxima-tion of the relationship. Data points in green are closest tothe best linear approximation. The English Wikipedia wasleft out, as it is considerably bigger than all other languageeditions.
4.1 Cultural Similarity
RelatedEuro.
Language LC Size Cuisines Editors ViewsBulgarian bg 158,130 Bulgarian 17.91 4669Bosnian bs 48,761 Bosnian 26.00 853Catalan ca 422,684 Catalan 28.58 2155Czech cs 289,551 Czech 31.54 11361Danish da 186,047 Danish 45.50 3122German de 1,692,696 Austrian
and Ger-man
146.34 108501
English en 4,462,417 British,English andIrish
222.77 468724
Spanish es 1,084,184 Spanish 77.78 137567Estonian et 121,329 Estonian 13.25 1221Hungarian hu 256,215 Hungarian 41.10 8603Croatian hr 143,375 Croatian 11.67 1362Finnish fi 342,384 Finnish 22.42 9467French fr 1,481,635 French 74.64 66692Italian it 1,103,118 Italian 46.23 39720Lithuanian lt 163,546 Lithuanian 18.33 4202Latvian lv 52,871 Latvian 13.67 1269Dutch nl 1,763,752 Belgian and
Dutch40.88 13125
Norwegian no 412,649 Norwegian 30.75 2544Polish pl 1,031,851 Polish 46.37 42972Portuguese pt 821,450 Portuguese 37.67 48972Romanian ro 241,239 Romanian 23.00 3575Russian ru 1,093,578 Russian 58.68 59685Slovak sk 190,907 Slovak 21.67 2243Serbian sr 243,268 Serbian 25.00 523Swedish sv 1,612,310 Swedish 32.76 16945Turkish tr 224,742 Turkish 41.71 17674Ukrainian uk 496,343 Ukrainian 28.66 9960
Table 1: Statistics of the Dataset: Language editionsof Wikipedia, their language codes, their sizes, the relatedEuropean cuisines, their average number of unique editorsof cuisine articles and the average monthly views of cuisinearticles (as of May 2014).
To assess the cultural similarity between two communitieswith respect to their cuisines, we use the overlap of conceptsthat are referenced when their cuisines are described.
4.1.1 ResultsFrom a global perspective the two communities which are
culturally most similar with respect to their cuisine are Rus-sia and the Ukraine, followed by Finland and Sweden andLithuania and the Ukraine. This shows that unlike in [32]where the authors report that the size of a language edi-tion was the strongest predictor of similarity, we do not findsize-effects and a manual inspections of language pairs showsthat many of them seem to be plausible and geographic close.This raises the question to what extent cultural similaritycan be explained by geographic distance. To address thisquestion, we compare the average cultural similarity of eachcommunity with neighboring and non-neighbouring commu-nities. Table 2 shows that geographic distances indeed canexplain in part cultural similarity between communities ac-cording to their cuisines and that each cuisine is roughly 1.5times more similar to its neighbors than to non-neighbours(with a standard deviation of 0.2). This finding is in linewith previous research which found that geographic distancehas a large impact on culinary similarities [35] and a slightimpact of similarity between language editions of Wikipedia[32]
4.1.2 ValidationTo validate whether the cultural similarities between com-
munities which we inferred from Wikipedia are plausible, weset up a crowd-sourcing task. From the list of communitypairs which were ranked by their cultural similarity inferredfrom Wikipedia, we randomly selected one pair out of the15 most and the 15 least similar pairs and presented bothpairs to the crowd workers. We asked them to judge whichpair of cuisines is more similar.6 Out of the 225 combina-
6The crowsourcing platform Crowdflower was used, where
Cuisine Sim Cuisine SimAustrian +1.51 Irish +1.50Belgian +1.41 Italian +1.23Bosnian +1.26 Latvian +2.00British +1.22 Lithuanian +1.51Bulgarian +1.42 Norwegian +1.77Catalan Polish +1.48Croatian +1.36 Portuguese +1.72Czech +1.51 Romanian +1.42Danish +1.62 Russian +1.50Dutch +1.58 SerbianEnglish Slovak +1.69Estonian +1.70 Spanish +1.86Finnish +1.73 Swedish +1.67French +1.64 Turkish +1.35German +1.36 Ukrainian +1.36Hungarian +1.14
Average +1.52 Std.Dev 0.20
Table 2: Cultural similarity with neighboring coun-tries: Ratio between the cultural similarity of neighboringcountries and non-neighbouring countries with respect totheir specific cuisine. A value of e.g. 2 indicates that neigh-boring countries are twice as similar as non-neighbouringones. Some values are missing since we defined that a min-imum of 3 neighboring countries are needed for calculatingthe mean.
tions of pairs which we presented to the crowd workers, allbut one where evaluated by the crowd according to our pre-dictions as informed by the article data (199 were selectedwith Crowdflower confidence scores over 0.9, and 25 withscores between 0.6 and 0.9)7.This means that for 99.56% ofall pairs, crowd workers agreed with the high similarity andthe low similarity pairs which our method produced.
