Project Assignment (1500 words)

* All authors contributed equally.

An Analysis of Categorical Biases in Word

Embeddings

Abstract— This work focuses on discovering forms of bias in

word embeddings through the use of the Word Embedding

Association Test (WEAT). A thorough categorical analysis is

performed to distinguish among different forms of biases, to

discover the extent and importance of it in different domains

and also uncover biases that are commonly not expected.

Moreover, a model specifically designed and trained to mitigate

bias in genders is also tested to see how much it can mitigate the

effect of embedded bias in training corpora. Finally, to discover

how those biases came to be, we perform an analysis on a

historic dataset which can illustrate how biases evolved and

whether there is a consistency in their current estimated form.

I. INTRODUCTION

Word Embeddings trained on Natural Language Processing

(NLP) and machine learning models are vulnerable to

rampant biases. These biases maybe desirable or undesirable.

Significant work has been done to develop de-biasing models

for word embeddings. But, making a model absolutely bias

free can be difficult. These biases are mostly generated due

lack of representational data, or as a direct influence of our in

bred societal biases that get incorporated in the data sets that

are used to train these models on.

With the advent of Artificial Intelligence and technologies

built on it, word embeddings find major real-world

application. For example; recommendation systems on e-

commerce websites find use of word embeddings to suggest

search-keywords or merchandises to its users. Websites for

professional networking can generate biased output too. If a

recruiter looks for suitable candidates for a position, biased

output can be generated based on the candidate’s gender,

location etc. If the recruiter enters the search keyword

“programmer”, the algorithm may tend to return resumes of

male candidates with higher priority, as dictated by the word

embeddings which associates the profession, “programmer”

more with the male gender. This is an undesirable bias which

should be eliminated from the model to make the application

fair. Again, when the results returned tend to be biased

towards candidates with residence in a particular location, to

some extent they might be desirable biases. For example;

when a company which is searching for candidates to fill a

position in London, if the results returned give low priority to

candidates from Tokyo, then it may not be completely

undesirable. This indicates, on top of eliminating undesirable

biases, algorithms must also be trained to identify and retain

the desirable biases.

Previous work done in this area have targeted biases like

gender bias, ethnic bias, temporal bias in historical data, etc.

The most used method is WEAT (Word Embedding

Association Test) to quantify and understand the extent of

biases in different word embedding models. Publicly

available datasets, like those from the areas of Journalism

(Google News data corpus), Social Media (Twitter data) as

well as Wikipedia data have been used for these analyses.

This paper will present a study on biases in word embedding

models by identifying sub-categories in areas like gender,

society, religion, racial and non-people categories. The biases

are measured against attribute sets like man vs woman, rich

vs poor, slow vs quick, etc. Four different word embedding

models trained have been used to measure the bias in each of

those target and attribute sets. The models used are: (a) the

micro-blogging site, Twitter’s data corpus trained on the

GloVe algorithm; (b) Google News data corpus trained on the

word2vec algorithm; (c) Wikipedia data corpus trained on the

GloVe algorithm and (d) Historical Data corpus from 1820 to

1990. Plots to visualize the biases will be constructed based

on the data from those derivations. The results of a de-biasing

algorithm run on the same dataset will be discussed along

with the causes of presence of such biases. To conclude, a

summary of future work required will be listed.

II. BACKGROUND AND RELATED WORK

This section provides a brief overview on word embedding,

their uses and applications. It also touches upon the previous

work done in this area and weighs in upon what are identified

as major problem areas. It concludes with a summary of w-

hat was aimed to be achieved in this paper and methodology

used.

At a higher level, word embedding can be described as vector

representation of words in a dictionary which is capable of

capturing the geometric distances between the words in it.

These distances can then be used to group together similar

words based on their proximity to one another. With the

Artificial Intelligence (AI) driven decision making tools

picking up pace in the real world, the biases in word

embedding models become more prominent. Prior work done

on biases in word embeddings show a wide variety of biases

pertaining to different categories. These categories may range

from gender, race and location to political concepts. These

works also reveal how the extent of bias in a category towards

an attribute can vary with the data corpus the word

embedding model has been trained on. For example; gender

bias in twitter data is lower than that observed in a word

embedding model trained on the Google News data corpus.

Such changes can also be temporal. This study [1] on models

trained on 100 years of historical data of the American society

reveals that, ethnic biases in word embeddings have changed

with respect to time. This may be attributed to changes in

societal constructs and diaspora of the country. We will talk

more about the causes of change in the analysis section of this

paper.

