=Paper=
{{Paper
|id=Vol-2865/paper5
|storemode=property
|title=Healthy Food Depiction on Social Media: The Case of Kale on Twitter
|pdfUrl=https://ceur-ws.org/Vol-2865/paper5.pdf
|volume=Vol-2865
|authors=Maija Kāle,Ebenezer Agbozo
|dblpUrl=https://dblp.org/rec/conf/dhn/KaleA20a
}}
==Healthy Food Depiction on Social Media: The Case of Kale on Twitter==
https://ceur-ws.org/Vol-2865/paper5.pdf
Healthy Food Depiction on Social Media:
The Case of Kale on Twitter
Maija Kāle1[0000-0002-6951-9009 ] and Ebenezer Agbozo2[0000-0002-2413-3815]
1 University of Latvia, Riga, Latvia
2 Ural Federal University, Yekaterinburg, Russian Federation
maijakale@gmail.com
Abstract. This food computing study analyzes Twitter microblog entries relat-
ed to food. A particular attention here is paid to one specific healthy food –
kale. Such an approach is chosen due to kale’s popularity as a healthy food. By
applying sentiment measurement, authors contribute to the understanding of
text-based stories behind healthy food, and conclude that kale offers clear bene-
fits for consumers via its specific health-related aspects (anti-inflammatory,
immune system boosting, etc.) while taste is being left outside the discourse.
The lack of any references to taste in kale-related social media stories leads to
the conclusion that healthy food descriptions in general are short of hedonistic
manifestations. Subsequently, this can be seen as one of the drawbacks of
healthy food presentation textually vis-a-vis comfort foods. With this study au-
thors add to the knowledge of the most efficient methods in shaping healthier
consumer behaviors and inspire for further research of text-based information
related to food.
Keywords: Food Computing, Kale, Sentiment-Token Bigram, NLP, Health,
Taste, Word Association, Food Blog, Social Media.
1 Introduction
The idea of a human as a rational wealth maximizer [1] can be best dissected when
looking at the mechanisms lying behind the food choice, which is a complex phenom-
enon impacted by factors such as context, culture, social class, emotions, mood et al.
but where the usually heavy-weight factor of health remains somewhat outside the
decision making. The rationality of a human can be questioned, considering the global
data on obesity and other long-term consequences of lifestyle-related illnesses [2].
Lifestyle-induced obesity, Type 2 diabetes and cardiovascular diseases are on the rise,
leaving public health policy makers with little success to report.
Meanwhile, the world has increasingly become a macroscopic sensor of our eating
habits and food cultures, evidenced via multiple photos, hashtags and documentations
of food and its background stories. “As food is central in our life, a significant frac-
tion of online content is about food” [3]. The evolution of homo sapiens from hunter-
gatherers to super-consumers [4] has led to a chicken-and-egg conundrum regarding
the main culprit – whether it is the consumer who gives preference to sweet and fatty
Copyright © 2021 for this paper by its authors. Use permitted under
Creative Commons License Attribution 4.0 International (CC BY 4.0).
52
fast food or whether we should put the blame on multinational corporations, i.e. food
industry actors that produce a plethora of increasingly sugary, salty and fat-intense
foods, widely advertising them on social media.
Food consumption is a complex phenomenon involving many aspects, identifiable
to a varying degree. When it comes to the consumer’s choice of food, the control over
decision-making is limited, leaving the food choice dependant on unpredictable out-
side triggers that appear in many formats, including those of food images and texts.
By analyzing texts on social media related to food, we can see that food choice can
serve as a marker of belonging to a certain social class, because “popular stereotypes
about who shops at farmer markets, makes their own kombucha, and drinks shade-
grown pour-over coffee reflect the reality that how we eat communicates important
information about our identities, including class background and status.” [5].
Through the social media text corpus, we can also look into the ways the food is
described and derive sentiments associated with it. At the same time, we bear in mind
that the field of food computing itself is a challenging area requiring a cross-sectoral
research approach where cognitive science and social science interlink with computer
science. This adds to the complexity of food computing choice analysis because the
notions prevalent in these science fields are often conflicting or at best not self-
explanatory. We will illustrate these challenges in our work where we analyze the
consumption of a particular food - kale, which is increasingly popular worldwide and
widely considered to be healthy.
