=Paper=
{{Paper
|id=Vol-2448/SSS19_Paper_Upload_225
|storemode=property
|title=A Cross-lingual Analysis on Culinary Perceptions to Understand the Cross-cultural Difference
|pdfUrl=https://ceur-ws.org/Vol-2448/SSS19_Paper_Upload_225.pdf
|volume=Vol-2448
|authors=Xuehui Leng,Masanao Ochi,Takeshi Sakaki,Junichiro Mori,Ichiro Sakata
|dblpUrl=https://dblp.org/rec/conf/aaaiss/LengOSMS19
}}
==A Cross-lingual Analysis on Culinary Perceptions to Understand the Cross-cultural Difference==
https://ceur-ws.org/Vol-2448/SSS19_Paper_Upload_225.pdf
A Cross-lingual Analysis on Culinary Perceptions
to Understand the Cross-cultural Difference
Xuehui Leng Masanao Ochi Takeshi Sakaki Junichiro Mori Ichiro Sakata
The University of Tokyo The University of Tokyo Hotto Link Inc. The University of Tokyo The University of Tokyo
snopict@gmail.com
Abstract tion as a possible substitute for traditional laborious investi-
gations, and many works proposed the framework for ana-
Native cuisines are usually acquired tastes for local people
lyzing food culture fusing data acquired from social media
but often appear to be unpalatable choices for foreign trav-
elers. This is attributable not only to taste preferences, but (Silva et al. 2014; Sajadmanesh et al. 2017). These works
also to the different dish perceptions between cultures. Un- succeed in comparing cuisines from a worldwide perspec-
derstanding culinary cultures is of great interest these days, tive, but they are still inadequate for providing detailed infor-
with many studies investigating the ingredient uses and tastes mation about experiences aside from those involving food.
among dishes, while little is known about the subjective per- Some framework is needed for comparing culinary cultures
ceptions people have toward them. In this paper, we intro- for dishes among local communities.
duce a new dimension to the understanding of culinary cul- As described in this paper, we introduce a new dimen-
tures. We specifically assess ‘cross-cultural perceptions of sion to the understanding of culinary cultures. We specifi-
food’, and develop a model to compare culinary perceptions
between cultures, especially different linguistic areas, using
cally assess ‘cross-cultural perceptions of food’, and develop
descriptions acquired from social media. To evaluate the va- a model to compare culinary perceptions between cultures,
lidity of the proposed model, we conducted a series of exper- especially different linguistic areas, using descriptions ac-
imental assessments using beverages as targets. We analyzed quired from social media. For example, an English-language
the perceptual differences between drinks in Japanese cul- tweet says that the poster does not like ramen but desires
ture and US culture. Results revealed social trends of culinary to eat a hot and spicy noodle, while a Japanese tweet says
cultures. For instance, the results show US residents’ prefer- that the poster wants to eat ramen with ‘umami’ at night in
ences for colorful foods, Japanese people’s general recogni- midwinter. What can be said from these posts? At least, the
tion of ‘bitter’ for coffee, US residents’ frequent mention of poster of the English tweet prefers hot and spicy noodles,
smoothie thickness, and the unique latte art culture in Japan. and the poster of the Japanese tweet imagines ‘midwinter’
Such perceptual differences are expected to be useful for
strategies of localization, and for recipe recommendation.
and ‘umami’ from ramen.
These descriptions used unconsciously on social media
reflect the preferences and values of a particular culture. In
Introduction the proposed framework, the cross-cultural attitude about a
Food is an integral part of our life and culture. Cuisines re- specific food is understood by analyzing the description pat-
flect culinary habits, lifestyles, and the regional cultures of terns used in a specific cultural area. We first gather descrip-
ethnic groups. Native cuisines are usually acquired tastes tions related to the targets of the analysis from Social Me-
for local people but often appear to be unpalatable choices dia. Next, descriptions with close meanings are categorized
for foreign travelers. This difference of preferences is at- as the same concept, so that descriptions can be compared
tributable not only to the tastes but also to a sense of dis- in a unified scale between languages. By comparing the fre-
comfort that arises from unfamiliar dish perceptions and en- quency of concepts used in each linguistic area, it is pos-
vironments in which the dishes are offered. People from sible to analyze the perceptual differences of food between
different areas differ in their sensitivities, perceptions of cultures.
things, and preferences and values. Even a cup of coffee To evaluate the validity of the proposed model, we con-
made from the same materials can elicit widely diverse com- ducted a series of experimental assessments using beverages
ments depending on the area in which it is tasted. Under- as targets. We analyzed the perceptual differences between
standing such cross-cultural culinary perceptions and prefer- drinks in Japanese culture and US culture with three exper-
ences is of great interest these days. Food companies trying iments. Results revealed social trends of culinary cultures.
