Introduction: How do non-Western recipes get whitewashed?¶

Online recipe databases and food media websites such as New York Times Cooking and Bon Appetit have helped make cooking from a wide variety of cuisines an increasingly accessible activity. However, food media remains overwhelmingly white--for example, within the New York Times Cooking recipe archive, even non-Western recipes are written by a majority white group of authors.^[1]

At the same time, the positioning of white authors as authority figures on non-white cuisines has increasingly been challenged, with critics pointing out the whitewashing and related harms that occur. In an interview^[2] with Yewande Komolafe, a Nigerian recipe developer and author, fellow recipe developer Priya Krishna discusses some of the frustrations they have experienced as BIPOC members of food media (emphasis mine):

We lamented the extra labor that non-white people are often asked to do—from including **additional explanations** of dishes in the head note to finding imperfect **substitutions** for ingredients—in service of making their recipes more accessible to a white audience.

In addition to the implied white audience that they must consider when adding explanations that would be redundant for non-white audiences, Krishna also observes how non-Western recipes are often framed in ways to make them appear more accessible:

It's why modifiers like “simple” or “weeknight” that many publications (including BA) have historically added always make me laugh. Indian is my weeknight food. There's no weeknight dal and then dal. It's all just dal. **It’s as if our food needs to be made more approachable**.

I highly recommend reading the full interview, as there are many other observations that I think you'll find relatable. Meanwhile, the interview also made me wonder what empirical evidence I could find for the whitewashing of recipes, and more generally, what biases may be present in how food media talks about Western versus non-Western recipes.

Data¶

Overview¶

To empirically test some of Komolafe's and Krishna's hypotheses (such as non-Western recipes attracting words that signal "approachability") and to identify large-scale patterns in how food media's treatment of Western and non-Western recipes differ, I assembled³ a dataset of 53K recipes from New York Times Cooking (NYT), Bon Appetit (BA), Food52 (F52), Epicurious (EPI), and Serious Eats (SE).⁴ I chose these five websites for the simple reason that they are the ones I visit most frequently. For the NYT recipes, I additionally scraped the text of the longer articles that sometimes accompany recipes, in which the author reflects on their inspirations and process for creating the recipe (see the highlighted "Featured in" box in the screenshot below for an example).

from IPython.display import Image
Image("./figures/nyt_ex.png")

Let's plot the distribution of recipes over websites (using our food-themed color "palate": "tomato", "orange", "lime", "wheat", "plum", "honeydew", "chocolate", "olive", "bisque", and "salmon"). Apart from Serious Eats, the distribution of recipes is relatively even across the other 4 websites:

Image(filename="./figures/gen_dist.png", width = 500)

Getting cuisine labels¶

Using existing cuisine categories¶

For the analysis, I need to assign each recipe to the label of either non-Western (nw) or Western (w). Since the 5 websites in our dataset provide cuisine categories such as "Chinese" and "Italian" for the majority of recipes, I created a mapping from each of these cuisine categories to either nw or w to get a label for these recipes. I mapped cuisines onto either w or nw based on the simple heuristic that w is associated with the continents and countries of Europe, North America, and Australia, whereas nw is associated with the continents/countries of Asia, Africa, and South America.

In total, there are 103 unique cuisine categories in use across the 5 websites, shown below. A lot can be said about the idiosyncrasies of the cuisine classification that these websites show, such as the varying granularities (e.g., asian vs. californian), overlap (e.g., asian, japanese), and presence of fusion (e.g., tex-mex).

Cuisine	Category
African	non_west
Afro-brazilian	non_west
American	fusion
Argentine	non_west
Ashkenazi	west
Asian	non_west
Australian	west
Australian-new-zealander	west
Austrian	west
Bangladeshi	non_west
Basque	west
Belgian	west
Brazilian	non_west
British	west
Cajun	fusion
Cajun-creole	non_west
Californian	fusion
Canadian	west
Cantonese	non_west
Caribbean	fusion
Central-american	non_west
Central-asian	non_west
Central-south-american	non_west
Chilean	non_west
Chinese	non_west
Colombian	non_west
Creole	fusion
Cuban	fusion
Czech	west
Danish	west
Dutch	west
East-african	non_west
Eastern-european	west
Eastern-european-russian	west
Ecuadorian	non_west
Egyptian	non_west
English	west
Ethiopian	non_west
European	west
Filipino	non_west
Finnish	west
French	west
French-provencal	west
German	west
Greek	west
Hungarian	west
Icelandic	west
Indian	non_west
Indonesian	non_west
Iranian	non_west
Irish	west
Israeli	non_west
Italian	west
Italian-american	west
Jamaican	non_west
Japanese	non_west
Jewish	fusion
Korean	non_west
Latin-american	non_west
Lebanese	non_west
Malaysian	non_west
Mediterranean	fusion
Mexican	non_west
Middle-eastern	non_west
Midwestern	fusion
Moroccan	non_west
Moroccan-north-african	non_west
New-england	fusion
Nigerian	non_west
North-african	non_west
Northern-italian	west
Norwegian	west
Nuevo-latino	non_west
Pacific-northwest	fusion
Pakistani	non_west
Persian	non_west
Peruvian	non_west
Polish	west
Portuguese	west
Provencal	west
Puerto-rican	non_west
Quebec	west
Russian	west
Scandinavian	west
Scottish	west
Sephardic	west
South-african	non_west
South-american	non_west
South-asian	non_west
Southeast-asian	non_west
Southern	fusion
Southern-italian	west
Southwestern	fusion
Spanish	west
Spanish-portuguese	west
Swedish	west
Swiss	west
Szechuan	non_west
Taiwanese	non_west
Tex-mex	fusion
Thai	non_west
Tibetan	non_west
Turkish	non_west
Tuscan	west
Vietnamese	non_west
Welsh	west
West-african	non_west

For the purposes of this project, I was concerned only with the macro-categories of "Western" (w) and "non-Western" (nw), which are of course slippery in their boundaries as well, often due to the presence of cross-cultural exchange, immigration, and other sociocultural phenomena resulting in what is broadly termed "fusion". To further simplify matters, I excluded recipes belonging to categories that are inherently at the intersection of w and nw, such as Tex-Mex and Mediterranean.