Significant work has been done to study and analyze gender

biases in word embedding models. In [2], the author studies

the prejudices of one gender against another in different

categories based on word embedding models trained on

different type of data sources. The data sets used are GAP,

Google News, Twitter and PubMed; and the categories

studied being Career vs Family, Science vs Arts, Math vs

Arts, Intelligence vs Appearance and Strength vs Weakness.

Different datasets showed different range of bias in those

categories- the Google News corpus showed bias in all the

categories unlike PubMed which showed relatively less bias

in all of the categories studied. Further, the author tries to

auto detect categories in word embeddings which display

gender biases based on the proximity of male- and female-

pronouns to other words in the dictionary. In [3], it is asserted

that, the use of unsupervised algorithms may lead to the

generation of bias when run on user generated or historical

data. It can lead to bias generation mimicking the bias in the

original documents later on, or, by identifying the biased

patterns in the original data as key concepts and them being

fundamental to the area. For example; as illustrated in [3], the

word Prime Minister was only associated with the male

gender prior to 1960. Word embeddings run on data till 1960

only associate the male gender with the Prime Minister as no

data is available for a woman holding the position till then. It

argues that the presence of genderless nouns in English can

be used to expose the correlation between he and she

stereotypes in data.

Reference [3] attempts to de-bias gender-neutral words by

eliminating the gender associations with them as a post

processing step. It is proposed two ways – (i) Hard de-

biasing, which involves manual intervention; (ii) Soft

Debiasing, where computer programs are responsible for de-

biasing sans any human intervention. According to the study,

hard de-biasing methods work better at this point in time as

they could not find an efficient enough de- biasing algorithm.

However, they do not consider the possibility of valid gender

associations with non-gender-definitional (or gender neutral)

words, while de-biasing - like the association of beard with

men. To overcome this problem, [4]proposes a model that

retains the desirable gender biases whilst doing away with the

undesirable ones. It identifies four sets of words in a

dictionary – (i) masculine, (ii) feminine, (iii) neutral and (iv)

stereotypical. Based on this classification, it models an

interaction that can retain the masculinity of a target, retain

the femininity target, protect the gender neutrality in the

target and remove the gender biases.

Though the elimination of desirable biases in [3] can be

overcome by implementing the method laid out in [4], the

models are not widely trained and it is unknown how the

model might run on other categories like race, location, etc.

Both the models try to categorically identify the biases with

respect to the attributes (in these cases, gender). But there

may be other correlated unidentified attributes which may be

biased towards these categories and may lead to the

production of biased results.

This paper focuses on the analysis of categorical biases in

word embeddings. Instead of focusing on a particular

category or attribute set, tests have been run on a diverse set

of sub-categories from broader areas and the biases they are

subjected to. The biases generated when these targets are

pitted against different set of attributes have been studied. For

example; we have studied the social (category) bias

pertaining to one’s ethnicity (sub category) using the WEAT

algorithm. We checked the bias for Germans vs Italians

(target) for the attribute set of lazy vs hardworking, which are

preconceived notions popularly associated with our target set.

We also study the variations in biases depending on the

training data sets for the word embeddings- we have used the

Google News data, Twitter data, Wikipedia data and

historical data from the 19th and 20th century. We will try to

reason the possible source of these biases at the discussion

section of this paper.

III. METHODOLOGY

a. Biased Models

The approach followed, attempts to answer the points

discussed in the introduction. We follow three main

directions and depending on the availability of data and code

packages, we perform a thorough analysis and plot the

results. The points we tackle can be summarized as follows:

• Biased models trained on multiple text corpora.

• Debiased model trained on text corpora to mitigate

bias.

• Historical development of biases through historic

corpora.

Figure 1: Three-part analysis followed

To recognize bias in word embedding models, first a selection

of available word embedding models on large corpora needs

to be done. For this purpose, three datasets are down-selected

to form the basis of our analysis.

The first dataset is Google News [5]. It is based on articles

from the same name website and consists of large size data.

It contains 100 billion words and totally 3 million different

trained words. The size of the word embeddings, meaning the

representation size of each word, is 300.

The second model used is that of Twitter which has 1.2

million different words and contains embedding sizes of 25,

50,100,200. [6]

The third model is that of Wikipedia which contains 400

thousand different words and embedding dimensions of

50,100,200 and 300.

