A machine learning approach to recognize bias and discrimination in job advertisements ↔ Effective Strategies for Mitigating Bias in Hiring Algorithms: A Comparative Analysis

LATER¶37-41

3.2 What are the different types of bias that need to be addressed? There are a number of different types of bias that need to be addressed within the realm of Artificial Intelligence. For instance, the measure of input bias (Ib) is important when it comes to assessing technological advancement7. This measure of technological scale bias (TS,) is given by the equation2,15, and is commonly used in epidemiology to detect bias due to measurement error12. Additionally, the total bias needs to be calculated for individual applications, which includes measurement error bias, response propensity and non-response bias13. Two different association tests have been designed to measure bias, one at sentence level (intra-sentence), and the other at discourse level (inter-sentence)11. Furthermore, electionby-election estimates of partisan bias can be calculated2, and a checklist has been provided to measure the race bias3. The criteria for measuring bias should satisfy the condition that if a set remains unchanged, it should be assigned a value of zero4. Finally, a language of measurement modeling has been introduced to uncover the implicit constructs that such systems rely on5. Thus, it is essential to take these various types of bias into account when developing and assessing AHSs to reduce any potential systemic discrimination. 3.3 What metrics are used to evaluate the performance of a hiring algorithm? In order to evaluate the performance of a hiring algorithm, various metrics are used. For instance, the measure of input bias (Ib) and the measure of technological scale bias (TS,) are two such metrics used to assess technological advancement. Additionally, the measure of bias due to measurement error2 is commonly used in epidemiology. Furthermore, when nonresponse is encountered, total bias needs to be measured9. For instance, two different association tests can be designed to measure bias at the sentence and discourse levels10, with election-by-election estimates of partisan bias used to assess methods for measuring bias and responsiveness12. Moreover, before considering the measurement of race bias13, a two-criterion approach for measuring bias must be adopted14. Finally, to uncover the implicit constructs of a hiring algorithm, measurement modeling can help15. Thus, to assess the performance of a hiring algorithm, a variety of metrics are used to evaluate the fairness and equity of recruitment practices. 4. Model Design 4.1 Feature engineering Linguistic features Several linguistic features were used for each token. Each feature represents a characteristic property of the word. (Table 1) presents an explanation for each linguistic feature. Semantic features We utilized various state-of-the-art pre-trained word embeddings as textual features for the machine learning classifiers. The different word embeddings which were used are: Word2Vec (Mikolov et al. 2013), BERT (Devlin et al. 2019), ELMo (Peters et al. 2018), GloVe (Pennington et al. 2014), Flair (Akbik et al. 2018) and FastText (Bojanowski et al. 2017). Pre-trained word embeddings were used because the word embeddings trained on the EMSCAD dataset did not 5 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 demonstrate sufficient semantic quality due to the smaller size of the dataset. (Table 2) shows the pre-trained models used for each word embedding.

LATER¶38-42

3.3 What metrics are used to evaluate the performance of a hiring algorithm? In order to evaluate the performance of a hiring algorithm, various metrics are used. For instance, the measure of input bias (Ib) and the measure of technological scale bias (TS,) are two such metrics used to assess technological advancement. Additionally, the measure of bias due to measurement error2 is commonly used in epidemiology. Furthermore, when nonresponse is encountered, total bias needs to be measured9. For instance, two different association tests can be designed to measure bias at the sentence and discourse levels10, with election-by-election estimates of partisan bias used to assess methods for measuring bias and responsiveness12. Moreover, before considering the measurement of race bias13, a two-criterion approach for measuring bias must be adopted14. Finally, to uncover the implicit constructs of a hiring algorithm, measurement modeling can help15. Thus, to assess the performance of a hiring algorithm, a variety of metrics are used to evaluate the fairness and equity of recruitment practices. 4. Model Design 4.1 Feature engineering Linguistic features Several linguistic features were used for each token. Each feature represents a characteristic property of the word. (Table 1) presents an explanation for each linguistic feature. Semantic features We utilized various state-of-the-art pre-trained word embeddings as textual features for the machine learning classifiers. The different word embeddings which were used are: Word2Vec (Mikolov et al. 2013), BERT (Devlin et al. 2019), ELMo (Peters et al. 2018), GloVe (Pennington et al. 2014), Flair (Akbik et al. 2018) and FastText (Bojanowski et al. 2017). Pre-trained word embeddings were used because the word embeddings trained on the EMSCAD dataset did not 5 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 demonstrate sufficient semantic quality due to the smaller size of the dataset. (Table 2) shows the pre-trained models used for each word embedding. For each token, the word embedding vectors were extracted from the corresponding word embedding model using the FlairFootnote1 library. Feature selection

LATER¶38-42

3.3 What metrics are used to evaluate the performance of a hiring algorithm? In order to evaluate the performance of a hiring algorithm, various metrics are used. For instance, the measure of input bias (Ib) and the measure of technological scale bias (TS,) are two such metrics used to assess technological advancement. Additionally, the measure of bias due to measurement error2 is commonly used in epidemiology. Furthermore, when nonresponse is encountered, total bias needs to be measured9. For instance, two different association tests can be designed to measure bias at the sentence and discourse levels10, with election-by-election estimates of partisan bias used to assess methods for measuring bias and responsiveness12. Moreover, before considering the measurement of race bias13, a two-criterion approach for measuring bias must be adopted14. Finally, to uncover the implicit constructs of a hiring algorithm, measurement modeling can help15. Thus, to assess the performance of a hiring algorithm, a variety of metrics are used to evaluate the fairness and equity of recruitment practices. 4. Model Design 4.1 Feature engineering Linguistic features Several linguistic features were used for each token. Each feature represents a characteristic property of the word. (Table 1) presents an explanation for each linguistic feature. Semantic features We utilized various state-of-the-art pre-trained word embeddings as textual features for the machine learning classifiers. The different word embeddings which were used are: Word2Vec (Mikolov et al. 2013), BERT (Devlin et al. 2019), ELMo (Peters et al. 2018), GloVe (Pennington et al. 2014), Flair (Akbik et al. 2018) and FastText (Bojanowski et al. 2017). Pre-trained word embeddings were used because the word embeddings trained on the EMSCAD dataset did not 5 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 demonstrate sufficient semantic quality due to the smaller size of the dataset. (Table 2) shows the pre-trained models used for each word embedding. For each token, the word embedding vectors were extracted from the corresponding word embedding model using the FlairFootnote1 library. Feature selection

