Towards Lexical Gender Inference: A Scalable Methodology using Online Databases

This paper presents a new method for automatically detecting words with lexical gender in large-scale language datasets. Currently, the evaluation of gender bias in natural language processing relies on manually compiled lexicons of gendered expressions, such as pronouns ('he', 'she', etc.) and nouns with lexical gender ('mother', 'boyfriend', 'policewoman', etc.). However, manual compilation of such lists can lead to static information if they are not periodically updated and often involve value judgments by individual annotators and researchers. Moreover, terms not included in the list fall out of the range of analysis. To address these issues, we devised a scalable, dictionary-based method to automatically detect lexical gender that can provide a dynamic, up-to-date analysis with high coverage. Our approach reaches over 80% accuracy in determining the lexical gender of nouns retrieved randomly from a Wikipedia sample and when testing on a list of gendered words used in previous research.


Introduction
There is a growing body of research on gender bias embedded in trained language models as well as on allocational and representational harms caused by the deployment of these models.There have moreover been increasing calls for early and thorough data description and curation in order to gain insights into how, for instance, gender stereotyping or quality of service bias is propagated from data into a language model.What both of these strands of research have in common is their reliance on pre-defined lexicons of terms related to gender.
In English, gendered words most commonly include pronouns (he, she, they, etc.), and also words that carry lexical gender, such as boyfriend, policewoman, or prince.Previous works on gender bias in language technologies often use manually compiled lists of words carrying lexical gen-der to, for example, mitigate gender stereotyping through data augmentation (Lu et al., 2020), assess trans-exclusionary bias in co-reference annotations (Cao and Daumé III, 2020) or evaluate gender inequalities in Wikipedia article titles (Falenska and Çetinoglu, 2021).However, curated lists are limited in their coverage of terms that contain lexical gender and can become outdated if not maintained.
To address this issue, we present a scalable algorithmic method to determine lexical gender by querying a word's dictionary definitions for a small subset of definitively gendered words.Our method allows for high-coverage, instantaneous detection of words carrying lexical gender, which eliminates the need to manually compile and maintain static lists of gendered words.This not only facilitates the extension of previous work on gender bias in NLP, but can also be used for a more detailed analysis on the representation of gender in large-scale language datasets used to train large language models like BERT (Devlin et al., 2018) or GPT-2 (Radford et al., 2019).
By combining the gender labels obtained from Merriam Webster Online (Merriam-Webster, 2022), WordNet ® (Princeton University, 2010) and Dictionary.com (Dictionary.com, LLC, 2022), our method reaches an accuracy of 84% in determining the lexical gender of words in a random sample of 1,000 Wikipedia articles and 87% accuracy on a list of words carrying lexical gender adapted from previous research.The code for the algorithm, evaluation methods and datasets are available 1 .
In the following section we first outline the conceptions of linguistic gender used in this research and secondly present an overview of research on gender in language technology that relies on curated lists of gendered words.Thirdly, we discuss prior approaches to algorithmic gender inference.Section 3 gives a detailed overview of the algo-rithm and Section 4 introduces the datasets used to assess our gender detection algorithm.We present quantitative and qualitative results in Section 5 and discuss limitations as well as avenues for future development.

Background
When dealing with the category of gender in language technology, it is important to make a distinction between the social category of gender and gender in a linguistic sense.While social gender relates to the complex property, performance and experience of one's own and others' gender within society (Ackerman, 2019), linguistic gender describes the expression of gender within grammar and language.In English, linguistic gender mainly encompasses ways to express gender as female, male or gender-indefinite (Fuertes-Olivera, 2007).Social gender, as an extra-linguistic category, includes a more fluid view of gender aside from male and female categories.This includes transgender, genderqueer and other non-binary experiences and expressions of gender (Darwin, 2017).As Bucholtz (1999) and Cao and Daumé III (2020) point out, there is no "one-to-one" mapping between social and linguistic gender.However, the two are influenced by each other: on one hand, expressions of gender in language are subject to changing norms in society (Fuertes-Olivera, 2007), on the other hand, the way gender is represented in language influences the conception of gender within society (Butler, 1990).Thus, being able to evaluate gendered expressions in language provides insights into societal conceptualisations of gender.
Since this research explicitly focuses on lexical gender in English, which is a linguistic category, we give an overview of linguistic gender in English in Section 2.1.Section 2.2 explores the role lexical gender information plays in different areas of research on gender bias in NLP, which simultaneously present possible areas of application for our method of lexical gender inference.Section 2.3 discusses two prior algorithmic systems for lexical gender inference in English.

