Version classiqueVersion mobile

Proceedings of the Seventh Italian Conference on Computational Linguistics CLiC-it 2020

 | 
Felice Dell'Orletta
, 
Johanna Monti
, 
Fabio Tamburini

Contributed Papers

Gender Bias in Italian Word Embeddings

Davide Biasion, Alessandro Fabris, Gianmaria Silvello et Gian Antonio Susto

Résumé

In this work we study gender bias in Italian word embeddings (WEs), evaluating whether they encode gender stereotypes studied in social psychology or present in the labor market. We find strong associations with gender in job-related WEs. Weaker gender stereotypes are present in other domains where grammatical gender plays a significant role.

Note de l’éditeur

Copyright 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).

Texte intégral

1. Introduction

1In the literature, the study of gender bias in word embeddings (WEs) is of interest for two main reasons: (i) WEs, as components of automatic decision systems (e.g. job search tools), may contribute to harm some user groups (De-Arteaga et al. 2019); (ii) WEs can be employed as a tool to measure the biases of text corpora (Garg et al. 2018) and systems for automatic text classification or information retrieval (Fabris et al. 2020). In both applications, it is important to isolate the gender-related information in a subspace (Bolukbasi et al. 2016) and subsequently (i) eliminate it via orthogonal projection or (ii) exploit it as a lens to study association of concepts with gender.

2A common taxonomy of bias in algorithms concentrates on the types of harm that they may cause (Barocas et al. 2017). Allocational harms happen when a limited resource (e.g. jobs) is assigned unfairly to subgroups of a population (e.g. women and men). Representational harms arise when groups or individuals are unable to determine their image, which is presented unfavourably or neglected. Autocomplete suggestions in search engines (Noble 2018; Olteanu, Diaz, and Kazai 2020) are a clear example of this situation. Query completion suggestions for “why are italian …” associate diverse concepts to the country and its inhabitants. Italians contribute very little to these results as they are unlikely to search information about themselves in English.

3Italian WEs have been developed (Berardi, Esuli, and Marcheggiani 2015; Bojanowski et al. 2017) and analyzed (Tripodi and Li Pira 2017), following seminal work in English; analysis of gender bias has unfortunately lagged behind. Our main contribution is to close this gap, by undertaking a systematic study of gender stereotypes in Italian WEs, adapting established approaches that assess gender bias in English WEs.

2. Related work

4Gender stereotypes are representational harms which influence the lives of women and men both descriptively and prescriptively, shaping the qualities, priorities and needs that members of each gender are expected to possess (Ellemers 2018). In seminal work, uncover problematic associations with gender in English WEs. Their approach to identify gender information is adapted to Italian in Section 3.1.1. study the stereotypical association of gender with dichotomies such as career and family, science and arts, following the Implicit Association Test (IAT - ).We recall their approach in Section 3.1.2. WEs of jobs have also been analyzed extensively due to their potential for allocational harms in resume search engines (De-Arteaga et al. 2019; Prost, Thain, and Bolukbasi 2019) and representational harms (Caliskan, Bryson, and Narayanan 2017), e.g. in general purpose search engines (Kay, Matuszek, and Munson 2015). Grammatical gender has been found to interact strongly with semantic gender in Spanish and German (McCurdy and Serbetci 2020), showing that the study of bias in gendered languages poses an additional challenge. We adapt these experiments to the Italian language, detailing our approach in Sections 3-5.

3. Gender in Italian WEs

3.1 Identifying gender information

3.1.1 Gender score

5To identify a vectorial subspace which encodes information about gender, we follow by building a list of gender definitional pairs: [lui (he), lei (she)], [uomo (man), donna (woman)], [padre (father), madre (mother)], [marito (husband), moglie (wife)], [fratello (brother), sorella (sister)], [maschio (male), femmina (female)].