2 Can Healthy Food Be Tasty?
In terms of healthy vs unhealthy food choices, the dichotomy between health and taste
is pronounced, as “many consumers still tend to over-consume energy-dense foods
because of two factors that work in combination: (1) foods that are “unhealthy” are
widely associated with being tasty” [6], and (2) taste is the main driver of food deci-
sions [7]. One of the key hindrances towards consuming healthy food is the un-
healthy=tasty intuition: “Previous attempts on the part of policy makers to induce
healthy food choices by educating consumers have not yet led to major improvements
or fundamental changes in actual behavior [..] the objectives of health and taste often
conflict - and [..] taste usually prevails in food decision making.” [8]
This intrinsically implies that the level of hedonism is higher in relation to un-
healthy foods, which are considered tastier than their healthy counterparts. Choosing
chocolate as one of the foods best depicting the level of hedonism, we can see that “a
‘reduced fat’ label, for example, negatively affects the acceptance of chocolate bars”
[8]. Meanwhile, descriptions of desserts and wine abound in sensuality. Taking into
account the central role food plays in human lives, “...it is no surprise the hashtag
#foodporn is one of the most popular hashtags on social media, especially on visual
media such as Instagram, often accompanied with a close-up of a tantalizing dish.”
[9]. In the world of #foodporn the arena of sensuality is dominated by cakes and
chocolate, and research reveals that “in the 1980s wine reviewers began to increase
their use of the body as a metaphor, starting to use words like fleshy, muscular, sin-
53
ewy, big-boned, or broad-shouldered.” [10] Not many of these connotations are found
in the healthy food discourse.
Imbuing healthy food with the desired portion of hedonism and presenting it as a
solution to unhealthy diets has its grounding in the multisensory field. “Research on
overeating assumes that pleasure must be sacrificed for the sake of good health. Con-
trary to this view, the authors show that focusing on sensory pleasure can make peo-
ple happier and willing to spend more for less food, a triple win for public health,
consumers, and companies alike” [11]. Thus, pleasure and hedonism matter, and there
is potential for embedding and highlighting those qualities in healthy food. Using text
as a proxy for such embeddings could be one of the solutions for enhancing more
healthy eating habits globally. Currently, however, we argue that the overall narrative
surrounding healthy food is not sufficiently engaging on an emotional level and lack-
ing in hedonistic connotations. As illustrated by Kāle and Agbozo [12], healthy food
is more associated with ease of preparation and seasonality while lacking any hedon-
istic attributes.
3 Cross-Sectoral Challenges of Discussing Food Choice
Determinants
There is a growing disagreement among the researchers across varying interlinked
research fields on what actually matters when consuming food. Much of the research
focus has lied on the juxtaposition between taste and healthiness since “the objectives
of health and taste often conflict - and (..) taste usually prevails in food decision mak-
ing.” [13]. Taste as the right combination of flavors has dominated almost all recipe-
related data analysis - “to test the “food pairing hypothesis,” a recent theory proposes
that tasty recipes are more likely to combine ingredients that share flavor com-
pounds.” [10]. Assumptions behind such food pairing research are that 1) taste is de-
termined by the chemical profile of the food, 2) taste is the most important determi-
nant in food choice. As the research group concludes: “Although many factors such as
colors, texture, temperature, and sound play an important role in food sensation, pal-
atability is largely determined by flavor, representing a group of sensations including
odors (due to molecules that can be bind olfactory receptors), tastes (due to molecules
that stimulate taste buds), and freshness or pungency (trigeminal senses). Therefore,
the flavor compound (chemical) profile of the culinary ingredients is a natural starting
point for a systematic search for principles that might underlie our choice of accepta-
ble ingredient combination.” [14].
Another stream of research, which is more qualitative and based on cognitive sci-
ence research methods rather than statistical analysis of big data, prioritizes all other
factors over taste in categorizing the food as tasty or not. “It is time to accept that
growing body of gastrophysics evidence demonstrating that the environment, not to
mention the plateware, dish-naming, cutlery and so on, all exert an influence over the
tasting experience.” [15]. It turns out that taste is a rudimentary attribute of the whole
food consumption process and is primarily related to the so-called food safety test,
which is the most important instinctive test humans perform when choosing to con-
54
sume food. “We mostly do not pay attention to what we taste. Rather, our brains just
do a quality check first, to ensure that there is nothing wrong with the food or drink
and that it tastes pretty much as we expected (or predicted) that it should” [15]. Here
the following assumptions are put into operation: 1) many other factors are more im-
portant than taste, 2) the perceived content of food prevails over the chemical content
(“perceived content” here means contextual factors - visual representation of the food
and its surrounding narrative).