Linguistic gender in English
The taxonomy of linguistic gender in this work builds upon the approach developed by Cao and Daumé III (2020) and incorporates work by Corbett (1991), Hellinger and Bussmann (2003) and Fuertes-Olivera (2007).
Within linguistic gender, Cao and Daumé III (2020) differentiate between grammatical, referential, and lexical gender.Grammatical gender refers to the distinction of noun classes based on agreement between nouns and their dependants.English, as a natural or notional gender language (McConnell-Ginet, 2013), does not have grammatical gender, but it has referential and lexical gender.Referential gender is used to refer to the social gender of a specified extra-linguistic entity.Thus, it "relates linguistic expressions to extralinguistic reality, typically identifying referents as 'female', 'male', or 'gender-indefinite.' " (Cao and Daumé III, 2020).In English, pronouns fall under the category of referential gender.Lexical gender, which this work focuses on, is non-referential but a semantic property of a given linguistic unit, which can be either masculine, feminine2 or genderindefinite/gender-neutral.Ackerman (2019) calls these words "definitionally gendered".Words that carry lexical gender can require semantic agreement in related forms, such as, for instance, using the pronoun his in connection with the word stuntman in the sentence 'Every stuntman needs to rehearse his stunts.'(Fuertes-Olivera, 2007).In English, lexical gender is usually not morphologically marked.Exceptions to this rule include e.g. the suffixes -man to denote masculine gender, such as in policeman, or -ess to denote feminine gender, such as in waitress.It should moreover be noted that lexical gender is exclusively a linguistic property.However, words containing lexical gender can be used to express referential gender if a concrete referent is specified (Cao and Daumé III, 2020).

Lexical gender in gender bias research
The evaluation and mitigation of gender biases in language datasets and models relies on referential expressions of gender, such as pronouns, but also words that carry lexical gender.These pieces of research vary in application, as well as the number of gendered expressions considered, which varies from two to around 120 words.Most works assess binary differences between male and female gender.However, an emergent strand of NLP research also focuses on non-binary gender expressions (Cao and Daumé III, 2020) and creating gender-neutral datasets and systems (Vanmassenhove et al., 2021).The following considers example use-cases of lexi-cons of lexically gendered words.These simultaneously represent a variety of applications for our lexical gender detection algorithm.

Dataset evaluation
The most straightforward form of using gendered words is to assess the distribution of gendered words in a corpus.Zhao et al. (2019) counted he/she pronouns in the One Billion Word Benchmark (Chelba et al., 2013) to show male skew in the training data for the ELMo language model (Peters et al., 2018), which is the primary focus of their analysis.This analysis addressed calls for better data evaluation (Bender et al., 2021;Rogers, 2021) prior to or alongside model bias analyses.
Retrieval for analysis Limited-scope lists of words that carry lexical gender were used by Caliskan et al. (2017) to retrieve Word2Vec embeddings (Mikolov et al., 2013) and perform the Word Embedding Association Test (WEAT).This test measured stereotyping by calculating implicit associations between eight male/female word pairs and words related to maths or science and arts.Guo and Caliskan (2021) used an adapted version of the WEAT, the CEAT, to asses intersectional biases in contextualized word embeddings (ELMo (Peters et al., 2018), BERT (Devlin et al., 2018), OpenAI GPT (Radford et al., 2019;Brown et al., 2020)).Another use-case in which gendered words were used for retrieval is research by Falenska and Çetinoglu (2021), who assessed gender bias in Wikipedia articles.As a first step, they filtered the article titles for a limited number of words that carry lexical gender.