LATER¶39-43

4. Model Design 4.1 Feature engineering Linguistic features Several linguistic features were used for each token. Each feature represents a characteristic property of the word. (Table 1) presents an explanation for each linguistic feature. Semantic features We utilized various state-of-the-art pre-trained word embeddings as textual features for the machine learning classifiers. The different word embeddings which were used are: Word2Vec (Mikolov et al. 2013), BERT (Devlin et al. 2019), ELMo (Peters et al. 2018), GloVe (Pennington et al. 2014), Flair (Akbik et al. 2018) and FastText (Bojanowski et al. 2017). Pre-trained word embeddings were used because the word embeddings trained on the EMSCAD dataset did not 5 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 demonstrate sufficient semantic quality due to the smaller size of the dataset. (Table 2) shows the pre-trained models used for each word embedding. For each token, the word embedding vectors were extracted from the corresponding word embedding model using the FlairFootnote1 library. Feature selection The aforementioned linguistic features were combined with one of the six semantic features (word embedding) to produce a unique feature set. As a result, six unique feature sets were produced as input to the machine learning classifiers.

LATER¶39-44

4. Model Design 4.1 Feature engineering Linguistic features Several linguistic features were used for each token. Each feature represents a characteristic property of the word. (Table 1) presents an explanation for each linguistic feature. Semantic features We utilized various state-of-the-art pre-trained word embeddings as textual features for the machine learning classifiers. The different word embeddings which were used are: Word2Vec (Mikolov et al. 2013), BERT (Devlin et al. 2019), ELMo (Peters et al. 2018), GloVe (Pennington et al. 2014), Flair (Akbik et al. 2018) and FastText (Bojanowski et al. 2017). Pre-trained word embeddings were used because the word embeddings trained on the EMSCAD dataset did not 5 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 demonstrate sufficient semantic quality due to the smaller size of the dataset. (Table 2) shows the pre-trained models used for each word embedding. For each token, the word embedding vectors were extracted from the corresponding word embedding model using the FlairFootnote1 library. Feature selection The aforementioned linguistic features were combined with one of the six semantic features (word embedding) to produce a unique feature set. As a result, six unique feature sets were produced as input to the machine learning classifiers. 4.2 Machine learning classifiers The machine learning classifiers were trained using the six unique feature sets on the training set of annotated job descriptions. The following classifiers were used:

LATER¶40-45

Semantic features We utilized various state-of-the-art pre-trained word embeddings as textual features for the machine learning classifiers. The different word embeddings which were used are: Word2Vec (Mikolov et al. 2013), BERT (Devlin et al. 2019), ELMo (Peters et al. 2018), GloVe (Pennington et al. 2014), Flair (Akbik et al. 2018) and FastText (Bojanowski et al. 2017). Pre-trained word embeddings were used because the word embeddings trained on the EMSCAD dataset did not 5 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 demonstrate sufficient semantic quality due to the smaller size of the dataset. (Table 2) shows the pre-trained models used for each word embedding. For each token, the word embedding vectors were extracted from the corresponding word embedding model using the FlairFootnote1 library. Feature selection The aforementioned linguistic features were combined with one of the six semantic features (word embedding) to produce a unique feature set. As a result, six unique feature sets were produced as input to the machine learning classifiers. 4.2 Machine learning classifiers The machine learning classifiers were trained using the six unique feature sets on the training set of annotated job descriptions. The following classifiers were used: For the baseline classifier, Scikit-learn’s Dummy classifier was utilized. By performing parameter optimization using GridSearch, Footnote2 it was possible to search for the optimal parameters for all the machine learning classifiers. For all the classifiers, the maximum iterations were increased to infinity to ensure that the models are able to converge. All parameters, including the default parameters utilized for model training, are presented in (Table 3).

LATER¶42-47

For each token, the word embedding vectors were extracted from the corresponding word embedding model using the FlairFootnote1 library. Feature selection The aforementioned linguistic features were combined with one of the six semantic features (word embedding) to produce a unique feature set. As a result, six unique feature sets were produced as input to the machine learning classifiers. 4.2 Machine learning classifiers The machine learning classifiers were trained using the six unique feature sets on the training set of annotated job descriptions. The following classifiers were used: For the baseline classifier, Scikit-learn’s Dummy classifier was utilized. By performing parameter optimization using GridSearch, Footnote2 it was possible to search for the optimal parameters for all the machine learning classifiers. For all the classifiers, the maximum iterations were increased to infinity to ensure that the models are able to converge. All parameters, including the default parameters utilized for model training, are presented in (Table 3). Table 1: Linguistic features. Table 2: Word embeddings characteristics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 6 Table 3: Parameter grid. 5. Results and Analysis In this section, we present the results of various machine learning models on the EMSCAD dataset. The dataset was divided into 80% training and 20% testing set. The evaluation metrics: accuracy, precision, recall and F1-score were computed for each model. Figure 1 presents the evaluation metrics for various classifiers with different feature sets. The results indicate that the RF classifier with BERT word embeddings as textual feature achieved the best performance. This illustrates that contextual word embedding representations such as BERT had a superior performance over the non-contextual word embeddings such as FastText and Word2vec. We also observe that tree-based (Random Forest and Decision Tree) classifiers had a better performance in classifying biased and discriminatory language