The controversy between these theoretical standpoints and the assumptions they
operate on stem first and foremost from the challenge of cross-sectionality per se.
While the research on recipes or any available large food-related datasets focuses on
statistical analysis and views the existing data in terms of their analytical utility, the
experiment-based research focuses instead on such cognitive aspects that are not yet
well-documented in large-scale data. This discrepancy raises an important question -
to what extent can we pair theories and ideas from one field (in this case cognitive
science) with another field’s methodology (big data analytics, statistics)? As illustrat-
ed via data analysis of food micro blogging entries on Twitter, the notion of com-
plexity as a proxy for food’s hedonic attributes was scantly represented in healthy
food discourse [12]. It shows that food computing can serve, albeit to a limited degre-
e, as an intersection point between cognitive science and data analytics.
4 Healthy Food Choice: the Case of Kale
25 BC Apicius stated that “we eat first with our eyes” (Apicius, 25 BC) [4], which
could be reformulated into “we eat first with our mind”. Here we mean that the mod-
ern human is consuming texts and stories about food as eagerly as images of it. Re-
cent studies have indicated the extent to which textual information or storytelling
about food influences the multi-sensory experience of food intake. “In particular,
according to one analysis of online restaurant menu descriptions, the average price of
a dish in the US was found to go up by 6 cents for each additional letter in the dish
description” [16]. The increasing impact of words is also well documented in the
growing number of food blogging and food-related texts on social media. Taking into
account the growing popularity of food blogging, we base our analysis on the textual
information available about food, which allows us to make conclusions about the way
food is being depicted in those texts.
We aim to dig into stories/texts available on social media and appraise to what ex-
tent the quest for hedonism is reflected in the healthy food domain judging from the
sentiments described in the texts. The object of our focus here is kale, which has ex-
perienced a dramatic hike in its consumption in a relatively short period of time – e.g.
in Denmark, where the consumption of kale has increased by 900% over the last ten
years [17]: “When telling the story of food policy in action, it makes sense to start
with the humble kale. Kale is a real superfood: sustainable, relatively cheap, packed
with vitamins and possible to grow in conditions as low as −20°C. However, for dec-
ades kale consumption was in decline, abandoned by consumers. Until recently. (..)
This is just one of many areas where the Nordic countries seem to have turned a cor-
55
ner and begun changing food habits – to the benefit of public health, the environment
and the local economies” [18].
Tracing the triggers for kale consumption in the Nordic countries and globally is
beyond the ambition of this article. We rather aim to understand the sentiments be-
hind kale consumption by looking at Twitter food microblog entries related to kale,
trying to pinpoint the factors behind its popularity and tracing the expressions of he-
donism that could appear in texts related to the consumption of this particular food.
5 Methodology
We identified text mining as the best technique for discovering patterns in unfamiliar
data (Visa, 2001) – as illustrated in figure 2. We have also employed Latent Semantic
Analysis (LSA), as it is known for extracting and representing the meaning of words’
contextual use by using statistical computations applied to large corpora of text, as
well as due to its ability to illustrate the number of word (term) occurrences in input
documents [19] [20]. Data representing food blogs was collected using R program-
ming language and Twitter API with the following keywords: “kale” and “food”. A
total of N = 10,000 observations was gathered spanning tweets from January 2020 to
June 2020 (figure 1). Only the English language tweets were used for analysis. As
seen in figure 1, there is a decline in the use of “kale” and “food”, which we attribute
to the SARS-CoV-2/COVID-19 pandemic that hit the globe during the first half of
2020.
Fig. 1. Distribution of Dataset over Time.
The dataset on kale was then preprocessed by lemmatization (where words are trans-
formed into their original form – e.g. Eating is transformed to its lemma, Eat) and
normalization (where stopwords, numbers, whitespaces, unnecessary symbols such as
56
urls are removed). With duplicates deleted, we obtained N=9504 observations in total.