To further corroborate the results and in order to validateto what extent the similarity between their cuisine descrip-tions on Wikipedia might approximate the overall similaritybetween two cultures, we compare the similarity ranking ofour communities with their corresponding ranking in the cul-tural similarity index [26] which is based on the EuropeanSocial Survey (ESS). ESS is a biennial 30-country survey ofattitudes, beliefs and behaviour. This comparison revealsa low but significantly positive Spearman rank correlation(ρ = 0.25, p < 0.001).
From our results we can conclude that: (1) the culinarysimilarities between communities produced by our approachseem to be plausible (cuisines which have a very high rank-ing are perceived more similar than those with a very lowranking) and (2) shared internal states such as beliefs andvalues that may define a culture according to [14, 2] are posi-tively correlated with shared cultural practices that have beenproposed as an alternative for defining a culture [5]. Therelatively low correlation is of course not surprising sinceESS captures various variables (e.g., social and public trust;political interest and participation; socio-political orienta-tions; media use; moral, political and social values), whileour Wikipedia based approach focuses on culinary practicesonly. Including further cultural practices such as literatureor art into our approach would most likely increase the cor-relation.
4.2 Cultural UnderstandingTo assess the cultural understanding of community A for
the food culture of community B, we measure how similarcommunity A describes the culinary practice of communityB compared to how B describes it. The understanding rela-tion is asymmetric.
4.2.1 ResultsFigure 2 shows the cultural understanding for all commu-
nity pairs. One can, e.g., see that the Catalan Wikipediarepresents the Portuguese cuisine more accurately than theSpanish cuisine which might reflect the difficult relation be-tween Catalonia and Spain. Also the good understandingof the Finnish and Swedish, but not the Norwegian cuisineby the French Wikipedia seems peculiar. Overall, the ma-trix is rather sparse, which could be interpreted as miss-ing cultural understanding. If, for instance, the HungarianWikipedia does not have an article about the French cui-sine, one could argue that the Hungarians are either notinterested in cuisines in general or they are not interested inthe French cuisine in particular. In both cases, this mightbe an indicator of missing cultural understanding, at least
we received at least 10 judgements per pair paying 4 centsper judgement. We received very positive worker feedback(4.4 out of 5).7The score lies between 0 and 1 and describes the levelof agreement between multiple contributors (weighted bythe contributors’ trust scores), cf. http://tinyurl.com/CFconfscore
bg bs ca cs da de en es et fi fr hr hu it lt lv nl no pl pt ro ru sk sr sv tr uk
Bulgarian cuisineBosnian cuisineCatalan cuisine
Czech cuisineDanish cuisine
Austrian cuisineGerman cuisine
British cuisineEnglish cuisine
Irish cuisineSpanish cuisine
Estonian cuisineFinnish cuisineFrench cuisine
Croatian cuisineHungarian cuisine
Italian cuisineLithuanian cuisine
Latvian cuisineBelgian cuisine
Dutch cuisineNorwegian cuisine
Polish cuisinePortuguese cuisine
Romanian cuisineRussian cuisineSlovak cuisine
Serbian cuisineSwedish cuisineTurkish cuisine
Ukrainian cuisine
Catalonia understands Iberian cuisinesCatalonia understands Iberian cuisines
France understands ScandinaviaFrance understands ScandinaviaFrance understands Scandinavia
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Figure 2: Understanding of food cultures by differentlanguage communities: The heatmap shows the overlapof concepts used to describe a cuisine by its associated lan-guage community and all other communities. The dots marksuch associations (e.g. the Bulgarian Wikipedia communitywith the Bulgarian cuisine). The crosses represent missingdata - i.e., cuisines which are not described by a certain lan-guage community. The absence of these articles might alsobe an indicator for (missing) cultural understanding.
in the culinary domain.Apart from the question whether the internal perception
of a cuisine differs from the external perception, it is alsointeresting to explore the variation of the perception of eachfood culture among different language communities. Forall possible combinations of language pairs, the overlap be-tween the concepts used by the language communities todescribe each cuisine is shown in Figure 3. One can see thatmore prominent cuisines such as the Italian and French cui-sine are more uniformly defined than less prominent cuisinessuch as the Bulgarian or Bosnian. It seems noteworthy thatthe Turkish cuisine is described in a similar way by manylanguage editions as well, although it is not one of the mostfamous cuisines in Europe. A potential explanation for thisobservation is the high migration rate of the Turkish pop-ulation in Europe. According to the bilateral migrationdatabase, they have the third highest negative migrationafter Russia and Poland.