LATER¶43-47

The aforementioned linguistic features were combined with one of the six semantic features (word embedding) to produce a unique feature set. As a result, six unique feature sets were produced as input to the machine learning classifiers. 4.2 Machine learning classifiers The machine learning classifiers were trained using the six unique feature sets on the training set of annotated job descriptions. The following classifiers were used: For the baseline classifier, Scikit-learn’s Dummy classifier was utilized. By performing parameter optimization using GridSearch, Footnote2 it was possible to search for the optimal parameters for all the machine learning classifiers. For all the classifiers, the maximum iterations were increased to infinity to ensure that the models are able to converge. All parameters, including the default parameters utilized for model training, are presented in (Table 3). Table 1: Linguistic features. Table 2: Word embeddings characteristics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 6 Table 3: Parameter grid. 5. Results and Analysis In this section, we present the results of various machine learning models on the EMSCAD dataset. The dataset was divided into 80% training and 20% testing set. The evaluation metrics: accuracy, precision, recall and F1-score were computed for each model. Figure 1 presents the evaluation metrics for various classifiers with different feature sets. The results indicate that the RF classifier with BERT word embeddings as textual feature achieved the best performance. This illustrates that contextual word embedding representations such as BERT had a superior performance over the non-contextual word embeddings such as FastText and Word2vec. We also observe that tree-based (Random Forest and Decision Tree) classifiers had a better performance in classifying biased and discriminatory language

LATER¶45-51

For the baseline classifier, Scikit-learn’s Dummy classifier was utilized. By performing parameter optimization using GridSearch, Footnote2 it was possible to search for the optimal parameters for all the machine learning classifiers. For all the classifiers, the maximum iterations were increased to infinity to ensure that the models are able to converge. All parameters, including the default parameters utilized for model training, are presented in (Table 3). Table 1: Linguistic features. Table 2: Word embeddings characteristics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 6 Table 3: Parameter grid. 5. Results and Analysis In this section, we present the results of various machine learning models on the EMSCAD dataset. The dataset was divided into 80% training and 20% testing set. The evaluation metrics: accuracy, precision, recall and F1-score were computed for each model. Figure 1 presents the evaluation metrics for various classifiers with different feature sets. The results indicate that the RF classifier with BERT word embeddings as textual feature achieved the best performance. This illustrates that contextual word embedding representations such as BERT had a superior performance over the non-contextual word embeddings such as FastText and Word2vec. We also observe that tree-based (Random Forest and Decision Tree) classifiers had a better performance in classifying biased and discriminatory language as compared to the remaining classifiers. Among the textual features, word embedding representations BERT, FastText and ELMo in combination with the RF classifier had the best performance. This was followed by FastText, ELMo and Flair word embeddings in combination with the DT classifier. We further evaluate the various machine learning classifiers with different word embedding representations as features using tenfolds cross-validation. (Figure 2) presents the macroaverages of the precision, recall and F1-score over tenfolds cross-validation. The results of the tenfolds cross-validation indicate that the RF classifier with FastText word embeddings had the best performance. (Figures 3, 4, 5 and 6) present the individual results for accuracy, precision, recall and F1-score for the various models. 7 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Figure 1: Machine learning models performance metrics. Figure 2: Cross-validated Machine learning performance metrics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 8 Figure 3: Machine learning models-accuracy. Figure 4: Machine learning models-precision.

LATER¶47-51

5. Results and Analysis In this section, we present the results of various machine learning models on the EMSCAD dataset. The dataset was divided into 80% training and 20% testing set. The evaluation metrics: accuracy, precision, recall and F1-score were computed for each model. Figure 1 presents the evaluation metrics for various classifiers with different feature sets. The results indicate that the RF classifier with BERT word embeddings as textual feature achieved the best performance. This illustrates that contextual word embedding representations such as BERT had a superior performance over the non-contextual word embeddings such as FastText and Word2vec. We also observe that tree-based (Random Forest and Decision Tree) classifiers had a better performance in classifying biased and discriminatory language as compared to the remaining classifiers. Among the textual features, word embedding representations BERT, FastText and ELMo in combination with the RF classifier had the best performance. This was followed by FastText, ELMo and Flair word embeddings in combination with the DT classifier. We further evaluate the various machine learning classifiers with different word embedding representations as features using tenfolds cross-validation. (Figure 2) presents the macroaverages of the precision, recall and F1-score over tenfolds cross-validation. The results of the tenfolds cross-validation indicate that the RF classifier with FastText word embeddings had the best performance. (Figures 3, 4, 5 and 6) present the individual results for accuracy, precision, recall and F1-score for the various models. 7 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Figure 1: Machine learning models performance metrics. Figure 2: Cross-validated Machine learning performance metrics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 8 Figure 3: Machine learning models-accuracy. Figure 4: Machine learning models-precision.