Google News is trained based on the Word2Vec method

while the others on the Glove method.

Word2vec [7] is a two-layer neural net that takes as input text

and vectorizes it into decimal numbers. Specifically, it

vectorizes words. It is similar to an autoencoder in a sense

that it is trained against other words that neighbor them in the

input corpus. This can be done with two ways. One way is

through Continuous Bag of Words which uses context to

predict a target word.

The other way is through Skip-Gram which given the word

predicts the neighbor words. When the feature vector or else

embedding assigned to word does not predict accurately the

context, the vector is adjusted. An accurate model with proper

training will place similar words close to each other.

Figure 2: CBOW and Skip-Gram as methods for embedding

learning

GloVe [8] is a model trained on non-zero entries of word to

word co-occurences which recognizes how often words co-

occur in a corpus. It is essentially a log-bilinear model for

unsupervised learning with a weighted least squares

objective. Semantic similarity is learned between words in

this way.

The characteristics of the models are shown in table 1. The

trained models that were used, are of dimension 50 for twitter

and Wikipedia and dimension 300 for Google News, while

they can be found at the gensim library where they are open

sourced [9].

b. De-biased Models

In order to uncover whether a model specifically trained to be

debiased, can perform better than those who are not, a popular

model for tackling gender bias is presented.

Reference [3] addresses the problem of bias by defining a

subspace which identifies the direction of the embedding that

captures the bias.

Then the algorithm has two options, being to neutralize or

soften. By neutralizing, it ensures that gender neutral words

are zero in the gender subspace. On the other hand, softening,

perfectly equalizes words outside the subspace and makes

sure that the any neutral word is equidistant to all words in

each equality set.

As an example, given the words grandmother and grandfather

and the words guy and gal, namely two equality sets, after

softening, the word babysit would be equidistant to

grandmother and grandfather and also the same to gal and

guy. A parameter also can control how much similarity is

maintained to the original embedding since this can be useful

in some applications.

For the purposes of this analysis, a trained model is used

which can be found here [10]. This model has been trained

using a combination of words to tackle gender specific

embedding bias on the aforementioned Google News dataset.

Datasets used Characteristics

Google News about 100 billion words,

300-dimensional vectors for

3 million words and phrases

Twitter 2B tweets, 27B tokens,

1.2M vocab, uncased, 25d,

50d, 100d, & 200d vectors

Wikipedia 6B tokens, 400K vocab,

uncased, 50d, 100d, 200d, &

300d vectors

c. Historical Models

In order to understand the language evolution and analyze

single and mutual biases in word usage throughout a certain

period, the lexical data about frequencies of word

appearances in the corpora for different years should be used.

Nowadays, this kind of time-wise analysis has become

possible due to development and dramatic expansion of

Google Books N-gram corpora, which comprises large

collections of books and similar materials printed between

1500 and 2008 in 8 different languages (English, Chinese,

French, German, Hebrew, Italian, Russian, and Spanish),

with a total of over 5 million books. The dataset includes

information about frequency of usage of n-grams (where n is

2 or higher). Such extensive text corpora allow researchers

for solving a wide range of natural language processing, and

exploring word embedding bias analysis is not an exception.

The authors of [11] provide an extensive statistical analysis

of words to examine two proposed statistical laws: the law of

conformity and the law of innovation. In this work,

researchers use 6 historical datasets to analyze historical

change, which are essentially subsets of Google Books N-

gram corpora. The datasets then have been used to obtain

different groups of word embeddings by applying several

techniques, one of which is word2vec based one. Given the

sparsity of data between 1500 and 1800, it is recommended

to exploit lexical data from the materials published after

1800.

In our work, we perform the analysis over various groups of

words from diverse semantic groups which might have

represented an expression of bias. Hence, it makes sense to

use a set of word embeddings with a sufficiently multivariate

set of words. For this reason, we decided to use the words

from “All English” dataset, which includes data from Google

books of all genres published between 1800 and 1999, with a

total of 8.5*1011 tokens.

We focused our attention on word2vec embeddings

pretrained on this dataset (SGNS) kindly provided by the

authors of the work on the HistWords project GitHub page 

[12]), which contains multiple tools and word embeddings.