Figure 3 also indicates that only a very small set of lan-guage communities define a food culture using the sameconcepts whereas the majority uses different ones. Thissupports the hypothesis that no globally true and objectivedescription of culturally relevant practices exists on collabo-rative online production systems like Wikipedia, but descrip-tions are highly influenced by the cultural background of thosewho produce them.
4.2.2 ValidationIdeally, to validate the cultural understanding between
0 50 100 150 200 250
ranked language pairs
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45ja
ccard
sim
ilari
ty
Turkish
French
Italian
Bulgarian
Latvian
Austrian cuisine
Belgian cuisine
Bosnian cuisine
British cuisine
Bulgarian cuisine
Catalan cuisine
Croatian cuisine
Czech cuisine
Danish cuisine
Dutch cuisine
English cuisine
Estonian cuisine
Finnish cuisine
French cuisine
German cuisine
Hungarian cuisine
Irish cuisine
Italian cuisine
Latvian cuisine
Lithuanian cuisine
Norwegian cuisine
Polish cuisine
Portuguese cuisine
Romanian cuisine
Russian cuisine
Serbian cuisine
Slovak cuisine
Spanish cuisine
Swedish cuisine
Turkish cuisine
Ukrainian cuisine
Figure 3: Ranked distribution of cultural similari-ties: Each data point represents the jaccard similarity oftwo language-specific descriptions of one cuisine. The pairsare ranked by how similar they describe that cuisine fromleft (highest) to right (lowest) on the x-axis. One can for ex-ample see that for the Bulgarian cuisine (light blue line), thetwo language editions which agree most on the description ofthat cuisine only reach a similarity of around 0.08. Popularcuisines such as the French and Italian cuisine are describedby more language communities in a similar way while lesspopular cuisines, such as the Bulgarian or Latvian, are lesssimilarly described by all communities.
community pairs in the culinary domain that we inferredfrom Wikipedia one would compare our results with exter-nal ground truth data. Since cultural understanding is acomplex concept, no such universal ground truth data ex-ists. However, it seems to be a plausible assumption thatcultural understanding is at least partially influenced by cul-tural similarity (e.g., communities that are similar are likelyto understand each other) and human mobility (e.g., coun-tries that are the target destination of immigrants are as-sumed to understand the source country and culture of thoseimmigrants rather well). For those influence factors externalground truth data exist, although not specifically related tocuisines.
We use the similarity index based on European Social Sur-
vey (ESS) [26] as a rough approximation of cultural simi-larity between countries and data from the global bilateralmigration database as a proxy for cross-country mobilitypatterns. The lists of country pairs which were presentin all three data sources (ESS, migration and wiki) wereranked as follows: The country pair AB with highest rankaccording to the migration statistics (Ukraine and Russia)is the pair where most people moved from country A to B.How many values and beliefs country A and B share de-termines the ranking of this pair in the ESS data. Finally,how well the language community associated with countryA understands the food culture of country B determines theranking of this pair according to Wikipedia. Table 3 showsthe Spearman rank correlations between these three rankedlists of pairs. The cultural understanding which we inferredfrom Wikipedia seems to be better explained by understand-ing due to migration than due to shared values and beliefs.Further, migration explains to a lesser extent shared valuesand attitudes captured in the ESS index than cultural un-derstanding inferred from Wikipedia. This is a noteworthyfinding, since it illustrates that the flow of people in Europemanifests more in shared understanding of culinary practicesrather than shared values and beliefs.
4.3 Cultural Affinity and BiasTo assess the cultural affinity between two language com-
munities we measure how much more attention one commu-nity pays to the food culture of the other one compared tohow popular this food culture is from a global perspective.