Though the majority of recipes are categorized as exactly one cuisine, about $1/3$ are missing an associated cuisine category (all of the BA recipes fall into this bin), and a handful of recipes are categorized as multiple cuisines. Excluding one outlier recipe ("homemade adobo seasoning mix & dry rub", which is associated with 42 different cuisines according to F52), the number of associated cuisines for recipes in our dataset are shown below.

Image(filename="./figures/num_cuisine_cats_dist.png",width=700)

It would be a shame to discard the third of our data that is missing a cuisine label, so I trained a classifier to label this unlabeled data for the analysis.

For the recipes with too many cuisine labels, given that the cuisine categories occur at multiple granularities, I first performed some manual inspection to see whether some categories can be collapsed. A heatmap showing the overlap percentages (calculated as $\frac{N_{c1,c2}}{N_{c1}}$, where $N_{c1,c2}$ is the number of recipes tagged as both $c1$ and $c2$, and $N_{c1}$ is the number of recipes tagged as $c1$) is shown below.

Image(filename="./figures/cuisine_overlap.png",width=1000)

Overall, the overlaps make intuitive sense: for example, there is a moderate degree of overlap between is_jamaican and is_caribbean. At the same time, it seems that most recipes do not get tagged as all cuisines that are applicable, i.e., there is imperfect overlap between cuisines and their hypernyms (like is_tuscan and is_italian) as one would expect.

To find potentially collapsible cuisine categories, I applied a threshold of $75\%$: if $75\%$ of recipes categorized as $c1$ are also categorized as $c2$, then I used a single label, $c2$, to label all recipes categorized as either $c1$ or $c2$. This scheme allows me to avoid counting a cuisine from a single macro-category twice, and resulted in the merging of the following categories:

Central-American, Latin-American, South-American
Northern-Italian and Italian

The resulting distribution of recipes over different cuisines is highly skewed, with recipes categorized as Italian, American, and French making up roughly 40% of all categorized recipes, as shown below:

Image(filename="./figures/dist_20_freq_cuisines.png",width=800)

Now that Ie have collapsed cuisine categories where possible, for the remaining recipes with multiple cuisine labels, I'll use the following labeling scheme:

Case 1: majority vote with the threshold at $0.67$ (e.g., a recipe with $2/3$ or more non-Western associated cuisines would receive the label nw)
Case 2: tie or no two-thirds majority--assign the label f for fusion, which I will exclude from later analysis

The resulting distribution over macro cuisine category labels is shown below:

Image(filename="./figures/dist_macro_cuisine_cats.png",width=500)

Comparing cuisine category-derived labels with demonym-derived labels¶

As a sanity check, let's compare our macro-category labels to explicit cues in the form of demonyms^[5] that are present in recipe titles, e.g. "Moroccan Lamb Stew". About $\frac{2281}{53946}=4.2\%$ of recipes have a demonym in the title. I used a manually created mapping of demonyms onto the macro-categories of w and nw, as follows:

Demonym	Category
Afghan	non_west
Africa	non_west
Alaskan	west
Algerian	non_west
America	west
American	west
Andalusian	west
Argentine	non_west
Armenian	non_west
Asia	non_west
Australian	west
Austrian	west
Baja	non_west
Bangladeshi	non_west
Bavarian	west
Belgian	west
Bengal	non_west
Bengali	non_west
Bohemian	west
Bosnian	west
Brazilian	non_west
Breton	west
Bulgarian	west
Canadian	west
Cantabrian	west
Chilean	non_west
Colombian	non_west
Corsican	west
Cretan	west
Cuban	non_west
Czech	west
Dalmatian	west
Dominican	non_west
Ecuadorian	non_west
Egyptian	non_west
Ethiopian	non_west
Galician	west
Genovese	west
Georgian	west
German	west
Ghanaian	non_west
Guatemalan	non_west
Haitian	non_west
Hungarian	west
Hyderabad	non_west
Indian	non_west
Indonesian	non_west
Iranian	non_west
Israel	non_west
Italian	west
Jamaican	non_west
Japanese	non_west
Javanese	non_west
Jordanian	non_west
Laotian	non_west
Liberian	non_west
Libyan	non_west
Ligurian	west
Lithuanian	west
Macedonian	west
Maharashtrian	non_west
Majorcan	west
Malagasy	non_west
Malaysian	non_west
Mexican	non_west
Moroccan	non_west
Nicaraguan	non_west
Nigerian	non_west
Norman	west
Parisian	west
Persian	non_west
Peruvian	non_west
Phoenician	fusion
Pole	west
Qatar	non_west
Roman	west
Russian	west
Salvadoran	non_west
Sardinian	west
Serbian	west
Sicilian	west
Silesian	west
Singaporean	non_west
Slovak	west
Somali	non_west
Syrian	non_west
Tahitian	non_west
Texan	west
Thai	non_west
Tunisian	non_west
Ukrainian	west
Umbrian	west
Uzbek	non_west
Uzbekistani	non_west
Valencian	west
Venezuelan	non_west
Victorian	west

Of those recipes with a demonym, the majority are nw demonyms:

Image("./figures/dist_macro_demonym_cats.png",width=500)

As we would expect, only a very small percentage of recipes with demonyms are categorized in the opposite category to their demonym: $1.8\%$ for recipes with non-western demonyms; $1.6\%$ for recipes with western demonyms.

Image("./figures/dem_cuisine_cat_confusion_matrix.png",width=600)

Some examples of disagreement between demonyms in a recipe and a website's cuisine categorization for a recipe include the following:

Recipe title	Demonym-based label	Cuisine-category-based label
Hand Held Syrian Spinach Pies	`nw`	`w`
Roasted Japanese Eggplant With Crushed Tomato, Pecorino and Thyme	`nw`	`w`
Vincent Hodgins's Moroccan Spiced Gravlax	`nw`	`w`
Fried Rice with Canadian Bacon	`w`	`nw`
An American Vegetable Soup	`w`	`nw`

As these examples of disagreement show, there are genuine reasons for considering such recipes as both nw and w. However, for better consistency, I ignored the signal provided by the demonym and continued to use solely the cuisine-categories to label these recipes.

Finally, there are 267 recipes that are missing a cuisine category but which do have a demonym, so in this case, we may as well label these recipes based on the macro-category of their demonyms. The final distribution of recipes over macro-categories is shown below:

Image("./figures/dist_macro_cuisine_cats_final.png",width=600)

I saved these labels for later use, including training the classifier.