to expand their businesses overseas expend vast amounts of For instance, the results show US residents’ preferences for
money and time to investigate the culinary habits of different colorful foods, Japanese people’s general recognition of ‘bit-
areas, which are usually based on communications or ques- ter’ for coffee, US residents’ frequent mention of smoothie
tionnaires administered to local people; a time-consuming thickness, and the unique latte art culture in Japan. Such per-
and expensive processes. Social media are attracting atten- ceptual differences are expected to be useful for strategies of
localization, and for recipe recommendation. Recipe Recommendation
The contributions of this paper can be summarized as pre- Many efforts have been undertaken for developing recipe
sented below. recommendation algorithms. User ratings of recipes are
• We introduce a model to compare culinary perceptions thought to be fundamentally important for recipe recom-
between languages using descriptions on social media. mendation (Freyne, Berkovsky, and Smith 2011; Kotonya,
• We introduce ‘perception’ as a criterion for analyzing the De Cristofaro, and De Cristofaro 2018). Most recipe recom-
distance of culinary habits. mendation algorithms are based on recipe similarities. In-
formation related to ingredients is a popular elements used
• We suggest the possibility of applying the model to recipe for calculating similarity (Teng, Lin, and Adamic 2012;
analysis and retrieval, and food and restaurant recommen- Kuo et al. 2012; Ueda et al. 2014). Other recipe attributes
dation. such as flavors and nutrients are attracting attention these
days (Min et al. 2017; Ahn et al. 2011). Furthermore, vi-
Related Work sual information can be a source for recipe recommendation
Our work is closely related to two research fields: culinary (Maruyama, Kawano, and Yanai 2012; Herranz, Min, and
habits and social media, and recipe recommendation. Jiang 2018).
Some studies evaluate the most suitable algorithms for
Culinary Habits and Social Media certain environments (Berkovsky and Freyne 2010). A per-
sonalized recipe suggestion system can support its users
With the spread of social media and review services, data in considering the balance of nourishment (Mino and
related to people’s food preferences and recipe information Kobayashi 2009), and in making health-aware meal choices
are becoming easier to access. Many studies use these data (Geleijnse et al. 2011; Ge, Ricci, and Massimo 2015; van
to analyze the culinary habits of people. Wagner et al. used Pinxteren, Geleijnse, and Kamsteeg 2011).
server log data to understand online food preferences and Our research provides a new dimension for recipe recom-
suggested the influence ingredients have on recipe prefer- mendation: people’s subjective perceptions of food. By con-
ences(Wagner, Singer, and Strohmaier 2014). Abbar et al. sideration of objective aspects of food, and subjective per-
examined the potential of Twitter to lend insight into US- ceptions of food, it is highly probable that the recommenda-
wide dietary choices (Abbar, Mejova, and Weber 2015). tion accuracy and generalization capability can be increased.
Location data from the social app ‘Untapped’ revealed the
drinking habits of numerous users and helped highlight im-
portant behavioral trends (Chorley et al. 2016). Laufer et al. Proposal Method
conducted cross-cultural analysis of food cultures by explor- This section proposes a framework to detect cross-
ing descriptions on Wikipedia (Laufer et al. 2015). cultural differences on culinary perceptions between lan-
Some studies have specifically examined cross-cultural guages. First, we gather descriptions related to the targets of
analysis. Kular et al. used network analysis to elucidate the the analysis from social media. To compare the descriptions
relation between cuisine and culture (Kular, Menezes, and in a scale beyond languages, descriptions with close mean-
Ribeiro 2011). They demonstrated that cultures can be de- ings are treated as the same concept. When defining con-
fined by similarities in how people prepare food. Sajad- cepts, we use a database showing correspondence between
manesh et al. analyzed ingredients, flavors, and nutritional words of multiple languages, construct a network with edges
values that distinguish dishes from different regions (Sajad- representing the closeness of meaning, and regard expres-
manesh et al. 2017). Silva et al. identified cultural bound- sions gathered at close positions on the network as one con-
aries by analyzing food and drink habits in Foursquare (Silva cept. Finally, by comparing the frequency of concepts used
et al. 2014). Min et al. performed cross-region recipe analy- in each linguistic area, it is possible to analyze the perceptual
sis by jointly using the recipe ingredients, food images, and differences of food between cultures.
attributes such as the cuisine and course(Min et al. 2018). Gathering Descriptions. The first task is formulating a
Silva et al. proposed a methodology to identify cultural target list and gathering descriptions about them from the
boundaries and similarities across populations at different internet. Microblogs such as Twitter and review sites such
scales (Silva et al. 2017). On a smaller scale, city size can as Yelp are suitable for the source. Then one gathers posts
indicate the influence of dietary choices (Cheng, Rokicki, that describe each target, and obtains as many descriptions as
and Herder 2017). possible. Because the purpose is to extract personal remarks,
These works compared cuisines from a worldwide per- one must exclude posts that are not user’s utterances (e.g.,
spective, but they are inadequate for providing detailed in- re-tweets and bot) before processing data.