4.3.1 ResultsPrevious work has shown that each Wikipedia edition
tends to describe their locally relevant information in moredetail than geographically distant concepts (cf. [12] [24][19]). Places like cities or monuments that are located inFinland are, for instance, described much more accuratelyand in greater detail on the Finnish Wikipedia than anyother language edition [12]. Our results suggest that theself-focus phenomenon can also be observed for cuisines (cf.Table 4). However, it is interesting to note that we foundlot of variation in the community-specific self-focus biases.Some countries and their corresponding communities revealonly small or moderate self-focus biases while others likeBulgaria, Catalonia or Hungary seem to be especially inter-ested in their own cuisines. Figure 4 shows the distribution
Pair ρ (p-value)wiki - ess 0.18 (� 0.001)wiki - migration 0.36 (� 0.001)ess - migration 0.22 (� 0.001)
Table 3: Correlations of cultural understanding withexternal validation data: The correlation values be-tween the cultural understanding extracted by our approach(wiki), the values of an index derived from the European So-cial Survey (ess) and the migration data from the WorldBank (migration). The cultural understanding betweencommunities as inferred from Wikipedia can in smaller partsbe explained by shared values, believes and behavior be-tween those communities and in larger parts by migration.Migration manifests less in shared values, attitudes and be-havior captured in the ESS index than in cultural under-standing inferred from Wikipedia.
of the biases, where the self-focus bias is clearly visible. Italso becomes apparent that the view data shows the highestself-focus bias, which means that consumers of Wikipedia areindeed much more interested in their own food culture, whileeditors seem to make an attempt to also represent relevantforeign food cultures.
In addition to the question whether a direct self-focus biasis visible, it is also a plausible assumption that geographicdistance might impact affinities between communities [29].Table 4 shows that the variance across communities with re-spect to their affinities and biases towards their neighbours ishigh. The Portuguese, Finnish and French Wikipedia com-munities seem to have stronger positive affinities towardstheir neighbouring communities. Contrarily, the Nether-lands, Bulgaria and the Ukraine do not show strong positiveor negative affinities towards their neighbours.
Our results show that cross-cultural affinities on Wikipediaare distributed around zero but are slightly skewed to theright. That means, most communities do not have specificaffinities towards each other, but some show much strongerpositive affinities towards a foreign food culture than wewould expect given the global popularity of that food cul-ture.
4.3.2 ValidationAssessing the validity of the cultural affinity relations for
cuisines between communities which we extracted from Wikipediais a difficult task since no obvious ground truth exists. Pre-vious research suggests that the historical votes for the Euro-
Self-focus Regional focusLanguage Outlinks Views Outlinks ViewsBulgarian +0.227 +0.612 -0.033 -0.004BosnianCatalan +0.200 +0.523Czech +0.062 +0.171 +0.000 +0.038DanishGerman +0.054 +0.047 -0.018 -0.005English +0.026 +0.036 -0.034 -0.000Spanish +0.192 +0.137 +0.005 +0.041Estonian +0.223 +0.384 +0.039 +0.018Finnish +0.032 +0.163 +0.038 +0.033French +0.138 +0.066 +0.008 +0.033CroatianHungarian +0.280 +0.441Italian +0.035 +0.053 +0.006 +0.014Lithuanian +0.019 +0.331 +0.005 +0.039LatvianDutch +0.057 +0.068 -0.096 -0.018NorwegianPolish +0.143 +0.166 -0.009 +0.013Portuguese +0.090 +0.122 +0.175 +0.128Romanian +0.218 +0.532Russian +0.008 +0.137 +0.007 +0.006SlovakSerbianSwedish +0.157 +0.212 +0.004 +0.006Turkish +0.081 +0.297Ukrainian +0.164 +0.350 -0.017 +0.004Average +0.120 +0.242 +0.005 +0.022Std.Dev 0.082 0.175 0.053 0.033
Table 4: Self-Focus Bias and Regional Bias: The biasof a community towards their own and neighbouring cuisinesranges between -1 and +1. It becomes apparent that nearlyall language communities on Wikipedia reveal a positive self-focus bias, while only half of them show a positive regionalbias. Some values are missing since some Wikipedia editionsdescribe less than three cuisines.
0 5 10 15
ranked language editions
0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
self-f
ocu
s bia
s
Words
Outlinks
Views
(a)
0 2 4 6 8 10 12 14
ranked language editions
0.10
0.05
0.00
0.05
0.10
0.15
0.20
regio
nal bia
s
Words
Outlinks
Views
(b)
Figure 4: Distribution of (a) self-focus and (b) re-gional biases: The horizontal gray line indicates that nobias is present – i.e., foreign (or distant) cuisines are consid-ered as important as the own (or neighbouring) cuisine. Onecas see that most language editions reveal a strong positivebias towards their own cuisine, but only half of them showa slight positive tendency towards cuisines of neighbouringregions.
vision Songcontest may expose politically motivated or cul-turally motivated affinities between communities [29, 9] andmight be used as a proxy for our use case.