LATER¶47-51

5. Results and Analysis In this section, we present the results of various machine learning models on the EMSCAD dataset. The dataset was divided into 80% training and 20% testing set. The evaluation metrics: accuracy, precision, recall and F1-score were computed for each model. Figure 1 presents the evaluation metrics for various classifiers with different feature sets. The results indicate that the RF classifier with BERT word embeddings as textual feature achieved the best performance. This illustrates that contextual word embedding representations such as BERT had a superior performance over the non-contextual word embeddings such as FastText and Word2vec. We also observe that tree-based (Random Forest and Decision Tree) classifiers had a better performance in classifying biased and discriminatory language as compared to the remaining classifiers. Among the textual features, word embedding representations BERT, FastText and ELMo in combination with the RF classifier had the best performance. This was followed by FastText, ELMo and Flair word embeddings in combination with the DT classifier. We further evaluate the various machine learning classifiers with different word embedding representations as features using tenfolds cross-validation. (Figure 2) presents the macroaverages of the precision, recall and F1-score over tenfolds cross-validation. The results of the tenfolds cross-validation indicate that the RF classifier with FastText word embeddings had the best performance. (Figures 3, 4, 5 and 6) present the individual results for accuracy, precision, recall and F1-score for the various models. 7 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Figure 1: Machine learning models performance metrics. Figure 2: Cross-validated Machine learning performance metrics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 8 Figure 3: Machine learning models-accuracy. Figure 4: Machine learning models-precision.

LATER¶50-54

7 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Figure 1: Machine learning models performance metrics. Figure 2: Cross-validated Machine learning performance metrics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 8 Figure 3: Machine learning models-accuracy. Figure 4: Machine learning models-precision. Figure 5: Machine learning models-recall. Figure 6: Machine learning models-F1. Full size image (Figures 7 and 8) present the confusion matrices of the two best performing models: (1) RF classifier with FastText word embeddings and (2) RF classifier with BERT word embeddings. The results in (Figs. 7 and 8) indicate that all the five classes of biased and discriminatory language were distinguishable from each other. Figure 7: Confusion matrix Random Forest-FastText. Figure 8: Confusion matrix Random Forest-BERT. 9 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Full size image We observe a linear improvement in the performance of a sample of our models as training size increases in our primary experiment. We wanted to see if the behavior was by any chance related to these particular models or statistically grounded across all our implemented models. To validate this improvement, we included additional data (3000 additional job descriptions) and ran a new experiment with for the lightweight classifiers (DT, LR and NB).

LATER¶50-54

7 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Figure 1: Machine learning models performance metrics. Figure 2: Cross-validated Machine learning performance metrics. J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 8 Figure 3: Machine learning models-accuracy. Figure 4: Machine learning models-precision. Figure 5: Machine learning models-recall. Figure 6: Machine learning models-F1. Full size image (Figures 7 and 8) present the confusion matrices of the two best performing models: (1) RF classifier with FastText word embeddings and (2) RF classifier with BERT word embeddings. The results in (Figs. 7 and 8) indicate that all the five classes of biased and discriminatory language were distinguishable from each other. Figure 7: Confusion matrix Random Forest-FastText. Figure 8: Confusion matrix Random Forest-BERT. 9 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Full size image We observe a linear improvement in the performance of a sample of our models as training size increases in our primary experiment. We wanted to see if the behavior was by any chance related to these particular models or statistically grounded across all our implemented models. To validate this improvement, we included additional data (3000 additional job descriptions) and ran a new experiment with for the lightweight classifiers (DT, LR and NB).

LATER¶52-57

Figure 5: Machine learning models-recall. Figure 6: Machine learning models-F1. Full size image (Figures 7 and 8) present the confusion matrices of the two best performing models: (1) RF classifier with FastText word embeddings and (2) RF classifier with BERT word embeddings. The results in (Figs. 7 and 8) indicate that all the five classes of biased and discriminatory language were distinguishable from each other. Figure 7: Confusion matrix Random Forest-FastText. Figure 8: Confusion matrix Random Forest-BERT. 9 Jazakallah Y., J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Full size image We observe a linear improvement in the performance of a sample of our models as training size increases in our primary experiment. We wanted to see if the behavior was by any chance related to these particular models or statistically grounded across all our implemented models. To validate this improvement, we included additional data (3000 additional job descriptions) and ran a new experiment with for the lightweight classifiers (DT, LR and NB). The results obtained from the new experiment are shown in Fig. 9 for the regular models (80% training set and 20% test set), and Fig. 10 for the tenfold cross-validated models. We see that the DT classifier which uses BERT word embeddings produced the best performance: 0.98977, 0.99587 and 0.99277 for the precision, recall, and F1-score respectively. In fact, when compared to the result initially obtained by the best performing model in our first experiment, i.e., BERT—RF as shown in Fig. 1, we can see an improvement since the previous performance scores obtained were 0.98557, 0.98862 and 0.98544 equally for the precision, recall, and F1-score respectively. Figure 9: Extended machine learning models performance metrics. Figure 10: Extended cross-validated Machine learning performance metrics. This analysis reinforces our belief that even the strong performance we have observed across the board can be further improved. However, it is not currently clear if, by any stroke of chance, the data samples utilized for evaluation might have been simplistic in the sense that they represent trivial cases. To the best of our knowledge, we have avoided cherry-picking by carrying out extensive random-sampling to select the evaluation set. However, given the small ratio of the entire evaluation set when compared to the size of our full data, we may not confidently rule out that this could have had an impact. In any case, we are leaving this for future study where we hope to perform a more comprehensive experiment on our entire dataset including carrying out an extensive ablation study with error analysis on the result. 7. Conclusion By combining machine learning techniques, statistical analysis, and ethical considerations, our research methodology offers a holistic approach to bias mitigation in hiring algorithms. This comprehensive framework enables us to identify, measure, and mitigate both overt and subtle forms of bias, promoting equitable and fair candidate selection practices. The synergy of these methodologies ensures that bias is not only detected but also actively addressed to enhance diversity and inclusion in the recruitment process.