6These pairs are built so that the second word denotes a female entity and the first word is, semantically, its male counterpart. Moreover, given we are interested in capturing semantic information about gender, while avoiding entanglement with grammatical gender, we ensure that the words in a pair do not derive from the same root via inflection. An example of pair discarded due to this criterion is [figlio (son), figlia (daughter)].

7Principal Component Analysis. We perform a Principal Component Analysis (PCA) on the six vector differences resulting from each gender definitional pair. The first eigenvalue dominates the remaining ones, with the first PC explaining 57% of variance. We normalize the first PC and consider it the main gender direction, denoted by gpca.

This is an established procedure to isolate the direction that captures most of the information about gender (Bolukbasi et al. 2016; Ethayarajh, Duvenaud, and Hirst 2019). In other words, by finding the direction that best fits the six vector differences (Image 10000000000000D60000001B7276D90CDE0E5612.jpg), we aim to obtain a direction that summarizes them.

Vector differences. To evaluate the robustness of this approach and highlight potential anomalies, we also consider each vector difference on its own, defining six unit length gender directions Image 10000000000000210000000EF5A35860F23D47DA.jpg:

Image 100000000000007C0000001BA8F7666DE7308D55.jpg

Image 10000000000000BA0000001B11978D40AA9C2A9B.jpg

Image 10000000000000AE0000001B62609BE93978D1C1.jpg

Image 10000000000000BE0000001BD3EAA7D8AEB61E4A.jpg

Image 10000000000000AE0000001B7A8BC7FDC6CD038B.jpg

Image 10000000000000D90000001B262392CCC9130849.jpg

8Gender score computation. [sec:gender-score] Given a word w, let us indicate with w its corresponding word vector. Let us consider any of the gender directions g defined above. We call gender score the normalized projection of w onto the direction g, defined as

Image 10000000000000B400000014679A280B086F7F9F.jpg              (1)

9This scalar captures associations of w along gendered lines. Informally, a highly positive value means that w is closer to the male terms of the pairs than to the female ones, while a strongly negative value entails the opposite.

3.1.2 WEAT

10The Implicit Association Test (IAT - ) is an assessment developed in cognitive psychology to measure subconscious associations between categories and concepts. It is commonly employed to assess implicit stereotypes in people. The Word Embedding Association Test (WEAT - ) is a technique inspired by the IAT to measure associations between concepts in WEs. Let X and Y be two equal-sized sets of target words and A and B two sets of attribute words, e.g., X=programmer, engineer, Y=nurse, teacher, A=man, male, B=woman, female. Let cos(ab) be the cosine similarity between the word vectors a and b. The differential association of a word w (taken from X or Y) with the attribute sets A and B is measured as

Image 10000000000001080000002D81D7A22D948D779D.jpg        (2)

11The normalized differential association between targets and attributes is defined as

Image 10000000000001600000002C969A0246D6C85032.jpg       (3)

12This is called effect size in statistics, and summarizes how different the quantity c(w, A, B) is, when evaluated on elements of target set X as opposed to target set Y. It is computed as a difference of means within each set, divided by overall standard deviation.

Gender score and WEAT. It is worth noting that, when |A|=|B|=1, WEAT is almost equivalent to the gender score defined in Section 3.1.1. Let Image 10000000000000470000001308E3429D7E71B9FB.jpgand Image 100000000000004700000013D479FD079D0EE3A3.jpgbe the sets of attribute words. Since we are using normalized vectors and the distributive property holds for the dot product, then

Image 100000000000011A0000002E23599F7122D47284.jpg        (4)

Handling grammatical gender

13Italian is a gendered language, wherein grammatical gender is assigned to all nouns. Within a sentence, each word is surrounded by other words of agreeing grammatical gender. This phenomenon, called grammatical gender agreement, in conjunction with the distributional hypothesis (Harris 1954), plays an important role when training WEs. Due to these properties, words that share the same grammatical gender to have similar vector representations. Accordingly, grammatical and semantic gender become entangled in WEs (McCurdy and Serbetci 2020; Gonen, Kementchedjhieva, and Goldberg 2019). As a consequence, when computing the gender score, we tend to obtain positive values for (grammatically) masculine terms and a negative score for feminine ones, making stereotypical association more noisy and harder to study.