Our results show that the view based affinities reveal thehighest agreement with the affinities extracted from the Eu-rovision voting data (up to 0.25 depending on the yearscovered). We further found that less affinity is exposed onWikipedia compared to the Eurovision Songcontest votingdata, which is not surprising since although different arti-cles on Wikipedia compete for a limited amount of atten-tion, there is no explicit competition going on like in theEurovision Songcontest.
4.3.3 SimulationTo reveal the potential effect of affinities between com-
munities on the observable outcome (the view data or thenumber of concepts per culturally relevant resource in differ-ent language editions), we simulate the process that mightgenerate the data. Our model receives as input a weightednetwork where all communities are connected. The weightsare uniformly distributed if no affinity biases between com-munities exist (Model 1 in Figure 5 (a)) or can be drawnfrom a normal distribution (Model 2 in Figure 5 (b)) if weassume community-pair specific biases. Further we assignto each community a popularity value which defines howpopular the food culture of a community is - i.e., how muchattention it will receive from all other communities inde-pendently of the presence of community-specific affinities.Finally, each community may reveal a self-focus bias whichis basically the affinity towards itself. In our simulations weset the self-focus bias to 0.242 since our empirical results re-vealed that this is the average self-focus bias of a communityon Wikipedia.
The first model (Figure 5 (a)) simulates a world whereonly the popularity of cultural practices determines howmuch attention they will receive from all other communi-ties. The Jensen-Shannon-divergence (JS-divergence) withrespect to the parameter λ of the exponential function isshown in Figure 5 (a) and reveals how well the affinities of
10-1 100 101
1 / λ
0.0
0.1
0.2
0.3
0.4
0.5
jense
n-s
hannon-d
iverg
ence
0.33
(a)
0 10 20 30 40 50 60 70
σ
0.0
0.1
0.2
0.3
0.4
0.5
jense
n-s
hannon-d
iverg
ence
0.08
(b)
Figure 5: Jensen-Shannon (JS) divergence betweenthe empirical data and two different models. Model(a) corresponds to a global popularity model where onlythe global popularity of a practice influences how much at-tention it received. Popularity values are drawn from anexponential distribution with parameter λ. Model (a) canapproximate the empirical data best if the popularity distri-bution is not too skewed and reaches a minimum divergenceof 0.33. Model (b) additionally considers community-pairspecific affinities as an influence factor that may drive at-tention. The affinity values are drawn from a normal distri-bution with fixed mean (µ = 10.0) and variance σ. One cansee that Model (b) can approximate the empirical data muchbetter than Model (a) and reaches a divergence of 0.08 whenthe affinities between communities has a variance σ ≥ 30.
the simulated view data approximate our empirical observedaffinity distribution. Our results show that if λ is too big(i.e., the popularity distribution is too skewed) the empiricaldata is approximated worse than if we assume a less skewedpopularity distribution. Nevertheless, the best simulationresult shows a JS divergence of 0.33. The JS-divergencewould be zero if the two distributions were identical andapproaches infinity the more they diverge.
The second model (Figure 5 (b)) uses the best popular-ity distribution (i.e., a popularity distribution which is nottoo skewed and therefore explains our empirical observationsbest) and draws affinities between communities from a nor-mal distribution and different variance values. If the vari-ance is zero, the second model is identical to the first modelin the sense that affinity values are uniformly distributed.However, if the variance is greater zero, the model simulatesa world where attention towards cultural practices is drivenby the global importance of practices and cross-communityspecific affinities. The greater the variance the stronger thecross-community-specific affinities.
One can see in Figure 5 that a simulation model whereaffinities between community-pairs vary (σ ≥ 30) approxi-mates the empirical data best (much better than the modelbased on popularity alone). This suggests that the processwhich generates the cross-cultural community attention onWikipedia is driven by both factors, global popularity ofcultural practices and cross-community specific affinities.
5. DISCUSSIONIn this paper we presented an approach for mining cultural
relations from Wikipedia. In order to demonstrate the util-
ity of our approach, we applied it to one specific cultural do-main, cuisines and their representations on Wikipedia. Al-though our empirical results are limited to this domain, ourapproach is general and can be extended to further culturaldimensions, such as music, literature, arts and others.