LATER¶54-58

Full size image We observe a linear improvement in the performance of a sample of our models as training size increases in our primary experiment. We wanted to see if the behavior was by any chance related to these particular models or statistically grounded across all our implemented models. To validate this improvement, we included additional data (3000 additional job descriptions) and ran a new experiment with for the lightweight classifiers (DT, LR and NB). The results obtained from the new experiment are shown in Fig. 9 for the regular models (80% training set and 20% test set), and Fig. 10 for the tenfold cross-validated models. We see that the DT classifier which uses BERT word embeddings produced the best performance: 0.98977, 0.99587 and 0.99277 for the precision, recall, and F1-score respectively. In fact, when compared to the result initially obtained by the best performing model in our first experiment, i.e., BERT—RF as shown in Fig. 1, we can see an improvement since the previous performance scores obtained were 0.98557, 0.98862 and 0.98544 equally for the precision, recall, and F1-score respectively. Figure 9: Extended machine learning models performance metrics. Figure 10: Extended cross-validated Machine learning performance metrics. This analysis reinforces our belief that even the strong performance we have observed across the board can be further improved. However, it is not currently clear if, by any stroke of chance, the data samples utilized for evaluation might have been simplistic in the sense that they represent trivial cases. To the best of our knowledge, we have avoided cherry-picking by carrying out extensive random-sampling to select the evaluation set. However, given the small ratio of the entire evaluation set when compared to the size of our full data, we may not confidently rule out that this could have had an impact. In any case, we are leaving this for future study where we hope to perform a more comprehensive experiment on our entire dataset including carrying out an extensive ablation study with error analysis on the result. 7. Conclusion By combining machine learning techniques, statistical analysis, and ethical considerations, our research methodology offers a holistic approach to bias mitigation in hiring algorithms. This comprehensive framework enables us to identify, measure, and mitigate both overt and subtle forms of bias, promoting equitable and fair candidate selection practices. The synergy of these methodologies ensures that bias is not only detected but also actively addressed to enhance diversity and inclusion in the recruitment process. The integration of machine learning techniques into the research methodology plays a pivotal role in detecting, quantifying, and mitigating bias in hiring algorithms. By utilizing a combination of supervised learning, de-biasing strategies, and fairness-aware representation learning, this approach ensures that the hiring algorithm not only improves its overall performance but also adheres to ethical considerations and promotes fairness and inclusivity in candidate selection.

LATER¶54-58

Full size image We observe a linear improvement in the performance of a sample of our models as training size increases in our primary experiment. We wanted to see if the behavior was by any chance related to these particular models or statistically grounded across all our implemented models. To validate this improvement, we included additional data (3000 additional job descriptions) and ran a new experiment with for the lightweight classifiers (DT, LR and NB). The results obtained from the new experiment are shown in Fig. 9 for the regular models (80% training set and 20% test set), and Fig. 10 for the tenfold cross-validated models. We see that the DT classifier which uses BERT word embeddings produced the best performance: 0.98977, 0.99587 and 0.99277 for the precision, recall, and F1-score respectively. In fact, when compared to the result initially obtained by the best performing model in our first experiment, i.e., BERT—RF as shown in Fig. 1, we can see an improvement since the previous performance scores obtained were 0.98557, 0.98862 and 0.98544 equally for the precision, recall, and F1-score respectively. Figure 9: Extended machine learning models performance metrics. Figure 10: Extended cross-validated Machine learning performance metrics. This analysis reinforces our belief that even the strong performance we have observed across the board can be further improved. However, it is not currently clear if, by any stroke of chance, the data samples utilized for evaluation might have been simplistic in the sense that they represent trivial cases. To the best of our knowledge, we have avoided cherry-picking by carrying out extensive random-sampling to select the evaluation set. However, given the small ratio of the entire evaluation set when compared to the size of our full data, we may not confidently rule out that this could have had an impact. In any case, we are leaving this for future study where we hope to perform a more comprehensive experiment on our entire dataset including carrying out an extensive ablation study with error analysis on the result. 7. Conclusion By combining machine learning techniques, statistical analysis, and ethical considerations, our research methodology offers a holistic approach to bias mitigation in hiring algorithms. This comprehensive framework enables us to identify, measure, and mitigate both overt and subtle forms of bias, promoting equitable and fair candidate selection practices. The synergy of these methodologies ensures that bias is not only detected but also actively addressed to enhance diversity and inclusion in the recruitment process. The integration of machine learning techniques into the research methodology plays a pivotal role in detecting, quantifying, and mitigating bias in hiring algorithms. By utilizing a combination of supervised learning, de-biasing strategies, and fairness-aware representation learning, this approach ensures that the hiring algorithm not only improves its overall performance but also adheres to ethical considerations and promotes fairness and inclusivity in candidate selection.