14Mean gender score. To compute the gender score (Equation 1) for gendered words that have both a feminine and a masculine version, we propose the following approach. Let us indicate with wi and wm the feminine and masculine version of a gendered word w. We define their gender score as

Image 1000000000000108000000152657F4229FB2CC3B.jpg       (5)

15Averaging the masculine and feminine version with equal weights corresponds to giving both versions of the word the same importance. Different approaches, based for instance on word frequency, may be applicable in other contexts.

16Orthogonal projection. Some nouns cannot be inflected into the opposite grammatical gender, making the above approach impractical. An example is ufficio (office). In this context, we propose to mitigate the effect of grammatical gender by re-embedding every word through an orthogonal projection. We build a list of 138 inflected word pairs. Each pair consists of the feminine and masculine inflections of the same root, such as cara and caro (dear), which only differ in grammatical gender. We take the embedding of both words in a pair and compute their difference.

We perform PCA on these vector differences. The resulting PCs span a subspace U that contains most of the variance due to grammatical gender. To reduce the influence of grammatical gender, we re-embed vectors by projecting them on the orthogonal complement of U. In other words, given a word embedding w, let us call Image 100000000000004000000011788585F55774591D.jpgits orthogonal projection onto the “grammatical gender subspace” U. We propose re-emdedding every word vector w to

Image 10000000000000A10000002A1F3EE05422DC7926.jpg         (6)

17By means of this procedure, we obtain a new set of WEs. By construction, in this new embedding space, grammatical gender should have a lower influence on the geometry of word vectors.

4. Datasets and embeddings

  • 1 The authors provide no information about which Wikipedia dump they use.
  • 2 Common Crawl is a corpus of web pages, aimed at representing “a copy of the internet” at a given ti (...)

18To study gender bias we use WEs trained on two different datasets for the Italian language, both made available by FastText (Bojanowski et al. 2017; Grave et al. 2018). The first group of vector representations, which we refer to as wiki, consists of word vectors trained on a 2016 Wikipedia dump (Bojanowski et al. 2017).1 The second group of word vectors (labeled wiki-cc) was trained on the May 2017 Common Crawl2 and the Wikipedia dump from September 11, 2017 (Grave et al. 2018).

19We compare our results from the analyses on Italian WEs with results on their English counterpart. To this end, we also download two sets of FastText WEs trained on the English version of the same corpora, i.e. the English counterparts of wiki and wiki-cc. Given Wikipedia is a more curated source, we expect to find weaker stereotypes in wiki than in wiki-cc for both languages. As a pre-processing step we normalize every word vector to unit length.

  • 3 The detailed list of sources is available upon request.

20Census data about the labor market is required to analyse the correlation between the gender gap in professions and the gender score of the respective WEs. The statistics on the American occupation and gender representation are readily available (Census Bureau 2019). For their Italian counterpart, we retrieve statistics about occupation participation from several institutions, including professional chambers (Comitato Unitario Permanente degli Ordini e Collegi Professionali, Confprofessioni) and academic databases (AlmaLaurea).3

21Finally, in order to perform the Word Embedding Association Tests (WEAT), we need sets of target and attribute words in Italian. The sets of target words for the gender-science WEAT (Section 5.2) are derived from the Italian version of IAT;4 those for the gender-career WEAT (Section 5.3) were unavailable and have been translated by the authors of this work from the original IAT (Greenwald, McGhee, and Schwartz 1998).

5. Experiments

5.1 Occupations

This experiment investigates gender representation for different jobs in Italy and their association with gender-realted information in WEs, following studies on the English language (De-Arteaga et al. 2019; Garg et al. 2018; Prost, Thain, and Bolukbasi 2019). For each occupation, we compute its gender score using the different gender directions defined in Section 3.1.1, namely gpca and Image 10000000000000210000000EF5A35860F23D47DA.jpg, Image 100000000000005B0000001385FE516F9DE846F2.jpg. We calculate the plain gender score for the ungendered occupations (Equation 1) and the mean gender score for occupations characterized by grammatical gender (Equation 1).