A cleaned dataset provided us with N= 9246 observations. The observations were
further tokenized into individual words/tokens followed by the part of speech (POS)
tagging in order to extract the adjectives required for analysis using the UDPipe Mod-
el [21]. SQL was used to query frequent tokens greater than two (2) for the purpose of
dimensionality reduction. Frequent bigrams with adjectives as the leading token fol-
lowed by a second token were then obtained from the dataset. From the tokenized
clean dataset a sentiment analysis of the adjective bigram was performed.
To evaluate the general emotions expressed per observation/Twitter post, the study
employed the Loughran-McDonald lexicon (Loughran and McDonald Sentiment
Word Lists), which was initially built for sentiment analysis in the financial context
[22]. The Loughran and McDonald Sentiment Word Lists were previously adopted in
the following studies:
- understanding sentiment disagreements with respect to bitcoin fluctuations [23],
- analyzing sentiments in financial microblogs [24],
- identifying moral hazards in restaurant hygiene inspections in New York [25],
The dictionary consists of tokens classified into “negative”, “positive”, “uncer-
tain”, “litigious”, “constraining”, “superfluous”, “interesting”, “modal words strong”
and “modal words weak”. The second semantic lexicon utilized was the AFFIN se-
mantic lexicon, which assigns a polarity score to each word [26] and thus consists of
words classified into positive or negative.
Fig. 2. Methodology and Research Workflow.
The following analytical techniques were employed:
a. discerning the most frequent bigrams of adjectives and tokens (i.e. adjective +
token)
b. discerning the most frequent bigrams of sentiments and tokens (i.e. sentiment +
token).
Choosing to focus on bigrams was justified by the advantage of bigrams over uni-
grams since bigrams are better suited for handling negated phrases [27]. Bigrams
contain more explicit information than words. They also have a relatively big seman-
tic granularity and can deal with unknown words [28]. Furthermore, bigrams provide
some understanding of the context since they have a better description capacity than
e.g. unigrams. [29]
57
Hereby we introduce a novel text analysis technique for representing co-
occurrence, which we term as Sentiment-Token Bigrams. Here, our algorithm match-
es the occurrence of a term from the AFFIN and Loughran-McDonald lexicons re-
spectively with the tokens in the ‘kale dataset’ (as the first component of the bigram).
This is followed by finding in the dataset (corpora) the subsequent token after the
sentiment-token. See the algorithm for Sentiment-Token Bigrams in Table 1.
Table 1. Algorithm for Sentiment-Token Bigram.
Algorithm: Sentiment-Token Bigram
Data: Twitter Posts Dataset (N observations)
Sentiment Lexicon: Dictionary of terms associated with sentiments
Result: Bigram with a Sentiment Lexicon Term as the first word
load SentimentLexicon
for each observation do
extract LexiconTerm and SubsequentToken
repeat until no observations left
end
count bigrams
delete bigrams with low frequency (user-defined threshold)
visualize top Sentiment-Token Bigram
6 Data Analysis
The Twitter microblog analysis of sentiments regarding kale shows the following
results:
- Bing sentiment analysis reveals that ‘anti-inflammatory’ (30), ‘benefit kale’
(15), ‘amaze food’ (15) and ‘amaze kale’ (12) are the most frequent bigrams of
the kale-related dataset
- Loughran-McDonald sentiment analysis yields ‘boost immune’ (30), ‘boost
food’ (25), ‘benefit kale’ (15) and ‘boost immunity’ (8).
Both sentiment graphs offer a wide range of bigrams containing the words ‘beauti-
ful’, ‘benefit’ and ‘awesome’ among other most frequent bigrams, and both of them
abound in verbs – e.g. ‘to benefit’ and ‘to boost’ – promising impressive results that
kale can deliver to one’s health.
58
Fig. 3. Most Frequent Bigrams (Bing Sentiment + Token).
Fig. 4. Most Frequent Bigrams (Loughran-McDonald Sentiment + Token).
Analyzing the most frequent bigrams with adjectives, we see that ‘sweet potato’ (249)
and ‘healthy food’ (245) dominate the discourse, implying that kale might only serve
as an ‘anti-inflammatory’ and ‘immunity boosting’ healthy addition to other foods.
59
If we look at the bigrams determining composition or the context of other foods
kale is consumed with, we can notice certain food-related value-based statements,
such as ‘fast food’ (136), ‘comfo food’ (116) and ‘vegan food’ (114). None of these
statements reflect any hedonistic attitudes associated with consuming this particular
food.