Classifier¶

Training and eval data¶

I experimented with several sources of training data for a logistic regression classifier:

Name	Description	Size
`recipes_labeled`	a random selection of $2/3$ the labeled data from `recipes` (labels from the `macro_cuisine_cat` column)	$16018$
`wiki`	data from the Wikipedia lists https://en.wikipedia.org/wiki/European_cuisine and https://en.wikipedia.org/wiki/Asian_cuisine to correspond to `w` and `nw` dishes, respectively^[6]	$8293$
`recipes_labeled_nouns`	same as `recipes_labeled`, keeping just the nouns from the recipe names (after POS-tagging with SpaCy)	$16018$
`wiki_nouns`	same as above, but with `wiki` as the base dataset	$8293$

I also created several different eval/dev sets:

Name	Description	Size
`recipes_labeled`	the remaining $1/3$ split of `recipes_labeled`	$7783$
`demonyms`	a subset of recipes with demonyms in the titles	$235$
`manual`	a manually-labeled subset of `recipes_labeled` with recipes from BA	$169$
`<X>_nouns`	each of the 3 datasets above, restricted to noun lemmas only

Some notes on the data splits:

Sizes shown are after excluding recipes with a fusion label.
I did not include wiki data in any eval or dev sets because the ultimate goal is to label non-Wikipedia recipes.
Observations on wiki data:
- There is some overlap ($N=113$) in w/nw dishes due to some common food terms as well as categories spanning Western and non-Western cuisine.
- e.g.: "beer", "caviar", "fruits", "ice cream", "kasha", "noodle soups with meat"
- I included the overlapping data in the training set (so, e.g., "ice cream" occurs twice, once with each label).
Since BA does not have any cuisine labels, I created labels for the eval/dev sets via manual annotation. For this reason, there is also no BA data in the training set.
I set all random seeds to 42.

And I applied the following pre-processing steps:

remove punctuation and non-alpha characters (keeping diacritics)
lowercase
deduplicate the training split of recipes_labeled from anything contained in manual and demonyms

Training a logistic regression classifier¶

Next, I trained logistic regression (LR) classifiers using the scikit-learn library, experimenting with all the training datasets and all possible combinations of word- and character-level features, with n-gram ranges of 1- to 4-grams for word-level features. I got the best cross-validated dev accuracy (across all dev sets) using recipes_labeled as training data and word-level features up to trigrams in combination with char-level features.

I then did a grid-search over the following hyperparameters, again with recipes_labeled as the eval and training sets:

Hyperparameter	Values	Best value
`solver`	`['newton-cg', 'lbfgs', 'liblinear']`	`'lbfgs'`
`c_value`	`[100, 10, 1.0, 0.1, 0.01]`	`10`
`max_iter`	`[100, 200, 500, 1000]`	`100`
`random_state`	`[1639, 18024, 16049, 14628, 9144]` (5 randomly generated seeds)	`1639`

A more detailed performance of the best-performing classifier is shown below (evaluated on the test split of recipes_labeled):

	precision	recall	f1-score	support
non_west	0.92	0.88	0.90	1958
west	0.92	0.94	0.93	2712

accuracy			0.92	4670
macro avg	0.92	0.91	0.92	4670
weighted avg	0.92	0.92	0.92	4670

Moreover, the classifier does pretty well on the manually-labeled test set, especially considering that this eval set comes entirely from BA-data not represented in the training data. Although performance on nw recipes is worse, the overall accuracy is still better than a majority-class baseline ($\frac{140}{169}=82.8\%$).

	precision	recall	f1-score	support
non_west	0.53	0.90	0.67	29
west	0.97	0.84	0.90	140

accuracy			0.85	169
macro avg	0.75	0.87	0.78	169
weighted avg	0.90	0.85	0.86	169

Qualitative error analysis¶

Next, I examined the incorrect predictions ($N=384$) and noticed some recurring misclassification types that largely make sense.

Error type	Examples	Justification
fusion recipes	"miso pork bread pudding", "shichimi togarashi granola", "super easy naan pizza "	there are conflicting signals, e.g. "miso", "shichimi", "naan" are `nw` but "pudding", "granola", and "pizza" are `w`
unigram recipes	"ayran", "cholay"	these recipes are not represented in the training data and contain no signal in the form of adjectives or ingredients for the classifier to use
vague recipes	"chicken stew", "roast pork", "spring soup"	these recipes are inherently impossible to classify without more context

We can also examine the feature weights to see which word- and char-features are most associated with each class. The top 15 features from each class are shown below.

Image("./figures/15_top_features.png",width=800)

For the w label, we notice many demonyms and words associated with Spanish and Italian cuisine ("paella", "bruschetta"). This is unsurprising, given the frequency of Spanish and Italian recipes in the training data. For the nw label, we also notice many demonyms as well as some common dishes associated with the top-weighted demonyms ("tacos", "tamale", "chutney").

Considering the intuitive errors and feature weights, I was pretty happy with the classifier's performance, and proceeded to the analysis with the fully-labeled dataset of recipes.

Analysis¶

I first applied the classifier to the unlabeled portion of recipes in order to get a prediction for macro_cuisine_cat for all datapoints. I conducted analysis on the fully labeled recipes dataset (combining existing website cuisine-category- and demonym-based labels and model predictions), as well as the a priori labeled subset of recipes (i.e., excluding model predictions), finding that similar patterns hold. All subsequent plots and figures are thus based on the full dataset.

Next, I used SpaCy (https://spacy.io/) to tokenize, lemmatize, part-of-speech tag, and dependency parse all text data (recipe title, ingredients, directions, free text, and full text of the associated "featured in" article in the case of some NYT recipes).

We can now do a few different kinds of analysis:

General: log odds ratios to determine which words are most associated with w vs. nw recipes
Approachability analysis: Do nw recipes get described more often as "fast", "easy" etc. to make them more approachable, as Yewande and Priya hypothesized?

General words¶

Using the results from spaCy pre-processing, I calculated the log-odds-ratio (LOR) for each word occurring with a Western vs. a Non-Western recipe using an informative Dirichlet prior (following Monroe et al. (2009)^[7]).