22We compute Pearson’s correlation r between the gender scores and the percentage of women employed in each profession. The same analyses are carried out on English WEs, restricting them to the same set of occupations considered in Italian. Results are summarized in Table 1 and Figure 1, showing that Italian WEs consistently capture information about different gender representation in jobs. Informally, this mean that ordering jobs by percentage of women and by projection on a gender direction yields similar results. The right pane of Figure 1 demonstrates the significant effect of grammatical gender.

Table 1: Results of the Occupation analysis. Statistical significance is marked as * for p<0.1, ** for p<0.05 and *** for p<0.01.

wiki-cc

wiki

IT

r(p)

r(p)

gpca

-0.634 (1.3x10^{-4})***

-0.589 (4.9x10^{-4})***

Image 10000000000000230000000EA24836BA346BD7F7.jpg

-0.664 (4.7x10^{-5})***

-0.490 (5.1x10^{-3})***

Image 10000000000000220000000E79E0E726FAD16805.jpg

-0.594 (4.3 x 10^{-4})***

-0.528 (2.3 x 10^{-3})***

Image 10000000000000220000000E28F73DB5299DF0AF.jpg

-0.575 (7.1 x 10^{-4})***

-0.537 (1.8 x 10^{-3})***

Image 10000000000000220000000EC2BF8A9AC346F08C.jpg

-0.401 (2.5 x 10^{-2})**

-0.160 (3.9 x 10^{-1})

Image 10000000000000230000000E93A30A3EF232A494.jpg

-0.658 (5.7 x 10^{-5})***

-0.599 (3.8 x 10^{-4})***

Image 10000000000000220000000E57776D45314BD3AF.jpg

-0.358 (4.8 x 10^{-2})**

-0.205 (2.7 x 10^{-1})

EN

r(p)

r(p)

gpca

-0.830 (2.0 x 10^{-6})***

-0.707 (2.3 x 10-4)***

Figure 1

Image 10000000000003B000000187F4D10A527D1E1974.jpg

(Left) Gender score of occupations gdiff0 , on the y axis, vs percentage of women in that occupation, represented on the x axis. (Right) Same variables, including both feminine and masculine versions of gendered jobs. Translations, omitted for readability, can be found in the left pane

5.2 Science and Arts

23In this WEAT, the sets of target words for Science and Arts, taken from the Italian version of the IAT, are: X = {biologia (biology), fisica (physics), chimica (chemistry), matematica (mathematics), geologia (geology), astronomia (astronomy), ingegneria (engineering)}, Y = {filosofia (philosophy), umanesimo (humanism), arte (arts), letteratura (literature), italiano (italian), musica (music), storia (history)}. The sets of male and female attributes are taken from the gender definitional pairs (Section 3.1.1): A = {lui, uomo, padre, marito, fratello, maschio\}, = {lei, donna, madre, moglie, sorella, femmina}. We compute the effect size d and the p-value using the whole attribute sets A and B, and label this analysis “all”. Moreover, we also perform the WEAT test over single word pairs, e.g. A= {lui}, B= {lei}. Results are reported in Table 2. We find no stereotypical association in the expected direction. We hypothesize that this is due to the feminine grammatical gender of all science-related target words, deferring a more detailed analysis to Section 5.4.