formation related to what experience people have aside from Then, descriptive expressions must be extracted from
those related to food. A framework is needed for comparing these data. Descriptive expressions include expressions
culinary cultures for each dish among local communities. about the state of the subject, the poster’s emotional expres-
Our work differs from those in that we specifically exam- sions about the subject, and the environmental characteris-
ine the ‘perception’ of food. We introduce ‘perception’ as a tics in which the object exists. For our research, we regard
new criterion for analyzing the distance of culinary habits. adjectives as descriptive expressions. For example, what is
We also propose a framework to discuss detailed culinary ‘pretty’ and what is ‘beautiful’ depend on the speaker’s per-
differences of specific dishes. ception. Adjectives are the easiest way to describe some-
thing. Therefore, we think that a frequency distribution of 2008) to calculate the modularity maximization.
adjectives can represent the tendencies of cross-cultural per- C0 = {cozy, cosy, snug, comf y, comf ortable(ja)}
ceptions. is an example of a concept cluster, which is a subset of
Constructing Description Network. In this procedure, a adjectives with green color in Figure 1.
description network is created by attaching weighted con- Calculating the Concept Frequency Distribution. In
nections to expression pairs that have close meanings. To this procedure, sum up the co-occurrence frequency by each
incorporate consideration of the cross-lingual word relation, concept. Calculate the ‘concept frequency distribution’. De-
a database giving the relation of words of multiple languages tails of processes are explained in the following text us-
is used as an input. ing relevant formulae. Presuming that the number of co-
Here, we used ConceptNet 5.5 (Speer, Chin, and Havasi occurrences between target ti and description dj is n(ti , dj ),
2017) as the multilingual database. For all adjective pairs in- the number of co-occurrence between target ti and concept
cluding English–English, Japanese–Japanese, and English– ck is calculated using the sum of co-occurrence between tar-
Japanese combinations, we assigned weighted links to ex- get and descriptions that belong to the concept cluster.
pression pairs with close meanings. There are several rea- X
sons for adopting ConceptNet. First, ConceptNet has suf- N (ti , ck ) = n(ti , dj ) (1)
ficient data volume and reliability for both English and dj ∈Ck
Japanese terms, which are the subjects of our research. In About target ti , define the co-occurrence probability of
addition, information related to the word relation can be ac- concept ck as Pti ,ck . Here, Pti ,ck is described using the fol-
quired easily using API. Simultaneously, information related lowing formula.
to closeness (weights of 0–1) can be acquired.
N (ti , ck )
Figure 1 presents an example of description network with Pti ,ck = Pkmax (2)
English and Japanese adjectives. Adjectives with similar k=1 N (ti , ck )
meanings are located in a near place irrespective of the lan- For target ti , the frequency distribution of the target’s co-
guage. In the figure, colors are displayed separately for re- occurrence with concept ck is calculable with the following
spective clusters. We visualized the graph using the Gephi formula. We define this distribution as Pti and designate it
software 1 . as the concept frequency distribution.
Pti = {Pti ,ck , ∀ck ∈ C}
P (3)
ck Pti ,ck = 1
Analyzing Cross-cultural perceptions using Concept
Frequency Distribution. In this procedure, we introduce
two methods that can be used to analyze cross-cultural per-
ceptions using the concept frequency distribution.
• JS Divergence of Concept Frequency Distribution:
Calculate the JS divergence of the concept frequency dis-
tribution. This figure is regarded as a barometer of the
closeness between targets. JS divergence approaches zero
if the two probability distributions are similar. In fact, it
gets closer to one for different shapes. Using this condi-
tion, JS divergence of two concept frequency distributions
can represent the distance between two targets. A small
Figure 1: Example of a description network. value of the JS divergence between two concept frequency
distributions indicates that two targets are close in their
perceptions, and vice versa.
Defining Concepts. For the description network, one
must then conduct clustering and categorize descriptions • Content of Concept Frequency Distribution: To see de-
into several clusters. Because the edge represents close- tails of the difference, one must look into the content of
ness in meaning, each cluster is a set of terms with similar the concept frequency distribution. One must then check
meanings. Because some edges connect different languages, the concept clusters with a large gap in their frequen-
terms from different languages with close meanings also be- cies between languages, which can suggest perceptual
long to the same cluster. Defining each cluster as a ‘con- differences related to food. For example, assuming that
cept cluster’ that represents a concept which exceeds lan- you want to compare the perceptual differences toward
guage, then similarity measurements and other analysis are ‘dog’ between English-speakers and Japanese-speakers.
performed using this concept cluster classification. You will find English-speakers are more familiar to ‘hand-
Modularity maximization (Newman and Girvan some’ dogs than Japanese-speakers if descriptions be-
2004) is used for network clustering. For our re- longing to ‘handsome’ cluster appear more in English
search, we used Louvain method (Blondel et al. context than in Japanese context. From here, the differ-
ences of general impressions toward dogs between En-
1
https://gephi.org/ glish and Japanese cultures can be inferred.