Image("./figures/gen_deltas.png",width=800)

Even after filtering to words with a significant LOR value, there are many which reflect inherent differences in terms of ingredients, cooking techniques, and proper nouns that exist between W and NW cuisine (for example, "ghee", "wok", and "indian" are unsurprisingly more associated with NW cuisine, and "pasta", "italian", and "bake" are more associated with W cuisine).

Thus, I used WordNet (https://wordnet.princeton.edu/) to compile a set of ingredients (hyponyms of the synsets ingredient.n.03, food.n.01, food.n.02, carbohydrate.n.01, edible_fat.n.01, edible_seed.n.01, herb.n.01, herb.n.02), cooking-related terms (hyponyms of kitchen_utensil.n.01, tableware.n.01), and proper nouns (hyponyms and member meronyms of state.n.04, country.n.02) to additionally filter out. The resulting log odds ratio plot for words that remain is shown below.

Image("./figures/gen_deltas_post_filtering.png",width=800)

Now that we have identified the words that are most associated with each cuisine type beyond these inherent cuisine characteristics, **what patterns do we see emerge among each cuisine type's words?**

To answer this question, I clustered the words into semantically coherent groups by retrieving word embeddings (GoogleNews, 300 dimensions) for each word, then applying KMeans clustering (I chose num_clusters=7 after manual inspection). The visualized results are shown below for each cuisine type:

Image("./figures/NW_clusters.png")

Image("./figures/W_clusters.png")

From manual inspection of the words contained in each cluster, I came up with over-arching semantic descriptions for each cuisine type's clusters. For both Western and non-Western recipes, POS-tag-related clusters emerge: both cuisine types have separate clusters for (what are predominantly) verbs and adjectives. Furthermore, both cuisine types have a cluster for proper nouns, as well as a cluster for common nouns that are unrelated to any of the other noun-clusters.

However, within similar clusters (e.g., adjectives, verbs, common nouns), we observe many contrasts between the two cuisine types, and we also observe that there are clusters unique to one cuisine type and not the other.

For example, one of the clusters unique to non-Western recipes consists of words relating to culture, family, and ethnic identity, e.g.: "mom", "family", "ancestor", "community", "paternal", "bride", "language", "indigenous", "colonial", "immigrant", "diaspora", "homesickness", "nationality". Many of the words in this cluster are suggestive of "immigrant story"-type narratives, suggesting that recipe writers tend to couch nw foods in terms of their ethnic symbolism.

On the other hand, Western recipes words have a cluster relating to terms indexing luxury ("gastronomic", "hotel", "butler", "gourmet") and rusticity ("countryside", "village", "shepherd", "farmhouse", "harvest", "forest", "hearth", "hunter", "peasant", "rustic"), and terms residing in the overlap ("manor", "heirloom") (it is perhaps also no coincidence that rusticity and the use of local, seasonal ingredients currently have significant cachet in fine dining).

Never mind that Western foods are also ethnic artifacts and intertwined in Western ethnic identities, or that non-Western recipes are also cooked by peasants and shepherds living in farmhouses and forests in their rustic countryside villages--it appears that for American food media, NW foods tend to be symbols of their cultures and the immigration that brings their cultures into relief, whereas W foods tend to be symbols of an upscale, farm-to-table aesthetic.

A similar bias is reflected in the Adjective clusters for each cuisine type: wherease Western recipes attract adjectives indexing fine dining such as "seductive", "stylistic", "grownup", "impressive", "timeless", "extravagant", "lavish", "fussy", "sexy", "sublime", "fashionable", "sleek", "fine", "fancy", "classic", and "elegant", non-Western recipes attract adjectives like "cheap", "affordable", and "everyday".

Also interesting is the greater presence of adjectives and lemmas involved in comparison ("different", "unfamiliar", "reminiscent", "exotic") within non-Western recipes--this suggests that Western recipes tend to be regarded as the default, against which recipe writers compare non-Western foods, ingredients, techniques etc.

Finally, it is interesting to note that adjectives relating to taste and smell (especially some that carry negative connotations) are more likely to be found in non-Western recipes (see the 'Foods and flavors' cluster, and some entries in the main Adjectives cluster, e.g. "bold", "funky"). Of course, this bias might be partially explained by the use of many flavorful ingredients and spices in non-Western cuisine that (stereotypically) blander Western cuisine lacks, but also by the subjective perception of non-Western cuisine as less bland.