Creation of synthetic evaluation data
In sentence-based analyses of gender-bias, lists of words with lexical gender can also be used to fill placeholders in sentence templates and thus create synthetic sentences with different gendered entities.For example, Kiritchenko and Mohammad (2018) created the Equity Evaluation Corpus (EEC) to analyse gender stereotyping in sentiment analysis systems.The EEC inspired the creation of the Bias Evaluation Corpus with Professions (BEC-Pro), which was used to analyse associations between gendered entities and professions in BERT (Bartl et al., 2020).Similarly, Sheng et al. (2019) used the word pair the man/the woman as fillers within sentence-start prompts for open-ended natural language generation (NLG) and the subsequent analysis of gender biases in the generated sentences.
In a rare instance of research on non-binary representations of gender in NLP, Cao and Daumé III (2020) used gendered lists of words to find and hide lexical gender in the GAP dataset (Webster et al., 2018).The dataset created in this way was used to measure gender-and trans-exclusionary biases in coreference resolution performed by both humans and machine-learning models.
Data manipulation Extensive lists of gendered words were used in the context of Counterfactual Data Augmentation (CDA), which replaces words with masculine lexical gender with their feminine variants and vice versa in a corpus.This is done in order to create training or fine-tuning data for gender bias mitigation.For instance, Lu et al. (2020) 'hand-picked' gender pairs to swap in CDA and Maudslay et al. (2019) added first names to the list of words to be swapped.
Another kind of data manipulation, this time aiming for neutral gender, was performed by Vanmassenhove et al. (2021).They used lists of unnecessarily gendered job titles (e.g.mailman/mailwoman) and feminine forms (e.g.actress), as well as generic uses of the suffix -man (such as in freshman) in the extended version of their Neutral Rewriter, which re-writes explicit mentions of gender into their gender-neutral variants (mail carrier, actor, first-year student).

Lexical gender inference
Previous approaches to automatic lexical gender inference used unsupervised and semi-supervised learning, drawing on the presence of gendered pronouns in the context of a given noun (Bergsma and Lin, 2006;Bergsma et al., 2009).While Bergsma and Lin (2006) created a large dataset of probabilistic noun gender labels, Bergsma et al. (2009) used these as basis for creating training examples for a statistical model that uses context and morphological features to infer lexical gender.
One major point of criticism here lies in the probabilistic determination of noun gender, which has the risk of mislabelling lexically neutral nouns, such as professions, as being gendered due to contextual distributions that are representative of stereotypes or the number of men and women holding the profession instead of the linguistic category of lexical gender.For example, since there are more female than male nurses (Bureau of Labor Statistics (BLS), 2022) and thus most nurses are referred to with female pronouns in text, the algo-rithm might infer that the term nurse has female lexical gender, when in fact it is neutral.

Method: Automatic Detection of Lexical Gender
The main goal of this work is to produce a dynamic, high coverage, scalable method to determine the lexical gender of a target word in order to replace previously used manually compiled lexicons.For this purpose, we leveraged the fact that the definition of a lexically gendered word includes words from a small set of definitively gendered words that carry the same lexical gender.In the following, we describe the main algorithm setup, additional parameters and heuristics, as well as the method to combine lexical gender labels from different databases.A schematic, exemplary overview of the algorithm is presented in Figure 1.

Algorithm construction
The method we outline utilises the increasing availability of machine-readable dictionaries, such as Merriam Webster Online, Dictionary.com, and the lexical database WordNet, in order to identify gendered terms.Examples ( 1) and ( 2) illustrate how lexical gender is captured within Merriam-Webster's (2022) definitions of nun and monk: (1) nun: a woman belonging to a religious order (2) monk: a man who is a member of a religious order and lives in a monastery Both definitions mention the lexical gender of the referent through a gendered word, in this case man and woman.Initial analyses showed that gendered words are more likely to occur at the beginning of a definition and definitions often used the words female/male or woman/man to specify lexical gender.In identifying gendered terms, we thus considered the presence and amount of up to eight definitively gendered words, such as male/female, man/woman etc., in the target word's definitions to draw inferences about its lexical gender.
For retrieval of the definitions, we accessed WordNet through the Natural Language Toolkit (NLTK, Bird et al., 2009) and Merriam Webster Online as well as Dictionary.comthrough HTTP requests.
Once the definitions for a given target word were retrieved, the process of obtaining lexical gender was the same for either dictionary.We determined whether a word has masculine, feminine, or neutral lexical gender by counting occurrences of a number of word pairs which have clearly defined feminine or masculine lexical gender, which are displayed in Table 1.If the combined definition texts contain more masculine than feminine terms, the word was labelled with masculine lexical gender, and vice versa.If the same number of masculine and feminine words was found within a set of definitions, which includes the case in which none of the predefined gendered terms can be found, the word was labelled with neutral lexical gender.We additionally obtained a combined label through a majority vote over the individual dictionaries' gender labels.In cases in which words could not be found in one dictionary and querying each of the other dictionaries returned different labels, a neutral gender label was assigned.

Parameters
Three variable parameters were used to limit the number of definitions and word tokens queried, as well as the number of definitively gendered words to use for the query.In order to determine the best combination of values for our parameters, we performed a grid search using our gold standard data (see Section 4.1) and combined labels to test performance.