Experimental Evaluation three analyses.
Based on the method of detecting differences in culinary per- (1) Comparing the general perceptions of beverages.
ceptions between languages introduced in the preceding sec- Is it possible to extract cultural culinary perception differ-
tion, we conducted an experimental evaluation. ences from descriptions? Is it appropriate to use adjectives
as descriptions? To address these questions, we used all En-
Datasets glish and Japanese tweets to evaluate the extractability of
We used Twitter to gather descriptions. There are several general perception differences. We analyzed the contents of
reasons for this. First, Twitter is extremely popular as pri- concepts with a wide gap separating English and Japanese
vate media to deliver opinions and feelings every day. Sec- in the concept frequency distribution. Because most English-
ondly, posts are mainly given by a short text that delivers speaking users on Twitter are from the US, we compared our
the poster’s opinion directly. These characteristics suit our results with generally known differences between Japanese
purpose of gathering descriptions. We chose targets of anal- and US residents.
ysis based on the categories of online recipe service. English (2) Comparing networks of beverage perceptions. In
targets are consulted from ‘Drinks Recipes’ category of all- our research, the JS divergence of concept frequency distri-
recipes2 . Japanese targets are consulted from the ‘Drinks’ bution is regarded as a barometer of the closeness between
category of Rakuten Recipe3 . Each target is used as the targets. To evaluate the validity, we draw a network based
query for gathering descriptions. They are presented in Ta- on JS divergence, for English and Japanese target list re-
ble 1. spectively, and assess whether it can represent the relation
of targets, or not. When drawing a beverage network, each
Table 1: English and Japanese targets for the experiment. target is regarded as a node. Edges are provided to target
Queries with a too small number of data acquired is excluded pairs with JS divergence under a certain threshold. As the
from the list. weight wij added on the edge eij linking node ti and tj , we
used the following value.
language targets
English cider, cocktail, coffee, eggnog, juice, lemon- wij = 1 − DJS (Pti k Ptj ) (4)
ade, liqueur, mulled wine, punch, shot, Here, Pti and Ptj respectively denote the concept frequency
shake, float, smoothie, tea distribution of target ti or target tj .
Japanese coffee, hot chocolate, tea, matcha, soy (3) Cross-lingual analysis of specific targets. In this
milk, yogurt, honey drink, mixed juice, paper, we introduce that by looking into the concept fre-
shake, smoothie, chai, beer, Japanese dis- quency distribution, perceptual differences can be extracted.
tilled spirit, plum wine, amazake, cocktail, To evaluate the validity, we compared the concept fre-
Mojito, healthy liquor quency distributions of specific drinks. Among the target list
items, English and Japanese share some targets: coffee, tea,
We acquired 1.04 million tweets and 34,000 tweets, re- smoothies, and juice. We used them for the experiment, and
spectively, for English and Japanese. As a preprocessing by comparing the results with social and cultural environ-
step, we excluded those tweets that include ‘RT’ at the be- ments around us, evaluated the method’s validity.
ginning of the content in the process of morphological analy-
sis. Additionally, we limited the users to ‘Twitter for iPhone’ Results and Discussion
and ‘Twitter for Android’ because the data needed are raw Comparing the General Perception of Beverages
descriptions by individuals. After preprocessing, 256,000
tweets for English and 12,000 tweets for Japanese remained. Results of comparisons of general perceptions of bever-
These were used for subsequent analyses. ages are presented in Figure 2. The value in the graph
For the adjectives acquired after morphological analysis, shows the co-occurrence frequency in English minus the co-
we limited their number to create description network so that occurrence frequency in Japanese for each concept cluster.
the calculation can be finished in time. After defining con- Concept clusters with positive values appear more in the En-
cepts, we also limited the concept clusters used for the cal- glish context. Concept clusters with negative values appear
culation of concept frequency distribution so that the sub- more in the Japanese context.
sequent analyses do not become too confusing. As a result, In English, Concept 13 (white, young, blue), Concept 17
30 concept clusters are acquired. Table 2 presents the rep- (refreshing, cool, icy), Concept 18 (thick, giant, massive),
resentative adjectives belonging to each concept cluster. For and Concept 26 (much, few, many) show higher frequency
the figures in the following sections, we will specify concept than in Japanese. Among these, the high frequency of Con-
clusters using the index shown in Table 2. cept 26 (much, few, many) can be attributed to the English
language system.
Evaluation Methods Concept 13 (white, young, blue) is a cluster of colors.