NW clusters¶

Proper nouns	Common nouns	Specialized food terms	Culture, family, nationality	Foods and flavors	Adjectives	Verbs
dal	steam	wrapper	southeast	spicy	fiery	skewer
char	heat	oriental	lunar	grill	thinly	shred
turkish	pressure	kernel	grocery	marinate	shredded	ferment
shanghai	dipping	charcoal	coal	masa	addictive	soak
latin	mortar	fresco	supermarket	aromatic	exotic	stir
mam	market	matchstick	mom	glutinous	grind	sate
monterey	many	seaweed	cultural	crispy	refreshing	turn
msg	lid	rinse	southwestern	cashews	authentic	condense
tet	traditional	devein	imperial	smoky	bright	combine
taiwanese	available	asafetida	comedian	flavor	favorite	blacken
bangkok	part	pod	indigenous	vegetarian	dim	splutter
lee	healthy	fermented	family	garnish	easy	burn
taiwan	through	thigh	eastern	flavorful	wisp	squeeze
philippine	pad	pestle	westerner	boneless	staple	find
newman	smoke	lotus	waterman	tandoor	comforting	adjust
yuan	own	asafoetida	colonial	salty	numbing	sit
lakshmi	national	cobs	polo	shiitake	milder	darken
sin	fire	clay	canton	cob	superman	distil
beijing	skirt	cleaver	monsoon	delicious	bold	sputter
anaheim	different	slurry	pasha	microwave	silken	shun
amarillo	distinct	cashew	maternal	eat	reminiscent	enjoy
ras	unique	cellophane	shark	tangy	colorful	inspire
caribbean	glove	agave	immigrant	pungent	amazing	replenish
seoul	section	barberry	religious	cook	funky	ram
bombay	super	shimmer	grocer	tasty	eyed	subscribe
chi	smoking	bonnet	dynasty	nutritious	incendiary	prepare
muslim	batch	cactus	pantry	cuisine	awesome	zap
tokyo	sum	finely	communal	spiciness	fantastic	serve
fresno	street	blister	diaspora	slurp	satisfying	add
chen	sticker	drumstick	pastor	skinless	script	subdue
lexington	online	floret	professor	cooked	steamy	vaporize
agua	smash	tilapia	tiger	vegan	tedious	flee
delhi	less	ml	editor	broil	ubiquitous	impart
mumbai	prep	fungus	canal	sizzle	familiar	grow
rajas	helpful	stalk	peninsula	steaming	fusion	cater
pacific	aside	perilla	household	sodium	picky	omit
tam	patrón	massage	subcontinent	grilling	mellow	chronicle
istanbul	everyday	nigella	trader	sear	subtle	submerge
xian	spam	cassia	refugee	basting	unseasoned	toss
zhou	vendor	fragrance	cub	vinegary	surprising	mop
chang	influence	lacquer	palace	dhal	vibrant	meld
han	contact	bruise	pup	gingery	palatable	wipe
hass	flushing	protein	lam	momos	squishy	discard
latino	carousel	packet	hawker	accompaniment	bumpy	customize
arabic	scramble	pleat	ancestor	rotisserie	shrimpy	inoculate
lan	trading	valve	coastal	pomfret	drunken	take
krishna	similar	filé	minister	yummy	unfamiliar	buy
info	leftover	tan	queen	unsweetened	snappy	build
hindi	indirect	hatch	paternal	sambar	auspicious	encapsulate
hindu	brand	briquette	archipelago	seared	chunky	explode
mr	lean	fin	unity	marinating	fleshed	deteriorate
diaz	once	soman	bride	sourness	moody	emit
ron	various	musky	prosperity	grilled	lech	introduce
li	commonly	aguas	uncle	flavourful	fun	oversimplify
naomi	able	bead	political	aroma	unrefined	transliterate
negro	can	unripe	phonetic	baste	opaque	unzip
lin	weeknight	jar	westernize	charred	crooked	sell
ngo	mesh	acidic	gringo	soupy	numb	blot
buddhist	cast	ginkgo	iguana	gobi	lively	wear
vega	preference	mosquito	eatery	curried	shriek	drain
rick	also	medicinal	immigrate	minced	intriguing	sterilize
kyoto	slow	diagonally	language	sweetness	blasphemous	recombine
punjab	win	kelp	student	capsicum	rare	stimulate
zulu	preserved	berber	census	piquant	cheery	sweeten
mao	depth	strainer	southeastern	spicing	intense	swirl
cuon	kick	desiccate	menopausal	deveined	pliable	recreate
tao	balance	resin	elderly	spiced	boring	borrow
yang	dice	reddish	immigration	sweetened	soaking	scorch
t	whistle	container	nationalist	tang	mystical	skip
shabu	roughly	mezcal	pueblo	coarsely	spattering	gobble
ba	general	tropic	smoker	nourishing	intricate	pronounce
eden	introductory	teflon	homesickness	tasso	zippy	comprise
min	share	fuchsia	screenplay	crisping	warmth	migrate
wichita	countless	alkaline	worker	toasted	lite	squirt
bach	towel	subtropical	nationality	kadai	wonderful	meshed
kwai	excited	ceramic	servant	comal	hungry	baffle
kennedy	immune	cylindrical	music	juicy	gracious	slake
mei	burner	burdock	outsider	broiled	sizzling	vent
amazon	indefinitely	brazier	racial	chewy	irresistible	synchronize
sally	typically	hibiscus	sorcery	spicier	powerful	inveigle
bali	full	diagonal	northeastern	tenderize	tingle	disabuse
nu	domino	headspace	exile	dishes	inauthentic	retrain
raja	tolerance	afro	ancestral	piquancy	popping	rip
bai	most	husked	mythological	zesty	perfect	energize
xe	morton	osmanthus	acreage	broiling	crappy	define
ip	basket	mealie	gambler	culinary	sourest	enliven
causa	complete	serape	conscionable	palate	hothouse	translate
ka	tempering	chrysanthemum	griot	recipe	interesting	depart
hindus	minimal	lactic	machete	griddled	lazy
atlanta	integral	frizzle	shrimping	jellied	eclectic
mama	fearing	insulation	community	fatty	omnipresent
wednesday	like	brine	mason	crosswise	exciting
lama	version	saran	ecological	minty	slimy
chennai	plank	laver	arid	puréed	mojo
xi	international	slit	caterer	crispiness	saucy