Number of definitions d
We limited the number of definitions, because definitions that occur early on have a higher likelihood of describing a more general sense of the word, while later definitions relate to very specific word senses.Therefore, we retrieved only the first d definitions that the dictionary lists for the word.During grid search, we tested integer values in the range d = [2..10], and the best value was determined to be d = 4.

Number of tokens t
We also experimented with limiting the number of tokens within a given definition to see whether definitively gendered terms were more likely to be mentioned earlier in a given definition.The definitions were tokenized using NLTK (Bird et al., 2009).We took the first t tokens of each definition.Regarding the number of tokens in a definition, we tested the algorithm with t = {5, 10, 15, 20, 25, 30, 35} in our experiments and found t = 20 to produce optimal results.

Number of gendered word pairs w
The word pairs used during experiments are listed in Table 1.The first two word pairs, woman/man and fe-  male/male, as well as the pair girl/boy, are most commonly used to describe the gender of a person or animal, while the rest of the words describes gendered family relations.The latter were chosen in order to account for cases in which the lexical gender of a person is described in relation to another person by using family terms.This is for example the case in the definition of baroness in Merriam Webster: "the wife or widow of a baron" (Merriam-Webster, 2022).The grid search was performed for integer values in the range w = [2..8] and best performance was obtained for w = 5 word pairs.Moreover, if a target word was included in the definitively gendered pairs or their plural forms, it was automatically classified with the respective lexical gender.

Morphological Heuristics
Aside from the lexical database method described above, we additionally applied heuristics relating to suffix-morphology and punctuation.Morphological heuristics were applied before querying the dictionaries, while the punctuation-related heuris-tic was applied when a word could not be found in a dictionary.The first heuristic was applied in order to handle gender-neutral definitions of words that carry gender-explicit markers, such as the word businessman, which carries the masculine suffix -man.Its definition in WordNet (Princeton University, 2010) is shown in (3).
(3) businessman: a person engaged in commercial or industrial business (especially an owner or executive) Even though businessman contains a masculine suffix, its definition is generic, most likely due to the fact that businessman was once used for business people of all genders.However, since feminine or neutral equivalents (business woman, business person) are widely used nowadays, the word businessman has become gender specific and defining it generically represents an outdated, maleas-norm viewpoint (Fuertes-Olivera, 2007).We thus classified words containing the suffixes -man and -boy or -woman and -girl into masculine  Note: for Wiki1000 full, combined predicted labels were used, because no gold labels exist for this dataset and feminine lexical gender, respectively.Regular expressions were used to ensure that feminine or neutral words ending in -man such as woman or human, as well as words that have the suffix -woman, were not classified as masculine.
Another heuristic was applied in order to account for spellings that differ in punctuation, e.g.grandfather vs. grand-father.We check for and subsequently remove punctuation within a word if it cannot be found within a dictionary.This also applies to the cases in which non-detection is caused by a whitespace character.

Data
We used two test datasets to evaluate and run the algorithm.The first dataset, which we call gold standard hereafter, contains nouns that have a clear lexical gender and were mainly sourced from previous research on gender bias.The second dataset contains 1,000 randomly sampled Wikipedia articles, which we used to extract gendered nouns.The following describes both datasets in detail.

Gold Standard
In order to gain insights into the performance of the dictionary-based algorithm for lexical gender retrieval, we compiled a list of words that have a nearly unambiguous lexical gender, which acts as the gold standard.The gold standard list was developed based on a lexical gender list by Cao and Daumé III (2020) with the addition of more words retrieved from online lists for learners of English 345 .Nouns retrieved from prior research and online sources were subsequently filtered for explicitness of lexical gender.For example, the pair actor/actress would not be considered since the word actor is nowadays used for both male and female referents.We moreover added neutral gender replacements for word pairs for which such an alternative exists.An example would be the triplet headmaster-MASC, headmistress-FEM, headteacher-NEUT.The final list is comprised of 53 masculine, 53 feminine, and 28 neutral words (see Table 4 in the Appendix).