To check the validity of the framework for detecting cross- The result demonstrating that English-speakers mention
cultural differences on culinary perceptions, we conducted color more agrees with the fact that the US has more col-
orful foods. From long ago, edible dyes have been used in
2
http://allrecipes.com/ the US to prevent food from bruising and to make food eye-
3
https://recipe.rakuten.co.jp/ catching, which also facilitates coloration of products by US
Table 2: Concept clusters and representative adjectives that belong to each cluster
index representative adjectives index representative adjectives
1 fast, snap, quick, quick(ja), early(ja) 16 spicy, bloody, bad-smelling(ja), acrid(ja)
2 salted, salty, salty(ja) 17 refreshing, cool, chilly, cold(ja), new(ja)
3 difficult, strong, bitter, sour, strong(ja), heavy(ja), 18 thick, massive, fat, great, deep, thick(ja), greasy(ja),
bitter(ja) deep(ja)
4 weak, soft, cheap, simple, light, soft(ja), light(ja), 19 like, -like(ja)
brittle(ja), cheap(ja)
5 little, poor, slight, thin(ja), poor(ja) 20 old, old(ja)
6 nervous, excited, warm, hot, hot(ja), affable(ja), 21 proper, decent, modest(ja)
warm(ja)
7 interesting, hilarious, curious, exciting, interest- 22 clean, refreshing(ja)
ing(ja), pleasant(ja), fun(ja)
8 asleep, sleepy, sleepy(ja) 23 brown, brown(ja)
9 fantastic, dangerous, awful, scary, dangerous(ja), 24 mad, crazy, insane, crazy(ja)
great(ja), amazing(ja)
10 bad, evil, tired, ugly, awkward, bad(ja), not deli- 25 rare, rare(ja)
cious(ja), annoying(ja)
11 blunt, boring, late, obtuse, late(ja), boring(ja), 26 much, few, many, multiple, several, few(ja),
lax(ja), blunt(ja) many(ja)
12 high, expensive, long, tall, short, short(ja), low(ja), 27 relaxing, comfy, soothing, friendly, cozy, comfort-
venerable(ja), shallow(ja), long(ja) able(ja), agreeable(ja)
13 white, young, blue, yellow, dark, black(ja), 28 busy, busy(ja)
white(ja), young(ja), dark(ja), yellow(ja)
14 near, close, near(ja) 29 sorry, sorry(ja)
15 pretty, adorable, cute, beautiful, lovely, pretty(ja), 30 flat, flat(ja)
nostalgic(ja), beautiful(ja)
food companies. US residents tend to regard colorful food (e.g., ‘cheap’ drinks and ‘light’ drinks). In Japan, more op-
as tempting. For example, Gatorade, which is a US brand tions are presented in beverage vending machines than in the
and which boasts the world’s top share among sports drinks, US. A greater variety of beverages are available at conve-
are colored with different colors for each type. Despite be- nience stores. Although further investigation is needed, one
ing drunk in more than 70 countries worldwide, Gatorade is might infer that Japanese people are more sensitive about
not currently produced in Japan. In Japan, loud colors from daily drinks because of the wide variety of beverages in their
edible dyes give a strong unhealthy and even harmful im- everyday environment.
pression. Loud coloring of food is not preferred.
In Japanese, Concept 3 (hard, strong, bitter), Concept 4
(soft, light, cheap), Concept 7 (interesting, fun), and Concept
9 (fantastic, great) show higher frequency than English.
The high frequencies of Concept 3 (hard, strong, bit-
ter), Concept 7 (interesting, fun), and Concept 9 (fantas-
tic, great) suggest that Japanese people post more nega-
tive opinions and emotions related to themselves. This re-
sult corresponds with those of earlier research. Vidal et al.
analyzed English-written Twitter data for breakfast, lunch,
dinner and snack occasions, and performed manual con-
tent analyses of tweets for deeper insights. They found con-
textual characteristics of eating occasions that were fre-
quently described in tweets, whereas emotions were rarely
described in tweets(Vidal et al. 2015). Acar et al. reported
that Japanese university students post mainly about them-
Figure 2: Gap of the concept frequency distribution between
selves, whereas many US students’ posts include questions
English and Japanese tweets. Concept clusters with positive
(Acar and Deguchi 2013). Our results indicate a similar ten-
values appear more in the English context; negative in the
dency for characteristics of social media usage for Japanese
Japanese context.
and US residents.
The high frequency of Concept 4 (soft, light, cheap) might
represent Japanese people’s interest in everyday drinks
(refreshing, cool) also appears more frequently in the En-
glish context. This frequent appearance is presumably be-
cause ‘refreshing’ and ‘cool’ belong to the same concept.
Further investigation separating these two concepts might
be needed. Higher frequency of Concept 11 (blunt, boring,
sharp) in English implies that US residents often see coffee
in a ‘boring’ context, and desire ‘exciting’ coffee. In fact,
many articles on the internet introduce how to make coffee
not boring.
In Japanese posts, however, Concept 3 (hard, strong, bit-
ter) and Concept 15 (dear, pretty, nostalgic) appear more fre-
(a) English (b) Japanese quently than in English. From the high appearance of Con-
cept 3 (hard, strong, bitter), one can infer the general percep-
Figure 3: Beverage networks. tion of ‘bitter’ related to coffee in Japan. Concept 15 (dear,
pretty, nostalgic) presumably reflects the Japanese unique
latte art culture. Latte art is not rare in the US, but no cus-
Comparing Networks of Beverage Perceptions tom of drawing animals and characters exists as in Japan.