doc	length	smell	festival	julienned	mildly
kampong	mott	bikini	geography	citric	sourer
poon	travel	splatter	ethnicity		dusty
para	dynamite	probiotic	ethnic		junky
phoenix	other	octagonal	strait		playful
austin	non	loofah	kin		wilted
mara	crunch	conical	consul		crave
pronounced	sick	colander	jungle		unstoppable
chino	numerous	impurity	modernist		sterile
god	podcast	pistachio	singleton		kid
diana	ceremonial	restorative	franc		distinctive
ramadan	rangoon	pummelo	author		synonymous
toledo	finishing	shade	hobo		sloppy
goma	level	souk	fertility		sprite
ana	dear	oxidant	brewer		untraditional
pei	visit	detox	harem		literal
shang	long	eddy	caravan		ornate
dodo	copyright	caffeine	ranch		indestructible
gao	fatima	milky	population		explosively
tod	gathering	alkali	civil		counterintuitive
jain	moderately	bulbous	doctor		bliss
lucknow	handful	fiddlehead	psychologist		unlit
ali	½	gelatinous	ecumenical		blistered
ra	here	lacquered	discernment		massaged
ming	systematically	hydrolyze	intermarried		licked
btu	uptown	proton	dance		slanted
ting	dress	hydroxyl	commander		unobtrusively
pb	explosive	muslin	idleness		slippery
jakarta	bicycle	mottled	cacique		occasionally
pong	stall	spherical	autobiographical		innocent
bong	manchurian	scissor	cremation		saturated
bangalore	promised	bin	ghillie		heady
tum	portable	wisps	avocation		liking
kip	ensure	gunpowder	prayer		mighty
ho	alongside	mouth	shrine		butchered
mick	welcome	earthen	ancestry		potent
som	below	facial	physician		beefy
devi	huge	rasp	martial		unreal
himalayan	grapevine	bud	mercantile		vivid
bc	vein	papery	adolescence		frightened
afghani	spike	thong	shire		wildest
anu	cornerstone	eelgrass	shrimper		sniffy
akan	manually	fridge	travelogue		nebulous
kali	excerpt	lb			gleeful
na	then	crosshatch			voluptuousness
rama	prospective	brightness			unmindful
maharashtra	technique	gingko			jarred
bollywood	type	adventitious			stormy
qi	player	longan			craggy
hua	stack	burr			feverish
maha	changer	carport			iconic
senor	arranged	collagen			contagious
sanchez	tat	penicillin			pleasantly
boston	firecracker	repellent			uncomfortable
epi	known	granule			smacking
chick	maximum	bacteria			artful
zee	glazer	halo			glowing
gan	flare	rosebud			authenticity
papas	inherent	antiseptic			versatile
mona	piper	unwaxed			devious
ji	valuable	incense			obscene
ge	block	receptor			hallowed
kashmir	element	krill			envious
lu	honored	lung			slaked
fen	eurasian	plaster			umpteenth
buddhism	step	garland			grandest
shah	regular	yellowfin			incredible
pang	timely	lacy			impatient
rivera	brewing	sumac			moxie
l	distinguished	desiccated			magical
santa	uncommon	gloved			sheer
dana	usually	terraced			characteristic
uma	essential	fuschia			richly
shawn	legitimate	digestibility			reflective
lutheran	live	silica			radiant
tehran	2	cornhusk			subterranean
moslem	healthiest	arco			unturned
lei	alternatively	aldehyde			senseless
perth	blessing	maroon			effortful
dharma	policy	sapphire			indiscernible
git	discouraging	floss			aimless
karachi	illness	herbaceous			symbiotic
imo	seventh	amino			brutally
rani	wrapping	beany			brighter
wu	cart	lengthwise			exquisiteness
tung	makeshift	slime			tangent
pachinko	traditionally	spire			shimmering
n	chase	flaked			superstitious
goldman	fulfilling				unruly
albuquerque	speedy				ravenous
cornish	call				subdued
peri	starting				rawness
david	irreversible				hedonistic
scum	trip				hypnotic
moore	administrative				lamentably
ko	prepackaged				faintly
	inter				lonely
	unheeded				swish
	diplomatically				authentically
	including				heaping
	cleanup				stoner
	complexity				gregarious
	exclusively				stinky
	requisite				fishy
	partially
	contrast
	appear
	aisle
	hit
	concession
	extensive
	shopping
	consistency
	carom
	unit
	honorary
	8th
	disposable
	unsanitary
	he
	grouping
	weill
	infusion
	originally
	hundred
	undisturbed
	viral
	leant
	inadequately
	ctc
	consultation
	attrition
	unfolding
	disrepair
	piñata
	fixing
	optimal
	blending
	formerly
	symbolic
	anti
	hung
	coincidentally
	cheap
	overwhelming
	stature
	note
	chopping
	outgoing
	underlying
	young
	easily
	routinely
	member
	backyard
	roadside
	inaccessible
	overcrowded
	suspicious
	courtesy
	smashed
	prefer
	wish
	ideal
	curbside
	nairobi
	unbrushed
	theoretical
	favored
	refining
	wraparound
	cassette
	gong
	intensive
	roulette
	ratio
	rut
	chipped
	abide
	absolute
	option
	latter
	package
	utmost
	steep
	function
	ultimate
	middle
	deeply
	eleventh
	immeasurable
	declaration
	notation
	homecoming
	unopened
	affordable
	effect
	lock
	know
	remiss
	kit
	drum
	tenet
	diversion
	finally
	exploratory
	diagraming
	deranging
	schematic
	unrestricted
	adaptive
	embedded
	trapped
	unconverted
	gravely
	lob
	refusal
	winning
	asean
	feasibility
	registration
	waterline
	interchangeably
	spoken
	method
	ploy
	appreciative
	overseas
	dirt
	selective
	cognitive
	trucking
	levy
	football
	safety
	important
	permission
	consecutive
	composite
	permissible
	fruitful
	finder
	sometime
	accordingly
	basic
	instantly
	weeklong
	propane
	similarity