We also see surprisingly little reference to tasty food. ‘Favorite food’ (101), ‘fresh
food’ (53) and ‘delicious food’ (49) could be related to taste aspects, but those are not
prevalent.
All in all, bigrams provide a meta-story of kale. By looking at the sentiment analy-
sis and the most frequent bigrams of adjectives and nouns, we are able to follow the
most frequently told ‘stories’ about kale on a more detailed, micro-level as well.
Fig. 5. Most Frequent Bigrams (Adjective + Token).
In conclusion, this analysis reveals that kale offers clear-cut and rather ambitious
promises for the immune system, thereby turning the focus on health. Hedonism and
tastiness are left out of the discourse, putting kale into the healthy food domain where
food consumption is governed by functional rather than sensual benefits.
7 Conclusions
With this research, we aimed to define some of the methods useful for facilitating the
consumption of healthy foods. Having made an analysis of texts describing food, we
have ascertained the lack of hedonism-denoting terms in healthy food descriptions.
We have also tried to distinguish the triggers behind preferences for healthy foods as
60
depicted in the texts of microblog entries available on Twitter. For the analysis unit
we chose kale - a food whose consumption over the last decade has grown significant-
ly.
Kāle and Agbozo [12] concluded that contemporary food blog entries on Twitter
related to healthy food do not focus on taste aspects and contain few references to
hedonic expressions, accentuating rather “simple and easy” qualities and to a lesser,
or barely discernible, extent - “complex and enjoyable”. The same holds true for the
case study analysis of kale, although the bigrams related to kale do feature more con-
crete promises for health improvement. Still, the notions of hedonism and tastiness
are left outside the sentiments associated with consuming the healthy food kale.
Though it may not have adversely influenced the result of the analysis, a minor
limitation of the study was the lack of a wider scope of dataset from previous years
(i.e. before 2020).
The study’s theoretical contribution is evident in the introduction of a novel Senti-
ment-Token Bigram algorithm which could be applied in the computational linguis-
tics.
Lastly, this research also contributes to the discussion of the extent to which con-
cepts that are used in cognitive science domain are operational in the quantitative
analysis of texts. Due to the challenge of cross-sectionality, this research has aimed to
refine the questions that could be posed by the research community with regard to
food and human relationships. Potentially, instead of asking ‘how tasty the food is?’
we should rather ask: ‘how interesting or entertaining the food is?’ or ‘to what extent
food consumption is related to social status?’ Instead of narrowing down the focus on
taste, as has been customary in computer science to date, the focus should be broad-
ened and encompass e.g. food’s associations with entertainment, sensuality and he-
donism, as well as its ability to elevate one’s social status or manifest other important
identity markers. The growing research of food computing which raises cross-
sectionally relevant questions brings us ever closer to understanding the phenomenon
of the food choice.
References
1. Amadae, S.M.: Rational choice theory. Encyclopaedia Britannica:
https://www.britannica.com/topic/rational-choice-theory, last accessed 2020/01/02.
2. Min, W., Jiang, S., Jain, R.C.: Food Recommendation: Framework, Existing Solutions and
Challenges. IEEE Transactions on Multimedia (2019).
3. Min, W., Jiang, S., Liu, L., Rui, Y., Jain, R.: A survey on food computing. ACM Compu-
ting Surveys (CSUR) 52 (5), 1–36 (2019).
4. Spence, C., Okajima, K., Cheok, A. D., Petit, O., Michel, C.: Eating with our eyes: From
visual hunger to digital satiation. Brain Cogn. 110, 53-63 (2016).
doi:10.1016/j.bandc.2015.08.006
5. Finn, S.: Can “Taste” Be Separated from Social Class? In: Ludington, C. & Booker, M.
(Eds.): Food Fights: How History Matters to Contemporary Food Debates, pp. 81-99.
Chapel Hill: University of North Carolina Press (2019).
61
6. Raghunathan, R., Walker Naylor, R., Hoyer, W.D.: The unhealthy= tasty intuition and its
effects on taste inferences, enjoyment, and choice of food products. Journal of Marketing
70 (4), 170–184 (2006).
7. Tepper, B. J., Trail, A.C.: Taste or health: A study on consumer acceptance of corn chips.
Food Quality and Preference 9 (4), 267–272 (1998).