Results of beverage networks of perception comparisons are One can infer that there is no general recognition of ‘cute’
shown in Figure 3. We visualized these networks using the for coffee and latte art in the US.
Gephi software.
tea In the English context, Concept 13 (white, young,
In the English beverage network, juice, lemonade, and tea
blue) and Concept 15 (pretty, very, adorable) appears more
appear in adjacent positions, reflecting the fact that these
than in Japanese. In the US, fruit tea is popular. Specialty tea
three beverages often have fruit among their ingredients. In
shops sell widely diverse kinds of tea, for which color also
the Japanese beverage network, hot chocolate, Amazake, tea
varies greatly. The color of tea leaves and the color of water
and honey drinks appear in proximate positions, all of which
when tea is poured are enjoyed. US residents have a culture
share the commonality of being warm drinks. In addition,
of enjoying tea with their eyes. Given the high frequency of
alcohol drinks including Mojito, Japanese distilled spirits,
Concept 15 (pretty, very, adorable) together, one can assume
plum wine, and cocktails are proximate in the network. Soy
more aspects of the US tea market. In the US, women com-
milk and smoothies are both perceived as healthy drinks.
monly give tea to their friends as a present. Colorful and
Both use milk as an ingredient. Therefore, it is not surprising
cute package decorations on tea caddies and bags are made.
that they share a close perception in Japan.
Pretty tea cutlery or accessories are manufactured and sold.
Comparison of these two networks reveals apparent dif-
ferences of positions among beverages. As an example, one smoothie In the English context, Concept 5 (little, poor,
can specifically examine the smoothie. In the English net- mini), Concept 13 (white, young, blue), Concept 15 (pretty,
work, a smoothie has edges with tea, cider, and coffee. Its very, adorable), and Concept 18 (thick, giant, massive) ap-
position is in the center of the network. However, in the pear more frequently than in Japanese. The high frequency
Japanese network, the only edge smoothie shares is that with of Concept 13 (white, young, blue) and Concept 15 (pretty,
soy milk: its position is at the network edge. very, adorable) shows the same tendencies as those for tea
This beverage network is based on the proximity of peo- and are presumably more affected by the colorful and
ples’ perceptions. The network represents a relation to our fruitful perception of smoothie in the US. According to
perceptions. Therefore, using JS divergence of the concept the high frequency of Concept 18 (thick, giant, massive),
frequency distribution to describe the distance of perception one can infer that US residents use thickness to evaluate
on the target is regarded as reasonable. smoothies. In fact, ways to make thick smoothies are intro-
duced on the Internet.
Cross-lingual Analysis on Specific Targets In the Japanese context, however, Concept 8 (sleepy),
The results of cross-lingual analysis of specific targets are Concept 9 (fantastic, great), Concept 10 (bad, not delicious,
presented in Figure 4. When comparing the contents of the troublesome), and Concept 19 (-like) appear more than in
concept frequency distribution for specific targets, we used English. The high frequency of Concept 10 (bad, not deli-
the result of Figure 2, and selected concept clusters with cious, troublesome) implies a general recognition of ‘not
larger gaps than the general perception because concepts that delicious’ in Japan. Using Concept 19 (-like) frequently
have large gaps after eliminating the influence of the general might mean a tendency to evaluate smoothies in the com-
tendency (e.g. usage of social media) are highly likely to parison with other beverages in Japan. This result sug-
differ in their perception by cultures. gests the possibility that smoothies are not so common and
popular in Japan, and that they are recognized as a substitute
coffee In the English context, Concept 6 (warm, excited, for other drinks in Japan.
nervous) appears more than in Japanese. Iced coffee origi-
nated in Japan. Because US residents are not familiar with juice In the English context, Concept 5 (little, poor, mini)
iced coffee, the impression of ‘hot’ or ‘warm’ might be Concept 17 (refreshing, cool, icy) and Concept 26 (much,
stronger in the US than in Japan. However, Concept 17 few, many) appear more frequently than in Japanese. High
Figure 4: Results of cross-lingual analysis of specific targets. The black line shows the gap of the concept frequency distribution
between English and Japanese tweets related to beverages, which reveals the same result as that portrayed Figure 2. Each bar
shows the gap of Concept Frequent Distribution between English and Japanese for each target. For each target, concept clusters
with larger absolute values than the whole beverages shown in the black line are selected for deeper investigation.
appearance of Concept 5 (little, poor, mini)and Concept 26 context and environment
(much, few, many) implies that US residents are concerned These are the context in which a target appears in a cul-
about the amount of juice. High appearance of Concept tural area. US residents’ mention of ‘boring’ for coffee
17 (refreshing, cool, icy) might indicate that US residents and Japanese people’s mention of ‘fun’ for mixed juice
generally regard juice as a cold drink. are salient examples.