W clusters¶

Proper nouns	Common nouns	Core cooking terms	Secondary cooking terms	Luxury and rusticity	Adjectives	Verbs
june	sufficiently	unflavored	amber	cinchona	amusing	stiffen
moscow	hawk	kneading	degrease	reindeer	weary	tremble
freeman	thicket	wholemeal	carmine	manor	heaped	look
portland	nature	poussin	pearled	sweater	childish	elude
october	excuse	brininess	braided	skiing	paean	leaven
crosby	tune	roasting	grating	pagan	charm	bother
m	narrower	glug	tentacle	vulture	toweling	rearrange
sir	pointer	diced	faucet	conger	seductive	suckle
adam	peaked	oxidize	unglazed	deer	clutch	scraping
berkeley	dispute	calorie	handkerchief	dory	frantic	destroy
sunday	counter	cupful	nitrate	carnival	witty	divvy
f	broke	runny	glazed	bath	stylistic	lure
catherine	perpetual	blanch	spacing	countryside	demeanor	absorb
leicester	seminar	panini	fahrenheit	pâtisserie	lightened	reinvent
herman	unreduced	floury	rubber	gastronomic	unstrained	revive
modifier	placement	rennet	wineglass	auberge	grownup	mull
september	trapping	calvados	doily	prefect	grinding	weave
jerusalem	cholesterol	tannic	blowtorch	godmother	contemporary	tear
cornwall	area	tannin	knuckle	restaurateur	cherished	subside
c	damage	filleted	bulky	furniture	magnificently	seduce
dublin	pleased	gooey	dense	sommelier	flashy	connect
atlantic	strive	dollop	sill	kahlua	impressive	ignite
vienna	tomorrow	ripened	centimeter	comte	timeless	trace
v	nonetheless	camembert	frosted	rioja	rosy	govern
schmidt	borough	gouda	textured	merino	extraneous	distribute
joel	forth	moisten	tumbler	springtime	extravagant	thaw
christ	longer	crumbly	spongy	mixology	uneven	shiver
rockefeller	rounded	refrigerate	sinew	quay	twist	prick
arnold	fee	lycopene	sculpin	mollusk	famed	miss
soho	report	boil	stainless	housewife	tad	hide
amish	active	deglaze	twine	lyonnaise	chilling	render
anderson	regulation	emulsion	envelope	pope	lavish	fortify
nancy	self	syrupy	translucent	chablis	festive	weigh
copenhagen	entry	yeasty	eggshell	cheesemonger	fussy	adhere
br	relation	spritz	molten	distillery	energetic	bind
somerset	shaping	tablespoonful	indent	sable	warmed	push
harry	is	salted	perforated	dogfish	homey	convince
mary	offer	gizzard	needle	boletus	pricey	sever
baldwin	amateur	chilled	flap	corpse	twisted	mar
arthur	stride	creamy	incision	votive	celebrated	crushed
albert	widest	fruity	cottony	rouge	bubbling	trim
kaufman	workable	ungreased	puffy	tuxedo	stuff	unroll
scott	accidental	freezer	genoa	deli	hardened	dredge
fanny	pay	flaxseed	nonmetallic	benedict	sexy	attach
milano	exceedingly	briny	cork	duchess	sensual	ooze
mls	immediately	curdle	pane	grandson	damp	sifted
judith	perhaps	emulsify	colored	carol	lent	spiral
anne	trash	quart	mallet	monastery	fiddle	enrich
cr	away	unbleached	milligram	woman	ultra	wilt
dover	shipment	sauteed	scape	yule	sublime	overflow
andrew	costly	cutlet	dye	village	fanciful	shrink
maria	test	browned	foam	delis	blond	lighten
michael	prize	foamy	feathery	lenten	sturdy	shave
hamilton	residual	moist	scalloped	alp	duo	blossom
bernard	conversion	glace	soil	lodge	indulgence	drape
keller	parameter	frosting	indentation	brut	musical	unfold
noel	blockbuster	teaspoonful	superfine	tasting	orthodox	tap
sheeps	lawsuit	brie	disc	symphony	fashionable	resemble
da	eight	stilton	greenish	fisherman	ethereal	slip
monte	cruise	semisweet	cavity	pounder	enthusiastic	lift
dolce	reasonably	rind	speck	hermitage	shower	crumble
slater	sorry	buttery	silicone	grove	lusty	decorate
franco	indeed	sangria	ink	mare	embellishment	slide
welsh	minuscule	drizzle	particle	automat	creamed	crimp
brittany	stamp	floured	hole	hotel	boning	remove
oliver	prestigious	gruyere	cutout	sucre	rocky	deflate
teresa	preliminary	dijon	snip	chapel	simplicity	loosen
allen	pain	marsala	hollow	jean	shiny	gather
barcelona	scored	buttered	liner	shop	airy	come
wellington	out	knead	serrated	noma	unpredictable	hook
patrick	nick	balsamic	sediment	distiller	blustery	greased
devon	council	breadcrumb	rosette	shepherd	cornucopia	freeze
halle	will	hazelnut	thickly	horseback	chic	poach
jonathan	prolonged	unsalted	sifter	riviera	generous	flatten
guinness	waning	sprinkle	oversize	dining	interior	lynch
monsieur	convinced	yeast	bra	farmhouse	forte	insert
louis	mistake	preheat	decorative	harvest	impeccable	pour
olivier	blind	bake	rim	forest	luscious	rotate
parker	question		cardboard	hearth	classy	pierce
europe	preparation		nest	salmonella	superb	choke
beck	thoroughly		flowerpot	robin	airtight	incorporate
jacob	locally		wheel	manzanilla	luxurious	scatter
les	commercial		swivel	swan	admire	hold
z	consumer		wreath	cellar	circus	fill
nero	partisan		dimple	brasserie	elegance	arrange
lyon	pattern		capri	bellini	whisking	scrape
madrid	selection		hipbone	cloudberry	underdone	fit
florence	subject		dome	patisserie	zigzag	shape
yorkshire	emmental		braid	framboise	autumnal	sift
milan	finish		gill	butler	tepid	melt
paul	owner		clump	linen	stern	beat
anna	idea		rimless	valentine	blend
finnish	mini		thickness	piazza	sleek
venice	rapid		fluted	decoration	grainy
pierre	faced		tubular	greenmarket	neatly
paris	grant		granulate	tapa	simmering
rome	heavily		tine	salade	luxury
christmas	title		inch	waiter	frothy
	resilient		decoratively	château	loosely
	stroke		casing	russet	waxed
	tunnel		bladed	geese	fashion
	visible		metal	medieval	fine
	cumberland		oval	stein	touch
	careful		tube	hunter	fragile
	attention		shallow	boar	wet
	appeal		roller	nursery	polish
	demand		rectangular	flute	flaky
	eighth		paddle	saba	excellent
	vision		grate	opera	cheesy
	certain		spray	ski	sparkle
	modest		scraper	peasant	decadent
	interval		lattice	hill	fancy
	trend		rope	heirloom	boozy
	6th		diameter	sundried	glossy
	motor		concentric	winery	oiled
	relentlessly		dot	coupe	generously
	base		grease	trotter	lukewarm
	suddenly		tin	chef	beater
	dean		tester	cirque	stirring
	untouched		frond	apprentice	zest
	continually		elastic	pub	classic
	blight		fingertip	granny	stale
	exhibition		frost	gourmet	chill
	suggest		bulb	tavern	elegant
	quite		ounce	thanksgiving	crusty
	nothing		gently	farmer	lightly
	recently		burgundy	tapas	stiff
	internal		foil	sauterne	warm
	very		disk	tent	freshly
	structure		brush	chanterelle	cool
	tilt		pipe	calf
	shrinkage		sprig	chianti
	shipping		removable	vineyard
	notion		temperature	chateau
	charge		thick	roquefort
	letter		rimmed	butcher
	stable		rectangle	pizzeria
	tough		dust	provencal
	core		coarse	alsatian
	presumably		surface	pinot
	fora		plastic	caesar
	xiv		melted	bakery
	incorporated		parchment	nordic
	defiance			bartender
	premier			processor
	speech			mediterranean
	allow			armagnac
	apprenticeship			bistro
	watchful			lavender
	nightcap			bordeaux
	halfway			pint
	run			countess
	scout			beaujolais
	male			madeira
	deputy			provence
	integrity			bouquet
	steady			baker
	success			cognac
	overlapping			gruyère
	mass			rustic
	tour			morel
	hunt			confectioner
	reputation			virgin
	alternately
	blemish
	assembled
	rate
	ramp
	apart
	career
	door
	force
	nineteenth
	inevitable
	tension
	book
	job
	crop
	intact
	aid
	receive
	age
	reduced
	seasonal
	portion
	almost
	verge
	midwestern
	update
	uniform
	meanwhile
	too
	concord
	reception
	creation
	manger
	child
	same
	stress
	boss
	lorraine
	there
	total
	vacation
	producer
	county
	size
	par
	today
	name
	morale
	recommendation
	domestic
	cauda
	moderate
	over
	possible
	ready
	accumulated
	closely
	spoke
	definitive
	aged
	have
	slump
	professional
	upside
	paring
	reduction
	inside
	financier
	recall
	fourth
	covering
	row
	vigorously
	flow
	guide
	off
	yukon
	holiday
	price
	victor
	say
	reserve
	seem
	sided
	offset
	strata
	reverse
	not
	topping
	volume
	overwork
	individual
	barely
	correct
	return
	be
	work
	carefully
	slightly
	shaving
	quarter
	pass
	stream
	hour
	equal
	irregular
	increase
	addition
	quality
	tie
	briefly
	space
	rapidly
	constantly
	bulk
	wide
	excess
	stretch
	necessary
	overlap
	enough
	ray
	reserved
	skate
	court
	yield
	do
	least
	additional
	remain
	border
	slotted
	scrap
	import
	approximately
	sharp
	time
	still
	completely
	press
	clean
	ahead
	log
	evenly
	cure
	continue
	attachment
	about
	stand
	weight
	position
	tender
	slowly
	overhang
	extra
	circle
	room
	low
	speed
	half
	fold
	edge
	form
	rise
	purpose
	gradually
	peak
	prepared
	degree