8. Mai, R., Hoffman. S., Helmert, R. J., Velichkovsky, M. B., Zahn, S., Jaros, D., Schwarz
E.H.P., Rohm, H.: Implicit food associations as obstacles to healthy nutrition: the need for
further research. Br. J. Diabetes Vasc. Dis., vol. 11 (4), Jul/Aug 2011, 182-186 (2011).
doi:10.1177/1474651411410725
9. Mejova, Y., Abbar, S., Haddadi, H.: Fetishizing food in digital age: #foodporn around the
world. In: Tenth International AAAI Conference on Web and Social Media (2016).
10. Jurafsky, D.: The language of food: A linguist reads the menu. WW Norton & Company
(2014).
11. Cornil, Y., Chandon, P.: Pleasure as a substitute for size: how multisensory imagery can
make people happier with smaller food portions. J. Mark. Res. Vol. LIII, 847-864 (2016).
doi:10.1509/jmr.14.0299
12. Kāle, M., Agbozo, E.: Tracing complexity in food blogging entries. In: CEUR Workshop
Proceedings 2612, pp. 51–62 (2020).
13. Mai, R., Hoffmann, S.: How to combat the unhealthy= tasty intuition: The influencing role
of health consciousness. Journal of Public Policy & Marketing 34 (1), 63–83 (2015).
14. Ahn, Y., Ahnert, S.E., Bagrow, J.P., Barabási, A-L.: Flavor network and the principles of
food pairing. Scientific reports 1, 196, (2011).
15. Spence, C.: Gastrophysics: The new science of eating. Penguin UK (2017).
16. Spence, C.: Complexity on the menu and in the meal. Foods 7, 158 (2018).
doi:10.3390/foods7100158
17. Nordic Food Policy Lab. Collecting, curating and co-producing policy tools for a consum-
er powered food transition, https://www.wur.nl/upload_mm/1/9/d/6f259c39-ddbb-453d-
9f46-290a7b721999_OECD%20-%20WUR%20-%20050919%20-%20PerssonM.PDF,
last accessed 2020/07/15.
18. Halloran, A., Fischer-Møller, M.F., Persson, M., Skylare, E.: Solutions Menu - A Nordic
guide to sustainable food policy. Nordic Council of Ministers (2018).
19. Wei, C. P., Yang, C. C., & Lin, C. M.: A latent semantic indexing-based approach to mul-
tilingual document clustering. Decision Support Systems, 45(3), 606-620, (2008).
20. Miles, I., Saritas, O., & Sokolov, A.: Intelligence: Environmental and Horizon Scanning.
Foresight for Science, Technology and Innovation, pp. 63-93. Springer, Cham. (2016).
21. Straka, M. 2018. UDPipe 2.0 prototype at CoNLL 2018 UD shared task. In: Proceedings
of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal De-
pendencies, 197–207, (2018).
22. Loughran, T., McDonald, B.: When is a liability not a liability? Textual analysis, dictionar-
ies, and 10-Ks. The Journal of Finance 66 (1), 35–65 (2011).
23. Ahn, Y., Kim, D.: Sentiment disagreement and bitcoin price fluctuations: a psycholinguis-
tic approach. Applied Economics Letters, 27(5), 412–416 (2020).
24. Cortis, K., Freitas, A., Daudert, T., Huerlimann, M., Zarrouk, M., Handschuh, S., Davis,
B.: Semeval-2017 task 5: Fine-grained sentiment analysis on financial microblogs and
news. Association for Computational Linguistics (ACL). (2017)
25. Mejia, J., Mankad, S., Gopal, A.: A for Effort? Using the Crowd to Identify Moral Hazard
in NYC Restaurant Hygiene Inspections (Jan 1, 2018). Kelley School of Business Re-
search Paper, 18-58 (2018).
62
26. AArup Nielsen, F.: A new ANEW: Evaluation of a word list for sentiment analysis in mi-
croblogs. arXiv preprint arXiv:1103.2903 (2011).
27. Chen, J.: Search engine For Twitter sentiment analysis (Doctoral dissertation). (2015).
28. Li, J., Sun, M.: Non-Independent term selection for Chinese text categorization. Tsinghua
Science and Technology, 14(1), 113-120 (2009).
29. Kaptein, R., Hiemstra, D., Kamps, J.: How different are language models and word
clouds? In: European conference on information retrieval, pp. 556-568. Springer, Berlin,
Heidelberg (2010).