Keeping in mind the much broader scope of the exter-nal datasets (ESS, migration, song contest) used for empir-ical evaluation of our approach, the found correlations, inconjunction with the results from our crowdsourcing experi-ment and simulation are robust indicators that our approachindeed allows to extract meaningful cultural relations fromWikipedia.
Our empirical results further show that communities thatshare more common values and beliefs according to the ESSindex are also more likely to share culinary practices. Thisillustrates that a weak but significant correlation exists be-tween shared internal states such as beliefs and values thatmay define a culture [14, 2] and shared cultural practices [5].
Not surprisingly, we found that prominent food cultures,such as the French and Italian cuisine are better under-stood (i.e., their descriptions are more consistent across dif-ferent language editions) than less prominent cuisines suchas the Bosnian one. However, we also find that migra-tion correlates with the cultural understanding between lan-guage communities as observed on Wikipedia to a certain ex-tent. For example, besides the Italian and French cuisine theTurkish cuisine is among the three globally best understoodcuisines and according to the bilateral migration database,Turkey has the third highest negative migration after Rus-sia and Poland. Note that e.g. in Germany the largest partof the immigrant economy belongs to the food sector [21]and therefore migration is often related with the number ofrestaurants of foreign cuisine that exist in a country.
Related to the concept of cultural similarity and under-standing is the concept of cultural affinity since one canhypothesize that communities which are e.g. more similarare also more interested in each other and vice versa. Ouranalysis of cross-cultural affinities on Wikipedia showed thataffinities are distributed around zero but are slightly skewed.That means, most communities do not have specific affini-ties towards each other, but some show much stronger pos-itive bias towards a foreign food culture than we would ex-pect given the global popularity of that food culture. Re-markably, no strong negative affinities can be observed onWikipedia - i.e. there are no communities which are sig-nificantly less interested in a foreign food culture that ispopular from a global perspective. Our simulations confirmthat affinities between communities are present and allowus to reproduce the empirically observed cross-cultural viewpatterns much better than a model that assumes that theinterest between communities only depends on the globalimportance of their cultural practices. Finally, our resultsalso show that all European communities reveal a substan-tial self-focus bias (which corroborates related results from[12]) and a moderate bias towards geographically close re-gions concerning cuisines (as in general also suggested by[24, 35, 32]).
Analyzing the relation between cross-cultural affinities,understanding and similarities on Wikipedia suggests that ahigh understanding between two cultural distinct communi-ties does not necessarily lead to positive affinities (Spearmanρ between understanding and affinity is −0.0393), but highsimilarity tends to correlate positively with both, under-
standing and affinities (Spearman ρ between similarity andunderstanding +0.1880 with p� 0.001 and between similar-ity and affinity ρ = +0.287 with p� 0.001). This suggests,that similar cultural groups (i.e., groups which have similarcultural practices) do not only understand each other, butalso show high interest in each other, while dissimilar cul-tural groups that understand each other do not necessarilyshow high interest in each other.
Implications: We proposed a viable method to modeland extract cultural relations from Wikipedia by analyz-ing cultural practices, on the example of culinary practices,on three distinct dimensions and showed how different lan-guage communities relate to each other in that domain. Byextending the presented approach to domains such as music,literature, art, etc., it seems likely that it could become apowerful tool to complement reactive data collection suchas surveys, which have been traditionally applied by socialscientists and, e.g., market research firms to study how dif-ferent cultural groups relate to each other. Complementingreactive data collections for cross-cultural studies is an im-portant issue since apart from other problems that notori-ously plague survey-driven data collection, such as social de-sirability bias [20], the cultural background of the researchermay also introduce bias [1].
Limitations: Language based comparisons of culturesare limited since language is only one aspect of culture andmany different cultures and subcultures may share the samelanguage. A widely used alternative is to rely on countryborders. However, cultures do not clearly map to countriessince e.g. ethnic minorities may present a cultural groupin a country which is different from the cultural group ofthe majority. This limitation is present in all current-state-of-art studies since no sources for cultural data at differentaggregation levels exist so far. Although many Wikipedialanguage editions can be associated with one or a small num-ber of countries where the language is predominantly spo-ken, the community (i.e., the active as well as the passivepart of the community) around a certain language editionof Wikipedia may not be representative for the populationof this country. As a consequence, there is selection-biaspresent in Wikipedia data. However, sophisticated statisti-cal methods exist that could in future work be applied to thedata presented here to detect and correct such biases [10].