Yewande and Priya's hypothesis: describing NW recipes as "fast" and "easy"¶

Returning to one of the motivating questions behind this project--**whether there is empirical support for Yewande and Priya's suspicion that American food media tends to qualify non-Western recipes as "fast" and "easy"** (to make them more approachable for white audiences)--I will show that non-Western recipes are indeed more likely to be associated with adjectives indexing speed and ease compared to Western recipes.

However, I also find that non-Western recipes take less time to prepare than Western recipes using the provided estimated cooking times, and that non-Western recipes have fewer steps than Western recipes.

Nevertheless, when comparing recipes binned by similar preparation time and number of steps, non-Western recipes are still more likely to be associated with adjectives indexing speed and ease.

Cooking time analysis¶

Here is the distribution of (log) recipe preparation times for recipes with cooking time information available ($N=3326$):

Image("./figures/cooking_time_dist.png",width=600)

Here is the distribution of (log) recipe preparation times, disaggregating w from nw recipes:

Image("./figures/cooking_time_dist_w_v_nw.png",width=600)

Here are the cooking times broken down across the 15 most common cuisines:

Image("./figures/cooking_time_boxplots.png",width=700)

From a one-tailed t-test, it looks like there is a meaningful difference between preparation time of w vs. nw recipes ($t=3.4, p<0.001$). But within each bin, are nw recipes still more likely to be described as quick? We bin recipes into one of 4 bins based on the cooking time provided in the recipe notes, as shown below:

Image("./figures/cooking_time_binned.png",width=700)

We find that among the words indexing "fastness", only instant is significantly more likely to be applied to either w or nw recipes for certain bins. However, it is always more likely to be applied to nw recipes, even when the cooking time is similar to w recipes.

Image("./figures/lor_boxplots_cooking_time.png",width=400)

Cooking ease¶

I used the number of steps as a proxy for ease of a recipe (the fewer steps, the easier the recipe). The distribution of recipes over (log) number of steps is shown below.

Image("./figures/cooking_ease_dist.png",width=500)

Here is the distribution of (log) cooking ease, disaggregating W from NW recipes:

Image("./figures/cooking_ease_dist_w_v_nw.png",width=600)

And the cooking ease broken down across the 15 most common cuisines:

Image("./figures/cooking_ease_boxplots.png",width=700)

From a one-tailed independent t-test, it also looks like w recipes do have significantly more steps ($t=18,p<1e-71$). But again, we can ask--within separate bins, are nw recipes still more likely to be described as easy? I binned recipes into one of 4 bins based on the number of steps in the recipe directions, as shown below:

Image("./figures/cooking_ease_binned.png",width=600)

Again, I found that among words indexing ease that are significantly more likely to be applied to either nw or w recipes, all are more likely to be applied to nw recipes, even though the number of steps are similar.

Image("./figures/lor_boxplots_cooking_ease.png",width=400)

Footnotes¶

[1] http://www.intersectionalanalyst.com/intersectional-analyst/2017/1/7/who-gets-to-be-an-authority-on-ethnic-cuisines ⤾

[2] https://www.bonappetit.com/story/recipe-writing-whitewashed ⤾

[3] I adapted code from recipe-box and this recipe scrapers module to retrieve URLs associated with recipes from each site's archive, in addition to each recipe's title, author, ingredients, directions, free text, and cuisine category, when available. I regularized all author and title text by keeping only alphanumeric characters (i.e. removing punctuation), lowercasing, but crucially retaining diacritics (as I assumed diacritics might be a useful signal in classifying recipes to a cuisine category). I deduplicated recipes using regularized author and title strings, in addition to the recipe's source website (so no two recipes in my dataset were contributed by the same author, with the same name, from the same website). I further removed duplicate recipes from Bon Appetit, which were listed twice due to a separate listing with the photographer as the author, and "test" recipes from Food52, where users can create and upload recipes freely, and so often create submissions for the purpose of testing out the interface. The latter can be identified by manually compiling a set of frequently occurring recipe titles to toss out, such as "test" and "cake". ⤾

[4] Analysis data, figures, and code notebooks for this project are available at this repository. ⤾

[5] I compiled a list of demonyms from Wikipedia (https://en.wikipedia.org/wiki/Demonym#Suffixation), excluding the demonyms "hamburger" and "frankfurter" as these have more prominent culinary meanings. ⤾

[6] For each list, I used the process below to extract names of individual dishes:

1. Get all hyperlinks with anchor text matching the pattern "X cuisine"
1. For each cuisine hyperlink:
  - A. Grab all items occurring in a list -> individual dishes/ingredients
  - B. Add to dictionary tracking ingredients/dishes of a cuisine ⤾

[7] http://languagelog.ldc.upenn.edu/myl/Monroe.pdf ⤾