LATER¶60-64

Ethical considerations play a pivotal role in shaping the research methodology for bias mitigation in hiring algorithms. By incorporating interdisciplinary perspectives, ensuring informed consent, promoting transparency, and fostering ongoing improvement, this approach aligns the research with ethical principles. Ultimately, the integration of ethical considerations helps create a framework that not only reduces bias but also upholds fairness, transparency, and social responsibility in the design and deployment of automated hiring systems. The proposed framework for fostering diversity and inclusion in the workplace is necessary to ensure that all employees are treated fairly and equitably. An effective framework should consider both the technical and legal aspects of recruitment and hiring, including laws and regulations that prohibit discrimination2. It should also include measures to prevent discrimination and ensure that every individual is given equal opportunities43. This is the only way to ensure that organizations can create a workforce that is diverse in terms of gender, ethnicity, age, and other factors44. Furthermore, organizations can use the framework to ensure that the recruitment and hiring process is transparent and fair5. This would help them create a positive work environment and foster inclusion and collaboration among employees46. A successful framework should also be capable of providing support and guidance to employees during the recruitment and hiring process7. Ultimately, the proposed framework should be comprehensive enough to provide a sense of closure and completeness to the recruitment and hiring process8. This would make sure that the process is conducted in a fair manner and that all applicants have an equal chance of being selected9. A proper conclusion should also be drawn upon the completion of the recruitment and hiring process5 so that the desired outcomes can be achieved5. This paper has presented a machine learning approach to identify five major categories of bias and discriminatory J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 10 language in job advertisements. We prepared a list of unique biased and discriminatory terms after examining the literature on behavioural works related to bias in recruitment. This list was used to semi-automatically generate an annotated corpus by the tagging the biased language terms (using a gazetteer-based approach) in the job advertisements of the publicly available Employment Scam Aegean Dataset, EMSCAD. This annotated corpus was used to train state-of-the-art machine learning classifiers to identify five different categories of biased and discriminatory language. We utilized a combination of linguistic features and most recent state-of-the-art word embedding representations as textual features to capture the natural language semantics of biased language. These features were fed into the machine learning classifiers. The results indicate that the Random Forest classifier with FastText word embeddings achieved the best performance with tenfold cross-validation. Overall, this work presents a major contribution in the attention phase of hiring and empowering recruiters by identifying and classifying discriminatory language in job advertisements using a machine learning-based approach. The output of this tool can be used to flag biased and discriminatory language and encourage recruiters to write more inclusive job advertisements. The findings of this research paper highlight the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. The study reveals that while automated hiring systems (AHSs) are being used to detect and address discrimination against protected groups, claims of ‘bias mitigation’ are rarely scrutinized and evaluated. The study emphasizes the importance of developing fair and equitable recruitment practices to promote greater diversity and inclusion in the workforce. The proposed framework for reducing bias in hiring algorithms aims to ensure that decision-making is done fairly and transparently, and provides support and guidance to employees during the recruitment and hiring process. However, the study also acknowledges that bias mitigation comes at a cost of efficiency and accuracy. The study suggests that ethical considerations should guide the development and implementation of AI-enabled recruitment practices, and that diverse groups should be represented in the development and testing of hiring algorithms. The study also highlights the importance of research methodology, including machine learning and statistical analysis, in providing greater accuracy and precision in results. Ultimately, this research contributes to the ongoing advancement of knowledge in the field of bias in hiring algorithms and provides valuable insights into effective strategies for mitigating bias and promoting equity in the workforce. Future research should continue to explore these issues and further develop the proposed framework for reducing bias in hiring algorithms. In conclusion, the recruitment and hiring process requires a proper conclusion to achieve the desired outcomes. This research paper highlights the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. While automated hiring systems are being used to detect and address discrimination against protected groups, claims of bias mitigation are rarely evaluated. Fair and equitable recruitment practices are essential to promote greater diversity and inclusion in the workforce.

LATER¶61-65

The proposed framework for fostering diversity and inclusion in the workplace is necessary to ensure that all employees are treated fairly and equitably. An effective framework should consider both the technical and legal aspects of recruitment and hiring, including laws and regulations that prohibit discrimination2. It should also include measures to prevent discrimination and ensure that every individual is given equal opportunities43. This is the only way to ensure that organizations can create a workforce that is diverse in terms of gender, ethnicity, age, and other factors44. Furthermore, organizations can use the framework to ensure that the recruitment and hiring process is transparent and fair5. This would help them create a positive work environment and foster inclusion and collaboration among employees46. A successful framework should also be capable of providing support and guidance to employees during the recruitment and hiring process7. Ultimately, the proposed framework should be comprehensive enough to provide a sense of closure and completeness to the recruitment and hiring process8. This would make sure that the process is conducted in a fair manner and that all applicants have an equal chance of being selected9. A proper conclusion should also be drawn upon the completion of the recruitment and hiring process5 so that the desired outcomes can be achieved5. This paper has presented a machine learning approach to identify five major categories of bias and discriminatory J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 10 language in job advertisements. We prepared a list of unique biased and discriminatory terms after examining the literature on behavioural works related to bias in recruitment. This list was used to semi-automatically generate an annotated corpus by the tagging the biased language terms (using a gazetteer-based approach) in the job advertisements of the publicly available Employment Scam Aegean Dataset, EMSCAD. This annotated corpus was used to train state-of-the-art machine learning classifiers to identify five different categories of biased and discriminatory language. We utilized a combination of linguistic features and most recent state-of-the-art word embedding representations as textual features to capture the natural language semantics of biased language. These features were fed into the machine learning classifiers. The results indicate that the Random Forest classifier with FastText word embeddings achieved the best performance with tenfold cross-validation. Overall, this work presents a major contribution in the attention phase of hiring and empowering recruiters by identifying and classifying discriminatory language in job advertisements using a machine learning-based approach. The output of this tool can be used to flag biased and discriminatory language and encourage recruiters to write more inclusive job advertisements. The findings of this research paper highlight the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. The study reveals that while automated hiring systems (AHSs) are being used to detect and address discrimination against protected groups, claims of ‘bias mitigation’ are rarely scrutinized and evaluated. The study emphasizes the importance of developing fair and equitable recruitment practices to promote greater diversity and inclusion in the workforce. The proposed framework for reducing bias in hiring algorithms aims to ensure that decision-making is done fairly and transparently, and provides support and guidance to employees during the recruitment and hiring process. However, the study also acknowledges that bias mitigation comes at a cost of efficiency and accuracy. The study suggests that ethical considerations should guide the development and implementation of AI-enabled recruitment practices, and that diverse groups should be represented in the development and testing of hiring algorithms. The study also highlights the importance of research methodology, including machine learning and statistical analysis, in providing greater accuracy and precision in results. Ultimately, this research contributes to the ongoing advancement of knowledge in the field of bias in hiring algorithms and provides valuable insights into effective strategies for mitigating bias and promoting equity in the workforce. Future research should continue to explore these issues and further develop the proposed framework for reducing bias in hiring algorithms. In conclusion, the recruitment and hiring process requires a proper conclusion to achieve the desired outcomes. This research paper highlights the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. While automated hiring systems are being used to detect and address discrimination against protected groups, claims of bias mitigation are rarely evaluated. Fair and equitable recruitment practices are essential to promote greater diversity and inclusion in the workforce. 8. References 1. Monedero JS, Dencik L, Edwards L. What does it mean to ‘solve’ the problem of discrimination in hiring?: social, technical and legal perspectives from the UK on automated hiring systems. FAT 2020;458-468.