SGNS dataset represents a group of files in a special format,

which is not compatible with gensim library, which is utilized

in our research. Thus, the following steps should be taken in

order to produce WEAT scores:

1. Word embedding files are to be converted into .txt

gensim-compatible format

2. Extract embeddings for words from our categories

and calculate mutual biases

3. Plot and discuss the results

d. Word Embedding Association Test

The Word Embedding Association Test is a statistical test

which aids in understanding the relation between words in

embedded in text corpora.

Considering that we have two sets of target words (X and Y)

and two sets of attribute words (A and B), we want to discover

if there is a difference between the former in terms of their

relative similarity to the latter.

Specifically, since this is a statistical test, we have the null

hypothesis that there is no relative difference between the

target words and the attribute words. We measure the

(un)likelihood of the null hypothesis by computing the

probability that a random permutation of attribute words

would produce the observed or greater difference in sample

mean.

The statistic of the test is derived by:

𝑠(𝑋, 𝑌, 𝐴, 𝐵) = ∑ 𝑠(𝑥, 𝐴, 𝐵)

𝑥∈𝑋

− ∑ 𝑠(𝑦, 𝐴, 𝐵)

𝑦∈𝑌

with:

𝑠(𝑤, 𝐴, 𝐵) = 𝑚𝑒𝑎𝑛𝑎 ∈𝐴 cos(𝑤, 𝑎) − 𝑚𝑒𝑎𝑛𝑏 ∈𝐵 cos(𝑤, 𝑏)

where cos the cosine distance, s(w,A,B) measuring the

association of w with the attribute and s(X,Y,A,B) measuring

the differential association of target words and attributes.

The test being performed is a permutation test and if {(𝑋𝑖 , 𝑌𝑖 )} denotes all the partitions of 𝑋 ∪ 𝑌 into two sets of equal size, the one-sided p-value of the permutation test is:

𝑃𝑟𝑖 [𝑠(𝑋𝑖 , 𝑌𝑖 , 𝐴, 𝐵) > 𝑠(𝑋, 𝑌, 𝐴, 𝐵)]

effect size = 𝑚𝑒𝑎𝑛𝑥∈𝑋𝑠(𝑥,𝐴,𝐵)− 𝑚𝑒𝑎𝑛𝑦∈𝑌 s(𝑦,𝐴,𝐵)

𝑠𝑡𝑑−𝑑𝑒𝑣𝑤∈𝑋∪𝑌𝑠(𝑤,𝐴,𝐵)

with the effect size being a normalized measure of the

separation of the two distributions of associations between

target and attribute.

By calculating the effect size of this statistical test, we follow

basically what is called “Cohen’s d” which calculates the

standard mean difference between two groups. An effect size

of 1 indicates that the groups differ by 1 standard deviation

and 2 indicates 2 standard deviations. Cohan has suggested

that d=0.2 is considered a “small” size, “0.5” medium and

above “0.8” large [13].

For performing the test, our code development was based on

a template to perform such analysis which can be found here

[14].

e. Analysis Procedure

The analysis performed aims primarily to measure and

compare the size of bias in word embedding in the

aforementioned models. In order to do that we focus on a

thorough categorical analysis when this is possible and try to

recognize in which cases there is evident bias.

The procedure for the analysis is illustrated in figure 3 and

with the main goals of:

• Quantify bias through the use of the WEAT.

• Compare between training corpuses.

• Compare between attributes of the same

subcategory.

• Compare between inter-category attributes.

• Compare between intra-category attributes.

• Discover whether the de-biased model achieves

better results.

• Discover the historical path of some formerly

discovered biases.

Figure 3: Processing pipeline for analysis

The categorical analysis focuses in the case of biased models

in specific categories which are outlined in table. The most

common suspected categories for bias are described and

include gender, race, religion, social and objects. Those

categories are split into subcategories.

Category Subcategory

gender work/education

character

sexual orientation

religious -

racial color

ethnicity

citizenship

social age group

prof./economic group

political

non-people -

Table 1: Categorical analysis

For each word set, a list of words are used which are

synonyms or close to its context. Below are some examples

for the two cases of comparison. For each case, we define a

set X, a set Y and their relation in comparison to A and B is

put to be tested through WEAT. Sets like these form the basis

of our comparison and multiple cases have been tested.