6. CONCLUSIONSIdentifying biases and distorted descriptions of cultures on
Wikipedia may help to guide the attention of the Wikipediacommunity towards areas where they need to invest effort(e.g., areas where the descriptions of cultural practices arebiased and/or where very high or low cross-cultural interestexist and/or where cultural practices are extremely dissim-ilar). Our results suggest that nearly no unbiased descrip-tions of cultural practices in the national food domain existon Wikipedia, leading to the suspicion that this might besimilar for other cultural practices. Unveiling those biasesmay on the one hand help to make people aware of them orcorrect them and may on the other hand reveal informationabout the relationship of different cultural communities.
Main findings: The main empirical findings of this workare: (i) shared internal states such as beliefs and values thatmay define a culture are positively correlated with sharedculinary practices, (ii) neighboring countries tend to have
more similar cultural practices than more distant countries,(iii) cultural understanding can in part be explained by theglobal importance of a food culture (e.g., the French cui-sine is ubiquitously better understood than the Bulgariancuisine) and by migration (e.g., Turkish food culture is thethird best understood food culture in Europe with an av-erage understanding of 0.04 across all language editions -after French and Italian food culture which both have anaverage understanding score of 0.06 - and Turkey has thesecond highest emigration in Europe), (iv) almost all Euro-pean language communities are most interested in their owncultural practice and only half of them show more inter-est in their neighbors’ food culture than in the food cultureof non-neighboring communities, (v) high understanding be-tween two cultural distinct communities does not necessarilylead to high interest or vice versa, but high similarity tendsto correlate positively with both, understanding and affini-ties. That means, similar cultural groups (i.e., groups whichshare cultural practices) do not only understand each other,but also show high mutual interest, while dissimilar culturalgroups that understand each other do not necessarily showhigh interest in each other.
Contributions: The contributions of this work are two-fold: (i) we present a novel approach for mining cultural rela-tions between communities using the access volume and thecontent of the description of cultural practices on differentlanguage editions of Wikipedia and present its applicationand results based on the example of the culinary practicesof 31 European countries; (ii) we validate our approach us-ing several external datasets, crowdsourcing methods andsimulations to show that the method is viable and meritsfurther research and extension, as it shows much potentialas a non-reactive way of empirically exploring cultural rela-tions online.
7. REFERENCES[1] G. Ailon. Mirror, mirror on the wall: Culture’s
consequences in a value test of its own design. Academy ofManagement Review, 33(4):885–904, 2008.
[2] S. B. Alavi and J. McCormick. Theoretical andmeasurement issues for studies of collective orientation inteam contexts. Small Group Research, 35:111–127, 2004.
[3] P. Bao, B. Hecht, S. Carton, and M. e. a. Quaderi.Omnipedia: Bridging the wikipedia language gap. InProceedings of the SIGCHI Conference on Human Factorsin Computing Systems, CHI ’12, pages 1075–1084, NewYork, NY, USA, 2012. ACM.
[4] P. Bourdieu. Outline of a Theory of Practice. CambridgeUniversity press, 1977.
[5] P. Bourdieu. Distinction: A Social Critique of theJudgement of Taste. Harvard University Press, 1984.
[6] E. Callahan and S. C. Herring. Cultural bias in wikipediacontent on famous persons. Journal of the Association forInformation Science and Technology, 62(10):1899–1915,2011.
[7] M. Calvo. Migration et alimentation. Social ScienceInformation, 21(3):383–446, 1982.
[8] E. Filatova. Multilingual wikipedia, summarization, andinformation trusworthyness. SIGIR workshop oninformation access in a multilingual world, 2009.
[9] D. Garcıa and D. Tanase. Measuring cultural dynamicsthrough the eurovision song contest. Advances in ComplexSystems, 16, 2013.
[10] C. Gary. Detecting and statistically correcting sampleselection bias. Journal of Social Service Research,3(30):19–33, 2004.
[11] R. O. G. Gavilanes, D. Quercia, and A. Jaimes. Culturaldimensions in twitter: Time, individualism and power. InInternational AAAI Conference on Weblogs and SocialMedia (ICWSM), 2013.
[12] B. Hecht and D. Gergle. Measuring self-focus bias incommunity-maintained knowledge repositories. In J. M.Carroll, editor, Proceedings of the Fourth InternationalConference on Communities and Technologies, pages11–20. ACM, 2009.
[13] B. Hecht and D. Gergle. The tower of babel meets web 2.0:user-generated content and its applications in amultilingual context. In Conference on Human Factors inComputing Systems, pages 291–300. ACM, 2010.