LATER¶61-65

The proposed framework for fostering diversity and inclusion in the workplace is necessary to ensure that all employees are treated fairly and equitably. An effective framework should consider both the technical and legal aspects of recruitment and hiring, including laws and regulations that prohibit discrimination2. It should also include measures to prevent discrimination and ensure that every individual is given equal opportunities43. This is the only way to ensure that organizations can create a workforce that is diverse in terms of gender, ethnicity, age, and other factors44. Furthermore, organizations can use the framework to ensure that the recruitment and hiring process is transparent and fair5. This would help them create a positive work environment and foster inclusion and collaboration among employees46. A successful framework should also be capable of providing support and guidance to employees during the recruitment and hiring process7. Ultimately, the proposed framework should be comprehensive enough to provide a sense of closure and completeness to the recruitment and hiring process8. This would make sure that the process is conducted in a fair manner and that all applicants have an equal chance of being selected9. A proper conclusion should also be drawn upon the completion of the recruitment and hiring process5 so that the desired outcomes can be achieved5. This paper has presented a machine learning approach to identify five major categories of bias and discriminatory J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 10 language in job advertisements. We prepared a list of unique biased and discriminatory terms after examining the literature on behavioural works related to bias in recruitment. This list was used to semi-automatically generate an annotated corpus by the tagging the biased language terms (using a gazetteer-based approach) in the job advertisements of the publicly available Employment Scam Aegean Dataset, EMSCAD. This annotated corpus was used to train state-of-the-art machine learning classifiers to identify five different categories of biased and discriminatory language. We utilized a combination of linguistic features and most recent state-of-the-art word embedding representations as textual features to capture the natural language semantics of biased language. These features were fed into the machine learning classifiers. The results indicate that the Random Forest classifier with FastText word embeddings achieved the best performance with tenfold cross-validation. Overall, this work presents a major contribution in the attention phase of hiring and empowering recruiters by identifying and classifying discriminatory language in job advertisements using a machine learning-based approach. The output of this tool can be used to flag biased and discriminatory language and encourage recruiters to write more inclusive job advertisements. The findings of this research paper highlight the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. The study reveals that while automated hiring systems (AHSs) are being used to detect and address discrimination against protected groups, claims of ‘bias mitigation’ are rarely scrutinized and evaluated. The study emphasizes the importance of developing fair and equitable recruitment practices to promote greater diversity and inclusion in the workforce. The proposed framework for reducing bias in hiring algorithms aims to ensure that decision-making is done fairly and transparently, and provides support and guidance to employees during the recruitment and hiring process. However, the study also acknowledges that bias mitigation comes at a cost of efficiency and accuracy. The study suggests that ethical considerations should guide the development and implementation of AI-enabled recruitment practices, and that diverse groups should be represented in the development and testing of hiring algorithms. The study also highlights the importance of research methodology, including machine learning and statistical analysis, in providing greater accuracy and precision in results. Ultimately, this research contributes to the ongoing advancement of knowledge in the field of bias in hiring algorithms and provides valuable insights into effective strategies for mitigating bias and promoting equity in the workforce. Future research should continue to explore these issues and further develop the proposed framework for reducing bias in hiring algorithms. In conclusion, the recruitment and hiring process requires a proper conclusion to achieve the desired outcomes. This research paper highlights the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. While automated hiring systems are being used to detect and address discrimination against protected groups, claims of bias mitigation are rarely evaluated. Fair and equitable recruitment practices are essential to promote greater diversity and inclusion in the workforce. 8. References 1. Monedero JS, Dencik L, Edwards L. What does it mean to ‘solve’ the problem of discrimination in hiring?: social, technical and legal perspectives from the UK on automated hiring systems. FAT 2020;458-468.