Wikipedia Sample
This research aims at providing a flexible, scalable, and high-coverage method for lexical gender detection.Therefore we additionally tested the approach on more naturalistic data, namely a random sample of 1,000 articles from English Wikipedia obtained through the wikipedia python library6 .We will abbreviate this sample corpus as Wiki1000 hereafter.
The articles were then cleaned and tokenized into sentences using NLTK (Bird et al., 2009) and subsequently processed with SpaCy to obtain partof-speech (POS) tags for each word.All singular and plural nouns (POS-tags: NN, NNS) were then extracted and analysed for lexical gender.Nouns that contained special characters due to cleaning and tokenization errors were dropped.This method provided us with 12,643 nouns, as illustrated under Wiki1000 in Table 2.
In order to test the performance of the algorithm, the instances of the Wiki1000 dataset needed true labels.A corpus size of 12,643 instances, however, was beyond the scope of this research to manually label.In fact, it represents the kind of corpus size that we aim to label automatically.We therefore filtered Wiki1000 for nouns that were labelled as either masculine or feminine by Merriam Webster Online, Dictionary.com, or WordNet.Like this, we specifically target gendered nouns and obtain a corpus similar to the gold standard corpus, but sourced from naturally occurring text.The resulting corpus of 515 nouns, which we call Wiki1000-sample, was subsequently labelled for 'true' lexical gender by members of the research team (Fleiss's κ ≈ 0.87).
The labels used for evaluation were determined by majority vote.The specifications of the Wiki1000sample dataset can be found in Table 2.
In line with previous research on gender bias in Wikipedia (Wagner et al., 2015;Falenska and Çetinoglu, 2021), which found an overrepresentation of male entities in the encyclopedia, Table 2 shows that there are approximately 1.5 times as many mentions of distinct entities with masculine lexical gender in our 1,000-article Wikipedia sample than there of entities with feminine lexical gender.

Quantitative analysis
An overview of algorithm performance on the gold standard dataset and the reduced Wiki1000 sample can be found in Table 3.We report the weighted average of precision, recall, and F1-measure due to unbalanced classes in our test data.
As seen in Table 3, our best performing approach on both the gold dataset (87% accuracy) as well as the sample of Wiki1000 (84% accuracy) was combining labels from all three sources by majority vote.Keeping in mind that the Wiki1000 sample is approximately three times the size of the gold standard, the relative consistency in performance here indicates robustness for our approach.It should also be noted that only querying Dictionary.com reached the same performance on the gold standard dataset (87% accuracy) while on the Wiki1000 sample, using only Merriam Webster reached a compa-rable accuracy score to the combined model (82%).
Table 3 moreover shows that on the gold standard dataset, which was used to fine-tune our parameter values using grid search, our method reached an accuracy of 77% or higher in each experiment configuration.Using the same parameter values for experiments on the Wiki1000 sample, only the combined approach as well as using only Merriam Webster reaches an accuracy of >77%.When using only WordNet or Dictionary.com, the performance drops from 84% to 63% and 61% accuracy, respectively.This shows that parameter configurations can be adapted to specific dictionaries and dataset sizes.
Figure 2 shows confusion matrices for the combined approach on both the gold standard dataset (2a) and the Wiki1000-sample (2b). Figure 2a shows that on the gold standard, the combined classifier mislabelled four feminine and 11 masculine instances as neutral, but did not mislabel any of the neutral instances as either masculine or feminine.In contrast, both these classification mistakes can be found on the Wiki1000 sample (Figure 2b).Here, the algorithm classifies more lexically neutral words as gendered than vice versa.
Cases in which lexically neutral words are classified as gendered include words that are traditionally related to specific genders, such as bikini or soprano, as well as patriarchy or testes.It is likely that dictionary definitions reflect this traditional gender association, leading to misclassification.Conversely, classifications of gendered words as neutral can e.g.be caused by definitions that do not mention gender, either because of presumed knowledge (pope) or because a lexically specific word was formerly used for all genders (landlord).Another reason for gendered-as-neutral misclassification can be the definition of one gendered term by using another, which 'cancel each other out'.For example, WordNet defines widow as "a woman whose husband is dead especially one who has not remarried" (Princeton University, 2010).
Another issue, which only occurred when testing on the gold standard dataset, concerns words that could not be found.The first is single person, which we chose as gender-neutral alternative for bachelor/spinster.The fact that it was not found could be due to the term single person being more of a composite phrase than a joined expression.Moreover, single people are often described using the adjective single in a predicative way, such as in the sentence 'He is single.',instead of 'He is a single person.'The other word that could not be found is child-in-law, which is the gender-neutral variant of son/daughter-in-law.Here, the issue could be frequency of use, since child-in-law is less established than its gender-specific variants.