X: career career

profession

work

successful

Y : family family

life

kids

wife

A: man man

male

he

himself

B: woman woman

female

she

herself

Table 2: Example of comparison set for gender bias

X: communism communism

communist

Y: socialism socialism

socialist

A: popular popular

rational

rational

B: unpopular unpopular

irrational

Table 3: Example of comparison set for social bias

IV. EXPERIMENTS AND DISCUSSION

a. Biased models

i. Gender bias

In investigating gender bias, an analysis was performed

based on three different subcategories, being

work/education, character and sexual orientation.

For work/education, there is a comparison between

engineering and humanities, doctor vs nurse, career vs

family, law vs medicine, author vs architect and cook vs

teacher. For the first four cases, there is a positive bias

which indicates man is more associated with the first

attribute. Specifically, engineering, doctor and career, all

of them with one exception show very strong bias as the

values are much larger than 1.

Interestingly, law is more associated with men than

medicine, perhaps owing this fact to medical professions

in general such as nurse, caretaker and others. A less

significant and not similar trend for all datasets trend is

shown in author vs architect which was also expected

since there is not a clear distinction also in reality. Cook

however is much more associated with women and

teacher with men. The cases where there is significant

bias averages absolute 1.35 (high), while the non-biased

case averages absolute 0.4 (small to medium).

Figure 4: Representation of gender bias

ii. Racial bias

For the racial bias category, the first six comparisons

have to do with skin color such as Europe vs Africa with

educated vs illiterate and African vs Asian with safe vs

dangerous. The words chosen are not color such as white

and black as these would contain noise from the actual

colors. When comparing Europe and Africa, we can see

that in most cases, bias is strong for safe and educated

towards the former. For Caucasian and Asian there is a

strong bias only in Wikipedia concerning rich vs poor.

African-American vs latino for lawful vs outlaw shows

small bias while African vs Asian for safe vs dangerous

shows negative strong or medium positive. Interestingly,

the word latino is found to be more associated to superior

than African- American in all datasets with strong

negative bias around -1.

The next sub-category, which is country, first compares

common stereotypes such as German vs Italian for

hardworking vs lazy. It is interesting again that the only

dataset that seems to be biased towards the stereotype is

neither the news nor social media but Wikipedia. The

same is true for the comparison of American vs Russian

and friend vs enemy. Very weak biases are found in

commonly rich countries when comparing between

them, such as Polish, Danish, Japanese, Korean, with

effect sizes less than 0.5.

Finally, citizenship shows illegal residents to be

significantly be associated with dangerous. There is no

clear however bias in immigrants and refugees though as

they are both moderately to low considered good and

bad.

Figure 5: Representation of racial bias

iii. Religious bias

When looking at religious bias, an effort was made again

to discover bias and compare known stereotypes with

combinations that probably also are not biased. The

biased results seem to come when comparing

Christianity vs Islam with peace vs violence and church

with mosque. A strong bias is found in news and

Wikipedia while twitter remains at non to small. In the

case of Sikhism vs Hinduism, only Wikipedia shows no

bias while the rest indicate that bad is associated with

Hinduism. Between protestant and orthodox another

comparison was made to compare with a non-biased as

expected comparison between rich and poor.

Figure 6: Representation of religious bias

iv. Social bias

In social bias, a split in comparison is first done for

different age groups. Old people are clearly strongly to

medium associated with slow and impolite.

Next, for social groups, quite unexpected, aristocrats

were not strongly associated with rich. Perhaps, there

was not much reference of some common adjectives.

Then, bankers and doctors were found to be rich and

educated correspondingly.

Finally, for political systems, democracy was connected

to happy, capitalism to unfair and libertarian popular in

Wikipedia.

Figure 7: Representation of social bias

iv. Non-people bias

Looking at non-people bias, interestingly we find that Books

are strongly associated with amusing. That can be explained

by the fact that perhaps when they are mentioned people

usually talk positively about them. For football and basketball

results were towards small with the exception of news that

favored football.

Figure 8: Representation of non-people bias

v. Inter-category comparison

Looking back at the overall results per category, it can be seen

that in the gender category, there are many comparisons that

are biased. That of course depends on the selected words but

when there is a bias it is very strong with values at around 1.5

or more. The other categories also show bias for specific

comparisons, however their strong values remain at a bit

lower levels of around 1 to 1.5 with fewer exceptions going

above 1.5.

It is also evident that many common stereotypes are

confirmed in most cases such as those about gender related

occupations, racial stereotypes such as white vs black and

illegals as well as social stereotypes about rich, poor and

happy. Those that had small effect size validate our expected

results and the validity of the method in finding biases, one

example being protestant vs orthodox.