Table 2: Results of the Science and Arts WEAT. Statistical significance is marked as * for p<0.1, ** for p<0.05 and *** for p<0.01

wiki-cc

wiki

IT

d(p)

d(p)

all

-0.172 (6.3 x 101)

-0.140 (5.9 x 101)

Image 10000000000000230000000EA24836BA346BD7F7.jpg

-0.464 (7.9 x 101)

-0.396 (7.5 x 101)

Image 10000000000000220000000E79E0E726FAD16805.jpg

-0.016 (5.1 x 101)

-0.064 (5.4 x 101)

Image 10000000000000220000000E28F73DB5299DF0AF.jpg

-0.408 (7.5 x 101)

-0.152 (6.1 x 101)

Image 10000000000000220000000EC2BF8A9AC346F08C.jpg

-0.002 (5.0 x 101)

0.271 (3.3 x 101)

Image 10000000000000230000000E93A30A3EF232A494.jpg

-0.127 (6.0 x 101)

-0.174 (6.2 x 101)

Image 10000000000000220000000E57776D45314BD3AF.jpg

-0.144 (6.1 x 101)

-0.195 (6.3 x 101)

EN

d(p)

d(p)

all

1.420 (1.5 x 103)***

1.304 (3.2 x 103)***

5.3 Career and Family

24In essence, the Career and Family WEAT is very similar to the Science and Arts WEAT; the only difference is in the sets of target words. The target sets are translated into Italian from the original English IAT as follows: X = {esecutivo (executive), management (management), professionale (professional), azienda (corporation), stipendio (salary), ufficio (office)}, Y = {casa (home), genitori (parents), bambini (children), famiglia (family), cugini (cousins), matrimonio (marriage), nozze (wedding), parenti (relatives)}. Results are summarized in Table 3. Stereotypical associations for wiki-cc WEs are present but weak, whereas they are more significant for wiki.

Table 3: Results of the Career and Family WEAT. Statistical significance is marked as * for p<0.1, ** for p<0.05 and *** for p<0.01.

IT

d(p)

d(p)

all

0.838 (5.3 x 10-2)*

1.351 (2.7 x 10-3)***

Image 10000000000000230000000EA24836BA346BD7F7.jpg

0.457 (2.0 x 10-1)

1.172 (8.7 x 10-3)***

Image 10000000000000220000000E79E0E726FAD16805.jpg

1.265 (4.7 x 10-3)***

1.512 (5.4 x 10-4)***

Image 10000000000000220000000E28F73DB5299DF0AF.jpg

0.614 (1.2 x 101)

1.181 (8.3 x 103)***

Image 10000000000000220000000EC2BF8A9AC346F08C.jpg

0.299 (2.9 x 101)

0.876 (4.4 x 102)**

Image 10000000000000230000000E93A30A3EF232A494.jpg

0.952 (3.2 x 10-2)**

0.898 (4.2 x 102)**

Image 10000000000000220000000E57776D45314BD3AF.jpg

0.713 (8.7 x 102)*

-0.566 (8.6 x 101)

EN

d(p)

d(p)

all

1.879 (0.0 x 100)***

1.568 (2.3 x 10-4)***

5.4 Mitigating the effect of grammatical gender

25In this section we quantify the extent to which the semantic gender information (Section 3.1) is influenced by grammatical gender, and test one approach designed to mitigate its influence (Section 3.2).

26The dataset used in the experiment about job-related WEs (Section 5.1) is suitable for this analysis, as it consists of words which have (i) a semantic association with gender, as measured objectively by the percentage of women in each profession and (ii) a grammatical association with gender, as half of those words admit a feminine and a masculine version.

  • 5 For the sake of brevity, we concentrate on gpca; the remaining gender directions (...)

We measure the relative strength of semantic and grammatical associations in the proposed gender directions as follows. Let us denote by Image 1000000000000012000000130E51D27F0337FE83.jpg, the set of job-related words which admit a feminine and a masculine version and by Image 10000000000000B30000001456989DE9E6023423.jpgthe difference in their gender scores.5 We compute the average influence of grammatical gender on direction \mathbf{g} (based on set Image 1000000000000012000000130E51D27F0337FE83.jpg) as

Image 10000000000000930000003168C3462057054073.jpg      (7)

27Visually, this corresponds to the average (signed) length of the vertical lines, in the right pane of Figure 1, connecting the feminine and masculine version of a job-related word.