[14] G. H. Hofstede. Culture’s consequences: Internationaldifferences in work-related values. Sage Publications,Beverly Hills, CA, 1980.
[15] J. M. Jermier, J. W. Slocum, L. W. Fry, and J. Gaines.Organizational Subcultures in a Soft Bureaucracy:Resistance behind the Myth and Facade of an OfficialCulture. Organization Science, 2(2):170–194, 1991.
[16] S. Kayan, S. R. Fussell, and L. D. Setlock. Culturaldifferences in the use of instant messaging in asia and northamerica. In Proceedings of the 2006 20th AnniversaryConference on Computer Supported Cooperative Work,CSCW ’06, pages 525–528, New York, NY, USA, 2006.ACM.
[17] D. E. Leidner and T. Kayworth. Review: A review ofculture in information systems research: Toward a theory ofinformation technology culture conflict. MIS Q.,30(2):357–399, June 2006.
[18] P. Massa and F. Scrinzi. Manypedia: Comparing languagepoints of view of wikipedia communities. First Monday,18(1), 2013.
[19] H. Maurer and J. Kolbitsch. The transformation of theweb: How emerging communities shape the information weconsume. Journal of Universal Computing Science,12(2):187–213, 2006.
[20] P. E. Meehl and S. R. Hathaway. The k factor as asuppressor variable in the minnesota multiphasicpersonality inventory. Journal of Applied Psychology,30(5):525, 1946.
[21] M. Mohring. Transnational food migration and theinternalization of food consumption: Ethnic cuisine in westgermany. Food and Globalization: Consumption, Marketsand Politics in the Modern World, 1(1), 2008.
[22] J.-B. Michel and Y. K. e. a. Shen. Quantitative analysis ofculture using millions of digitized books. Science,331(6014):176–182, 2011.
[23] K. Nemoto and P. A. Gloor. Analyzing cultural differencesin collaborative innovation networks by analyzing editingbehavior in different-language wikipedias. Procedia-Socialand Behavioral Sciences, 26:180–190, 2011.
[24] S. E. Overell and S. M. Ruger. View of the world accordingto wikipedia: Are we all little steinbergs? Journal ofComputational Science, 2(3):193–197, 2011.
[25] M. Rask. The reach and richness of wikipedia: Iswikinomics only for rich countries. First Monday, 13(6),2008.
[26] J. Roose. Der Index kultureller Ahnlichkeit - Konstruktionund Diskussion. Freie Universitat Berlin, 2010.
[27] E. H. Schein. How Culture Forms, Develops, and Changes.In R. H. Kilmann, M. J. Saxton, and R. Serpa, editors,Gaining Control of the Corporate Culture, pages 17–43+.Jossey Bass, 1985.
[28] T. H. Silva, P. O. S. V. de Melo, J. Almeida, M. Musolesi,and A. Loureiro. You are what you eat (and drink):Identifying cultural boundaries by analyzing food and drinkhabits in foursquare. In International AAAI Conference onWeblogs and Social Media (ICWSM), 2014.
[29] L. Spierdijk and M. Vellekoop. Geography, culture, andreligion: Explaining the bias in eurovision song contest
voting, 2006.[30] T. Talhelm, X. Zhang, S. Oishi, S. Chen, D. Duan, X. Lan,
and S. Kitayama. Large-scale psychological differenceswithin china explained by rice versus wheat agriculture.Science, 344(6184):603–608, 2014.
[31] W. Tobler. A computer movie simulating urban growth inthe detroit region. Economic Geography, 46(2):234–240,1970.
[32] M. Warncke-Wang, A. Uduwage, Z. Dong, and J. Riedl. Insearch of the ur-wikipedia: Universality, similarity, andtranslation in the wikipedia inter-language link network. InProceedings of the Eighth Annual International Symposiumon Wikis and Open Collaboration, WikiSym ’12, pages20:1–20:10, New York, NY, USA, 2012. ACM.
[33] J. West and J. L. Graham. A linguistic-based measure ofcultural distance. MIR: Management International Review,44(3), 2004.
[34] K. Yanai, K. Yaegashi, and B. Qiu. Detecting culturaldifferences using consumer-generated geotagged photos. InLocWeb, ACM International Conference Proceeding Series,page 12. ACM, 2009.
[35] Y.-X. Zhu, J. Huang, and Z.-K. e. a. Zhang. Geographyand similarity of regional cuisines in china. ComputingResearch Repository (CoRR), 2013.