As far as the datasets are compared, Twitter contains less

biases and Google News and Wikipedia seem to be higher.

This conclusion, however, takes into account the overall

picture since there are many observations that show otherwise

and only specific words are tested here, which of course

cannot generalize to the overall datasets.

b. De-biased model

When comparing the gender specific de-biased model with

the simple model some improvements were found. In some

cases, such as doctor vs nurse and author vs architect the

effect is reversed.

To split them down however, there was a reduction of 50%

or more in engineering vs humanities, doctor vs nurse, law vs

medicine, cook vs teacher and beautiful vs ugly.

In the other, it remained almost same or increased.

Totally, it decreased in 10 out of 13 cases.

Figure 9: Comparison of biased and de-biased models

c. Historical models

For the purpose of analyzing evolution of biases within our

sub-categories, pairs with high and prominent WEAT metric

values have been chosen, which are interesting in terms of

analysis in historical perspective.

i. Gender bias

In this sub-category, we focused our attention on two groups

of words. For the first one (Career vs Family / Man vs

Woman), we can clearly observe that the idea of men being

more inclined towards career aspirations rather than

dedicating most of their effort and time to family compared

to women has been held in the literature throughout the whole

period of time, as the comparative value fell below 1.0 only

in 1850, with small fluctuations, but general tendency to

increase slowly over time.

The second pair (Straight vs Gay / Right vs Wrong) shows an

unexpected general rapid growth from 1820 to 1920, and the

values stay high until the year 1950, after which it starts

decreasing slowly. A possible explanation is that a negative

attitude towards the homosexual males might had been

increased for political and/or economic reasons, and in

addition, the word “gay” was assigned its current meaning in

the middle of 20th century, which could also cause the rise of

comparative WEAT score. Later, it started declining,

possibly because of the changes in public outlook on different

sexual orientations. It should be mentioned that the results

can also be influenced by the fact that some of the words may

possess several meanings (as with the word “straight”).

Figure 10: Temporal changes in bias of gender-related

words

ii. Racial bias

Here, we looked at the groups and compared them in terms

of bias in educational level and wealth. First pair (Europe vs

Africa / Educated vs Illiterate) there is no strong trend

towards descent or ascent of mutual score: if interpolated, the

score would be insignificantly below zero, as most of the

values fall within the range [0.5; 0.5] with a majority of points

below a zero line. That means that there is no consistent

strong bias in English literature in a given corpus regarding

illiteracy of Africans compared to Europeans, although in

some cases values fall below –0.5 (years 1940 and 1960).

In the second case, we can easily distinguish the stereotype in

the literature that white people are generally more well-off

than Asian people, with some exceptions in the data (years

1820 and 1900), which might have happened due to lack of

close topics. Generally, the fluctuation of the plot reduces in

time, which is possibly explained by a higher confidence of

writers about higher level of life of Caucasian people.

Nevertheless, all the inferences which are being made here

are just our own hypotheses, and the true reason may differ

from the average trend for particular years and cases.

In addition, we analyzed the potential bias between two

nations (Americans vs Russians) being treated as friends or

enemies. Generally, we see a huge variation of values in the

positive region of Y-axis values which means that Russians

are not treated as friends compared to Americans in English

literature. For some certain points, the inclination may be

viewed as a result of global historical events, such as a Cold

War which might be a reason of dramatic increase of the

mutual score from 1950 to 1960. Nevertheless, there is no

common pattern that can be reproduced from this plot.

Figure 11: Temporal changes in bias of race-related words

Figure 12: Temporal changes in bias of nationality-related

words

iii. Religious bias

Our third group is bias in religion. First pair (Christianity vs

Islam) shows a slight bias towards Christianity being treated

as a more peaceful religion that Islam. For some years, the

score value reflects a high bias (in the 19th century), but it is

mitigated over time, staying close to 0.5

For the second pair (Protestant vs Orthodox / Rich vs Poor),

we see a very interesting pattern: until the middle of 20th

century, orthodox branch is more associated with the wealth

that protestant, but then the dramatic growth is occurring. It

can be understood along the lines of development of market

economy in first world countries, where many protestant

followers have been residing, as well as by Protestant work

ethic, which does not forbid and even promote trading in

some sense, unlike with orthodox morals.

Figure 13: Temporal changes in bias of religion-related

words

iv. Social bias

In the “Social bias” category, we will consider groups related

to age and political preferences.