In the Japanese context, on the other hand, Concept 7 (in- unique genre
teresting, fun), Concept 9 (fantastic, great), and Concept 15 This is a genre unique to a culture. Latte art in Japan and
(pretty, very, adorable) co-occur more frequently than in En- hot cider in the US (not described in the results) are note-
glish. One can assume that there for Japanese people, juice worthy examples.
is a drink appearing in scenes that evoke positive emo-
tions. These perceptual differences can be useful for localiza-
tion strategies of food companies. Localization strategies are
Conclusion processes of understanding local culture, habits and prefer-
In this work, we propose a framework to compare cross- ences used to assess what changes are necessary to spread
cultural culinary perceptions by grouping descriptions with a company’s products. Our model can at least perform well
the meanings of words. To evaluate the validity of the pro- as a cost-effective and time-effective way to clarify cross-
posed method, we conducted a series of experimental eval- cultural characteristics of food culture, and give advice for
uations using beverages as targets. We analyzed the percep- strategies.
tual differences of drinks in Japanese culture and US culture In the context of recipe recommendation, we suggest a
using three experiments. The results revealed social trends of new criterion: culinary perception. With cultural bias re-
culinary cultures. Cultural differences that can be acquired garded together with objective attributes such as ingredients,
with the proposed framework can be categorized into the fol- flavors, and visual appearance, the recommendation algo-
lowing four groups. rithm would move on to the next stage. This progression is
not merely for recipe recommendation; it applies widely to
general recognition a recommendation of any kind. Methods to detect and ma-
This is the general recognition of targets in a cultural area. nipulate within-group biases properly are needed.
Japanese popular impressions of ‘not delicious’ assigned As future work, we plan to build a recipe recommenda-
to smoothies and ‘bitter’ for coffee, and US residents’ im- tion algorithm based on culinary perceptions. We also wish
pressions of ‘cold’ for juice are salient examples. to investigate how elements of food (ingredients, flavors, nu-
values and standards trients, visual appearance, and perceptions) are mutually re-
These are values regarded as important in a cultural area lated and how they influence our culinary preferences. Find-
and as standards for evaluating the target in a cultural ing the answer to this question is expected to provide a better
area. US residents’ frequent mention of smoothie thick- recommendation system with high generalization capability,
ness, juice amount, and tea color are salient examples. which can suggest a choice with healthier, more satisfactory
and more diverse experience for any person in the world. Maruyama, T.; Kawano, Y.; and Yanai, K. 2012. Real-time mo-
bile recipe recommendation system using food ingredient recogni-
References tion. In Proceedings of the Second ACM International Workshop
on Interactive Multimedia on Mobile and Portable Devices, 27–34.
Abbar, S.; Mejova, Y.; and Weber, I. 2015. You tweet what you ACM.
eat: Studying food consumption through twitter. In Proceedings of Min, W.; Jiang, S.; Wang, S.; Sang, J.; and Mei, S. 2017. A deli-
the 33rd Annual ACM Conference on Human Factors in Computing cious recipe analysis framework for exploring multi-modal recipes
Systems, CHI ’15, 3197–3206. New York, NY, USA: ACM. with various attributes. In Proceedings of the 2017 ACM on Multi-
Acar, A., and Deguchi, A. 2013. Culture and social media usage: media Conference, MM ’17, 402–410. New York, NY, USA: ACM.
analysis of japanese twitter users. International Journal of Elec- Min, W.; Bao, B.-K.; Mei, S.; Zhu, Y.; Rui, Y.; and Jiang, S. 2018.
tronic Commerce Studies 4(1):21. You are what you eat: Exploring rich recipe information for cross-
Ahn, Y.-Y.; Ahnert, S. E.; Bagrow, J. P.; and Barabási, A.-L. 2011. region food analysis. IEEE Transactions on Multimedia 20(4):950–
Flavor network and the principles of food pairing. Scientific Re- 964.
ports 1:196. Mino, Y., and Kobayashi, I. 2009. Recipe recommendation for a
Berkovsky, S., and Freyne, J. 2010. Group-based recipe recom- diet considering a user’s schedule and the balance of nourishment.
mendations: Analysis of data aggregation strategies. In Proceed- In Intelligent Computing and Intelligent Systems, 2009. ICIS 2009.
ings of the Fourth ACM Conference on Recommender Systems, IEEE International Conference on, volume 3, 383–387. IEEE.