Moreover, let us denote by Image 100000000000000C0000000DAA720472CF730333.jpgthe complete set of job-related words wj and by xj the percentage of women in job wj. Let us indicate with Image 100000000000003100000014013B1E4D78C263D5.jpgthe respective gender score, computed according to Equation 1 or 5, depending on whether wj admits different masculine and feminine inflections. We define Image 100000000000004200000013A008EBCB2FEF6E46.jpg(Image 100000000000003E0000001347E231296E79F611.jpg) as the maximum (minimum) percentage of women in a job from set Image 100000000000000C0000000DAA720472CF730333.jpg. Furthermore, let us call m the angular coefficient computed by (linearly) regressing Image 100000000000003100000014013B1E4D78C263D5.jpgonto xj over set Image 100000000000000C0000000DAA720472CF730333.jpg. We compute the full-scale influence of semantic gender on direction g (based on set Image 100000000000000C0000000DAA720472CF730333.jpg) as

Image 10000000000000F20000001356CAB281BA667693.jpg       (8)

Visually, this corresponds to the vertical component of the blue regression line in Figure 1, clipped between Image 100000000000003E0000001347E231296E79F611.jpgand Image 100000000000004200000013A008EBCB2FEF6E46.jpg.

28Finally, we compute the relative strength of semantic and grammatical associations in the proposed gender direction as the ratio

Image 100000000000003F0000002886DCFA7C9BCC0B58.jpg      (9)

  • 6 In this experiment, the grammatical gender subspace U is spanned by the first PC.

The first three rows of Table 4 report Image 1000000000000016000000132818DE00FF71F023.jpg, Image 10000000000000150000001036B75D7A35E28CBE.jpgand k for wiki-cc (first column) on the job dataset described in Section 4. The second column concentrates on a set of word embeddings derived from wiki-cc by removing information about grammatical gender from every word, via Equation 6.6 We label this new set of word embbeddings Image 100000000000004C000000103C00FA21266C9649.jpg. In going from wiki-cc to Image 100000000000004C000000103C00FA21266C9649.jpg, Image 1000000000000016000000132818DE00FF71F023.jpgis reduced by over 40% while Image 10000000000000150000001036B75D7A35E28CBE.jpgdecreases by less than 10%. This indicates that the orthogonal projection procedure reduces the influence of grammatical gender while retaining semantic information which is present in the original version of the WEs, hence the value of k decreases.

The final three rows of Table 4 report summary statistics for stereotypical associations described in Sections 5.1-5.3. Interestingly, the significance of each association is larger for Image 100000000000004C000000103C00FA21266C9649.jpgthan for wiki-cc. In particular, the effect size for the Science-Arts WEAT becomes positive, in accordance with the stereotype. We interpret these results as evidence for the hypothesis that grammatical gender confounds and outweighs stereotypical associations in Italian WEs, in line with prior work on gendered languages (McCurdy and Serbetci 2020).

Table 4

wiki-cc

Image 100000000000004C000000103C00FA21266C9649.jpg

Occupations (gpca)

Image 1000000000000016000000132818DE00FF71F023.jpg

0.41

0.22

Image 10000000000000150000001036B75D7A35E28CBE.jpg

0.23

0.20

k

1.79

1.09

r (p)

-0.63 (1.3 x 104)***

-0.68 (2.2 x 105)***

Science & Arts (all)

d (p)

-0.17 (6.3 x 101)

0.73 (9.7 x 102)*

Career & Family (all)

d (p)

0.84 (5.3 x 102)*

1.21 (6.1 x 103)***

Importance of semantic and grammatical gender before (wiki-cc) and after (Image 100000000000004C000000103C00FA21266C9649.jpg) projecting WEs onto the orthogonal complement of the grammatical gender subspace (Equation 6). Where applicable, statistical significance is marked as * for p<0.1, ** for p<0.05 and *** for p<0.01

6. Discussion

29We successfully replicated prior analyses about gender-stereotypical associations in English WEs, finding them to be consistently stronger when computed on WEs trained on a weakly curated corpus. To the best of our knowledge, this is a novel result.