Qualitative analysis
The following section discusses some classification errors in more detail.We focus on errors that occur due to gender-exclusive definitions in the lexical databases caused by historically close associations of words to a single gender.
In our first example, an outdated definition in WordNet (Princeton University, 2010) causes the misclassification of the word crew, a neutral term, as masculine.We show the first and fourth definitions in Example (4) in order to illustrate how the masculine label was obtained.
(4) crew 1. the men and women who man a vehicle (ship, aircraft, etc.) 4. the team of men manning a racing shell In the first definition, the words men and women are used to define the crew of any vehicle while in the fourth definition, which describes the crew of a racing shell (a type of rowing boat), only the word men is used.This leads to a masculine lexical gender label, since the definitions taken together contain more masculine than feminine words.However, the fourth definition could have been worded like the first, or used the word people, since racing shells can be crewed by people of any gender.
A similar classification error occurred for the words soprano, menopause and nurse, which were all classified as feminine by the combined model, even though they have neutral lexical gender.These terms are all closely associated with female social gender due to anatomical and hormonal differences between sexes (soprano and menopause), historical biases of women performing care-work, as well as current gender distributions in certain professions (nurse; Bureau of Labor Statistics (BLS), 2022).While using gender-exclusive wording to define lexically neutral terms could inform readers of a word's traditional relation to social gender, it can also reproduce gender stereotypes and exclude those who do not identify as female but still sing in soprano voice or work as a nurse.Moreover, using feminine words in the definition of words like menopause can be seen as a form of transexclusionary bias, since people assigned female at birth, whose body can cease to menstruate, might not identify as female.

Limitations and Future Developments
We have selected dictionaries to obtain the lexical gender of a word, because they represent a relatively objective resource that is expected to list neutral and non-stereotypical definitions of words.However, as shown in Section 5.2, dictionaries are after all a human-curated resource and as such still carry human biases and outdated definitions, which in turn lead to biased or outdated results.
We would moreover like to point out that we are explicitly working with English, which does not mark gender grammatically.In languages that mark grammatical gender, our method would most likely be obsolete, because here gender can e.g.be inferred from formal features such as morphology or agreement for most nouns (Corbett, 1991).What is more, English, as a lingua franca and the language most focused on by the NLP community (Bender et al., 2021), has a plethora of high-quality and high-coverage resources available.Since our method is reliant on lexical resources, adapting the method to low-resource languages could prove challenging.However, while more complex lexical resources like WordNet might not yet exist for some languages, it is likely that online dictionaries do exist.Therefore, we still believe that our method can be adapted to other notional gender languages (McConnell-Ginet, 2013).
Another limitation of the present work concerns word sense disambiguation, since the presence of lexical gender depends on the word's sense in context.As an example, the word colt, can either mean a young male horse or a brand of pistol.In the sense of a male horse, the lexical gender of colt is clearly masculine while in the sense of the pistol, it is neutral.Differences in the lexical gender of word senses can also be caused by semantic shifts, such as for the word master, which traditionally refers to a man who is in control of e.g.servants or a household.However, in an academic context its meaning has shifted and now refers to an academic degree, or more broadly to a person of undefined gender who has reached a high level of skill in a given discipline.Therefore, future work will integrate word sense disambiguation within the algorithm.

Conclusion
We have presented a method to automatically determine the lexical gender of a given word by querying its dictionary definitions.The performance of the algorithm on a gold standard dataset of gendered nouns based on related literature, as well as a set of nouns sampled from 1,000 randomly selected Wikipedia articles, reached up to 87% accuracy.Previous research on gender bias in NLP used manually compiled lists of gendered words for data evaluation, retrieval, manipulation, and the synthetic creation of data.In contrast, our method is scalable and has a high, dynamic coverage, which gives it a variety of applications within past and future research on gender bias in NLP.These include e.g. the assessment of gender representations in large-scale corpora, the retrieval of gendered words for which gender-neutral replacements need to be found, as well as determining whether male-centric language such as epicene he is used in coreference resolution clusters.

Figure 1 :
Figure 1: Simplified exemplary flowchart of gender detection algorithm

Figure 2 :
Figure 2: Confusion matrices for combined labels words that were not found in (a): single person, child-in-law

Table 1 :
Words carrying explicit lexical gender; w = number of pairs used for experiments

Table 2 :
Composition of evaluation corpora for lexical gender detection algorithm.

Table 3 :
Quantitative results for lexical gender detection of gold standard and Wiki1000-sample