RecSys ’10, 111–118. New York, NY, USA: ACM. Newman, M. E., and Girvan, M. 2004. Finding and eval-
Blondel, V. D.; Guillaume, J.-L.; Lambiotte, R.; and Lefebvre, E. uating community structure in networks. Physical Review E
2008. Fast unfolding of communities in large networks. Journal of 69(2):026113.
Statistical Mechanics: Theory and Experiment 2008(10):P10008. Sajadmanesh, S.; Jafarzadeh, S.; Ossia, S. A.; Rabiee, H. R.; Had-
Cheng, H.; Rokicki, M.; and Herder, E. 2017. The influence of city dadi, H.; Mejova, Y.; Musolesi, M.; Cristofaro, E. D.; and Stringh-
size on dietary choices. In Adjunct Publication of the 25th Confer- ini, G. 2017. Kissing cuisines: Exploring worldwide culinary
ence on User Modeling, Adaptation and Personalization, 231–236. habits on the web. In Proceedings of the 26th International Con-
ACM. ference on World Wide Web Companion, WWW ’17 Companion,
1013–1021. Republic and Canton of Geneva, Switzerland: Inter-
Chorley, M. J.; Rossi, L.; Tyson, G.; Williams, M. J.; et al. 2016. national World Wide Web Conferences Steering Committee.
Pub crawling at scale: Tapping untapped to explore social drinking.
Silva, T. H.; de Melo, P. O. V.; Almeida, J. M.; Musolesi, M.; and
In ICWSM, 62–71.
Loureiro, A. A. 2014. You are what you eat (and drink): Iden-
Freyne, J.; Berkovsky, S.; and Smith, G. 2011. Recipe rec- tifying cultural boundaries by analyzing food and drink habits in
ommendation: accuracy and reasoning. In International Confer- foursquare. In Proceedings of the Eighth AAAI International Con-
ence on User Modeling, Adaptation, and Personalization, 99–110. ference on Weblogs and Social Media (ICWSME4).
Springer. Silva, T. H.; de Melo, P. O. V.; Almeida, J. M.; Musolesi, M.; and
Ge, M.; Ricci, F.; and Massimo, D. 2015. Health-aware food rec- Loureiro, A. A. 2017. A large-scale study of cultural differences
ommender system. In Proceedings of the Ninth ACM Conference using urban data about eating and drinking preferences. Informa-
on Recommender Systems, 333–334. ACM. tion Systems 72:95–116.
Geleijnse, G.; Nachtigall, P.; van Kaam, P.; and Wijgergangs, L. Speer, R.; Chin, J.; and Havasi, C. 2017. Conceptnet 5.5: An open
2011. A personalized recipe advice system to promote healthful multilingual graph of general knowledge. In AAAI, 4444–4451.
choices. In Proceedings of the 16th International Conference on Teng, C.-Y.; Lin, Y.-R.; and Adamic, L. A. 2012. Recipe recom-
Intelligent User Interfaces, 437–438. ACM. mendation using ingredient networks. In Proceedings of the Fourth
Herranz, L.; Min, W.; and Jiang, S. 2018. Food recognition and Annual ACM Web Science Conference, WebSci ’12, 298–307. New
recipe analysis: integrating visual content, context and external York, NY, USA: ACM.
knowledge. arXiv preprint arXiv:1801.07239. Ueda, M.; Asanuma, S.; Miyawaki, Y.; and Nakajima, S. 2014.
Kotonya, N.; De Cristofaro, P.; and De Cristofaro, E. 2018. Of Recipe recommendation method by considering the users prefer-
wines and reviews: Measuring and modeling the vivino wine social ence and ingredient quantity of target recipe. In Proceedings of the
network. arXiv preprint arXiv:1804.10982. International MultiConference of Engineers and Computer Scien-
tists, volume 1.
Kular, D. K.; Menezes, R.; and Ribeiro, E. 2011. Using network
analysis to understand the relation between cuisine and culture. In van Pinxteren, Y.; Geleijnse, G.; and Kamsteeg, P. 2011. Deriv-
Proceedings of the 2011 IEEE Network Science Workshop, NSW ing a recipe similarity measure for recommending healthful meals.
’11, 38–45. Washington, DC, USA: IEEE Computer Society. In Proceedings of the 16th International Conference on Intelligent
User Interfaces, 105–114. ACM.
Kuo, F.-F.; Li, C.-T.; Shan, M.-K.; and Lee, S.-Y. 2012. Intelligent
Vidal, L.; Ares, G.; Machı́n, L.; and Jaeger, S. R. 2015. Using
menu planning: Recommending set of recipes by ingredients. In
twitter data for food-related consumer research: A case study of
Proceedings of the ACM Multimedia 2012 Workshop on Multime-
”what people say when tweeting about different eating situations,”.
dia for Cooking and Eating Activities, 1–6. ACM.
Food Quality and Preference 45:58–69.
Laufer, P.; Wagner, C.; Flöck, F.; and Strohmaier, M. 2015. Mining
Wagner, C.; Singer, P.; and Strohmaier, M. 2014. The nature and
cross-cultural relations from wikipedia: A study of 31 european
evolution of online food preferences. EPJ Data Science 3(1):38.
food cultures. In Proceedings of the ACM Web Science Conference,
WebSci ’15, 3:1–3:10. New York, NY, USA: ACM.