30For Italian WEs, the picture is more nuanced and tied to grammatical gender. WEs for occupations, which are ungendered or admit a dual form, are robustly associated with gender along a stereotypical direction. Compared against the other stereotypes analysed in this work, this is the strongest association, confirming results from prior work on English WEs (Fabris et al. 2020). In the Science-Arts WEAT, science-related words are all feminine nouns, departing from the expected stereotypical association. Semantic associations with gender are outweighed by grammatical gender in this WEAT, in accordance with prior work on gendered languages (McCurdy and Serbetci 2020). Our analysis in Section 5.4 demonstrates the importance of grammatical gender in Italian. On the other hand, the Career-Family WEAT features a more balanced distribution of grammatical gender, resulting in a differential association which is in line with gender stereotypes, especially for wiki, less so for wiki-cc.

In Italian WEs, we find that wiki embeddings contain stronger stereotypical associations than wiki-cc embeddings for the Career-Family WEAT. This disconfirms our hypothesis that WEs trained on a less curated corpus (wiki-cc) would encode stereotypes more strongly. Finally, we find no consistent property connected to specific gender directions Image 10000000000000210000000EF5A35860F23D47DA.jpg. Across different corpora and stereotypes, the aggregated analyses (labelled “all” and gpca) provide a reasonable summary of the stereotypical associations encoded in the single gender directions Image 10000000000000210000000EF5A35860F23D47DA.jpg.

7. Conlusion

31Overall, we have analyzed gender bias in Italian WEs, adapting existing techniques and gathering data where required. We looked for stereotypical associations with gender-imbalanced professions, Career and Family, Science and Arts, finding significant associations in 2 out of 3. As expected from prior work (Gonen, Kementchedjhieva, and Goldberg 2019; McCurdy and Serbetci 2020), grammatical gender is a strong confounder in these analyses.

32We draw the following preliminary conclusions: (i) Italian WEs seem to have less potential than their English counterparts to systematically reinforce the tested gender stereotypes, mostly due to grammatical gender. However, (ii) the influence of grammatical gender on WEs may cause different harms. As an example, in the context of job search, masculine is likely to be the default choice for queries of recruiters (male as norm - e.g. “psicologo” [psychologist]). Those queries would likely be closer to male candidates’ CVs than equivalent female ones, in some embedded text representations, potentially putting women at a systematic disadvantage. Both points above require further analysis of text retrieval/classification systems based on Italian WEs. Finally, (iii) isolating stereotypical concepts and gendered associations in Italian WEs along a single direction is challenging. The tested WEs show little promise as a reliable measurement tool for gender-stereotypical associations, unless combined with approaches to mitigate the influence of grammatical gender.

Bibliographie

Solon Barocas, Kate Crawford, Aaron Shapiro, and Hanna Wallach. 2017. “The Problem with Bias: Allocative Versus Representational Harms in Machine Learning.” In 9th Annual Conference of the Special Interest Group for Computing, Information and Society.

Giacomo Berardi, Andrea Esuli, and Diego Marcheggiani. 2015. “Word Embeddings Go to Italy: A Comparison of Models and Training Datasets.” In IIR.

Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. “Enriching Word Vectors with Subword Information.” Transactions of the Association for Computational Linguistics 5: 135–46.

Tolga Bolukbasi, Kai-Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam T. Kalai. 2016. “Man Is to Computer Programmer as Woman Is to Homemaker? Debiasing Word Embeddings.” In Advances in Neural Information Processing Systems, 4349–57.

Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. “Semantics Derived Automatically from Language Corpora Contain Human-Like Biases.” Science 356 (6334): 183–86.

Census Bureau. 2019. “Current Population Survey: U.S. Bureau of Labor Statistics.” https://www.bls.gov/cps/cpsa2019.pdf.