The pair (Old vs Young / Polite vs Impolite) does not indicate

a specific bias, as most of the values fall into the range [-0.5;

0.5], where the bias does exist, but not so extreme to represent

high value for researchers. That being said, in the literature

sources stored in the database there was not a distinct

relationship between the age and the degree of politeness.

Similar tendency is observed with the second pair

(Libertarian vs Conservative / Popular vs Unpopular), where

the majority of values fluctuate around zero and do not

surpass 0.5 in absolute value. It means that none of mentioned

political views can be considered as being inclined towards

being popular in comparison to another based on the text

corpora used.

Figure 14: Temporal changes in bias of society-related

words

V. CONCLUSION AND FUTURE WORK

This work focused on analyzing biases in word embeddings

and performed a category based analysis using WEAT to

measure it. The results indicated that bias in present in many

popular datasets and sometimes strong bias too. Particularly

in cases with strong bias, it is evident in most datasets.

Gender bias was significant, followed by other, close together

such as racial, social and religious. The study went a step

head and split each category into different categories to

illustrate where exactly most bias is observed when talking

about gender and others. The same was then done when

comparing a biased model as well as specific cases of historic

development.

There are some limitations in our study such as the fact that a

limited amount of words could be chosen to represent each

set for comparison. For example, in specific occupations,

such as orthodox, the word could be specified by only a few

synonyms. Having a larger collection of appropriate words

would perhaps remove some of the noise introduced by

having less words. Conversely, some words may represent a

broad range of meanings which can create problems for the

approach to make sensible inferences (we observed such

examples during analysis of WEAT score temporal changes

for some words pairs, like “straight vs gay”).

Our current findings can be extended to be part of larger study

in the future that can attempt to develop a new method for

debiasing, in a way that it can eliminate bias in multiple

categories at the same time. Also, the thorough categorization

that was performed could be one of the ways to perform such

model testing. The historical analysis could give an

interesting outlook at how trends occur and where more

debiasing is needed for a model now or in the future

following a specific trend.

REFERENCES

[1] L. S. D. J. J. Z. Nikhil Garg, "Word embeddings

quantify 100 years of gender and ethnic stereotypes,"

in PNAS, 2018.

[2] A. M. Kaytlin Chaloner, "Measuring Gender Bias in

Word Embeddings across Domains and Discovering

New Gender Bias Word Categories".

[3] T. B. a. K.-W. C. a. J. Z. a. V. S. a. Adam, "Man is to

Computer Programmer as Woman is to Homemaker?

Debiasing Word Embeddings," 2016.

[4] D. B. Masahiro Kaneko, "Gender-preserving

Debiasing for Pre-trained Word Embeddings," in

Association for Computational Linguistics, 2019.

[5] https://code.google.com/archive/p/word2vec/,

"word2vec".

[6] R. S. C. D. M. Jeffrey Pennington, "GloVe: Global

Vectors for Word Representation," 2015.

[7] "Wikipedia," 1 5 2020. [Online]. Available:

https://en.wikipedia.org/wiki/Word2vec. [Accessed 5

5 2020].

[8] R. C. D. M. Jeffrey Pennington, "GloVe:

GlobalVectorsforWordRepresentation".

[9] "Gensim topic medelling for humans," 1 11 2019.

[Online]. Available:

https://radimrehurek.com/gensim/index.html.

[Accessed 1 5 2020].

[10] tolga-b, "GitHub," 2016. [Online]. Available:

https://github.com/tolga-b/debiaswe. [Accessed 1 5

2020].

[11] W. L. H. a. J. L. a. D. Jurafsky, "Diachronic Word

Embeddings Reveal Statistical Laws of Semantic

Change," in Association for Computational Linguistics

, 2016.

[12] williamleif, "histwords," 25 10 2015. [Online].

Available: https://github.com/williamleif/histwords.

[Accessed 1 5 2020].

[13] "SimplyPsychology," [Online]. Available:

https://www.simplypsychology.org/effect-size.html.

[Accessed 1 5 2020].

[14] "compare-embedding-bias," GitHub, 27 5 2019.

[Online]. Available:

https://github.com/hljames/compare-embedding-bias.

[Accessed 1 5 2020].

[15] M. L. Salvador, "Text analytics techniques in the

digital world: Word embeddings and bias," Irish

Communication Review, vol. 16, no. 1, 2018.