Maria De-Arteaga, Alexey Romanov, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnaram Kenthapadi, and Adam Tauman Kalai. 2019. “Bias in Bios.” Proceedings of the Conference on Fairness, Accountability, and Transparency - FAT* ’19.

Naomi Ellemers. 2018. “Gender Stereotypes.” Annual Review of Psychology 69 (1): 275–98.

Kawin Ethayarajh, David Duvenaud, and Graeme Hirst. 2019. “Understanding Undesirable Word Embedding Associations.” In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 1696–1705. Florence, Italy: Association for Computational Linguistics.

Alessandro Fabris, Alberto Purpura, Gianmaria Silvello, and Gian Antonio Susto. 2020. “Gender Stereotype Reinforcement: Measuring the Gender Bias Conveyed by Ranking Algorithms.” Information Processing & Management 57 (6): 102377.

Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. “Word Embeddings Quantify 100 Years of Gender and Ethnic Stereotypes.” Proceedings of the National Academy of Sciences 115 (16): E3635–E3644.

Hila Gonen, Yova Kementchedjhieva, and Yoav Goldberg. 2019. “How Does Grammatical Gender Affect Noun Representations in Gender-Marking Languages?” In Proceedings of the 23rd Conference on Computational Natural Language Learning (Conll), 463–71.

Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. 2018. “Learning Word Vectors for 157 Languages.” In Proceedings of the International Conference on Language Resources and Evaluation (Lrec 2018).

Anthony G. Greenwald, Debbie E. McGhee, and Jordan L. K. Schwartz. 1998. “Measuring Individual Differences in Implicit Cognition: The Implicit Association Test.” Journal of Personality and Social Psychology 74 (6): 1464–80.

Zellig S. Harris. 1954. “Distributional Structure.” Word 10 (2-3): 146–62.

Matthew Kay, Cynthia Matuszek, and Sean A Munson. 2015. “Unequal Representation and Gender Stereotypes in Image Search Results for Occupations.” In Proceedings of the 33rd Annual Acm Conference on Human Factors in Computing Systems, 3819–28. ACM.

Katherine McCurdy, and Oguz Serbetci. 2020. “Grammatical Gender Associations Outweigh Topical Gender Bias in Crosslinguistic Word Embeddings.” arXiv Preprint arXiv:2005.08864.

Safiya Umoja Noble. 2018. Algorithms of Oppression: How Search Engines Reinforce Racism. nyu Press.

Alexandra Olteanu, Fernando Diaz, and Gabriella Kazai. 2020. “When Are Search Completion Suggestions Problematic?” In Computer Supported Collaborative Work and Social Computing (Cscw). ACM.

Flavien Prost, Nithum Thain, and Tolga Bolukbasi. 2019. “Debiasing Embeddings for Reduced Gender Bias in Text Classification.” Proceedings of the First Workshop on Gender Bias in Natural Language Processing.

Rocco Tripodi, and Stefano Li Pira. 2017. “Analysis of Italian Word Embeddings.” In CLiC-It.

Notes

1 The authors provide no information about which Wikipedia dump they use.

2 Common Crawl is a corpus of web pages, aimed at representing “a copy of the internet” at a given time. The authors train WEs on pages written in Italian, exploiting language identification as preliminary step for their pipeline.

3 The detailed list of sources is available upon request.

4 https://implicit.harvard.edu/implicit/italy/takeatest.html

5 For the sake of brevity, we concentrate on gpca; the remaining gender directions ) yield similar results.

6 In this experiment, the grammatical gender subspace U is spanned by the first PC.

Auteurs

Università degli Studi di Padova – biasiondav@dei.unipd.it

Università degli Studi di Padova – fabrisal@dei.unipd.it

Università degli Studi di Padova – sustogia@dei.unipd.it

© Accademia University Press, 2020

Conditions d’utilisation : http://www.openedition.org/6540

Lire

Open access

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search