LATER¶61-65

The proposed framework for fostering diversity and inclusion in the workplace is necessary to ensure that all employees are treated fairly and equitably. An effective framework should consider both the technical and legal aspects of recruitment and hiring, including laws and regulations that prohibit discrimination2. It should also include measures to prevent discrimination and ensure that every individual is given equal opportunities43. This is the only way to ensure that organizations can create a workforce that is diverse in terms of gender, ethnicity, age, and other factors44. Furthermore, organizations can use the framework to ensure that the recruitment and hiring process is transparent and fair5. This would help them create a positive work environment and foster inclusion and collaboration among employees46. A successful framework should also be capable of providing support and guidance to employees during the recruitment and hiring process7. Ultimately, the proposed framework should be comprehensive enough to provide a sense of closure and completeness to the recruitment and hiring process8. This would make sure that the process is conducted in a fair manner and that all applicants have an equal chance of being selected9. A proper conclusion should also be drawn upon the completion of the recruitment and hiring process5 so that the desired outcomes can be achieved5. This paper has presented a machine learning approach to identify five major categories of bias and discriminatory J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 10 language in job advertisements. We prepared a list of unique biased and discriminatory terms after examining the literature on behavioural works related to bias in recruitment. This list was used to semi-automatically generate an annotated corpus by the tagging the biased language terms (using a gazetteer-based approach) in the job advertisements of the publicly available Employment Scam Aegean Dataset, EMSCAD. This annotated corpus was used to train state-of-the-art machine learning classifiers to identify five different categories of biased and discriminatory language. We utilized a combination of linguistic features and most recent state-of-the-art word embedding representations as textual features to capture the natural language semantics of biased language. These features were fed into the machine learning classifiers. The results indicate that the Random Forest classifier with FastText word embeddings achieved the best performance with tenfold cross-validation. Overall, this work presents a major contribution in the attention phase of hiring and empowering recruiters by identifying and classifying discriminatory language in job advertisements using a machine learning-based approach. The output of this tool can be used to flag biased and discriminatory language and encourage recruiters to write more inclusive job advertisements. The findings of this research paper highlight the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. The study reveals that while automated hiring systems (AHSs) are being used to detect and address discrimination against protected groups, claims of ‘bias mitigation’ are rarely scrutinized and evaluated. The study emphasizes the importance of developing fair and equitable recruitment practices to promote greater diversity and inclusion in the workforce. The proposed framework for reducing bias in hiring algorithms aims to ensure that decision-making is done fairly and transparently, and provides support and guidance to employees during the recruitment and hiring process. However, the study also acknowledges that bias mitigation comes at a cost of efficiency and accuracy. The study suggests that ethical considerations should guide the development and implementation of AI-enabled recruitment practices, and that diverse groups should be represented in the development and testing of hiring algorithms. The study also highlights the importance of research methodology, including machine learning and statistical analysis, in providing greater accuracy and precision in results. Ultimately, this research contributes to the ongoing advancement of knowledge in the field of bias in hiring algorithms and provides valuable insights into effective strategies for mitigating bias and promoting equity in the workforce. Future research should continue to explore these issues and further develop the proposed framework for reducing bias in hiring algorithms. In conclusion, the recruitment and hiring process requires a proper conclusion to achieve the desired outcomes. This research paper highlights the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. While automated hiring systems are being used to detect and address discrimination against protected groups, claims of bias mitigation are rarely evaluated. Fair and equitable recruitment practices are essential to promote greater diversity and inclusion in the workforce. 8. References 1. Monedero JS, Dencik L, Edwards L. What does it mean to ‘solve’ the problem of discrimination in hiring?: social, technical and legal perspectives from the UK on automated hiring systems. FAT 2020;458-468.

LATER¶61-65

The proposed framework for fostering diversity and inclusion in the workplace is necessary to ensure that all employees are treated fairly and equitably. An effective framework should consider both the technical and legal aspects of recruitment and hiring, including laws and regulations that prohibit discrimination2. It should also include measures to prevent discrimination and ensure that every individual is given equal opportunities43. This is the only way to ensure that organizations can create a workforce that is diverse in terms of gender, ethnicity, age, and other factors44. Furthermore, organizations can use the framework to ensure that the recruitment and hiring process is transparent and fair5. This would help them create a positive work environment and foster inclusion and collaboration among employees46. A successful framework should also be capable of providing support and guidance to employees during the recruitment and hiring process7. Ultimately, the proposed framework should be comprehensive enough to provide a sense of closure and completeness to the recruitment and hiring process8. This would make sure that the process is conducted in a fair manner and that all applicants have an equal chance of being selected9. A proper conclusion should also be drawn upon the completion of the recruitment and hiring process5 so that the desired outcomes can be achieved5. This paper has presented a machine learning approach to identify five major categories of bias and discriminatory J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 10 language in job advertisements. We prepared a list of unique biased and discriminatory terms after examining the literature on behavioural works related to bias in recruitment. This list was used to semi-automatically generate an annotated corpus by the tagging the biased language terms (using a gazetteer-based approach) in the job advertisements of the publicly available Employment Scam Aegean Dataset, EMSCAD. This annotated corpus was used to train state-of-the-art machine learning classifiers to identify five different categories of biased and discriminatory language. We utilized a combination of linguistic features and most recent state-of-the-art word embedding representations as textual features to capture the natural language semantics of biased language. These features were fed into the machine learning classifiers. The results indicate that the Random Forest classifier with FastText word embeddings achieved the best performance with tenfold cross-validation. Overall, this work presents a major contribution in the attention phase of hiring and empowering recruiters by identifying and classifying discriminatory language in job advertisements using a machine learning-based approach. The output of this tool can be used to flag biased and discriminatory language and encourage recruiters to write more inclusive job advertisements. The findings of this research paper highlight the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. The study reveals that while automated hiring systems (AHSs) are being used to detect and address discrimination against protected groups, claims of ‘bias mitigation’ are rarely scrutinized and evaluated. The study emphasizes the importance of developing fair and equitable recruitment practices to promote greater diversity and inclusion in the workforce. The proposed framework for reducing bias in hiring algorithms aims to ensure that decision-making is done fairly and transparently, and provides support and guidance to employees during the recruitment and hiring process. However, the study also acknowledges that bias mitigation comes at a cost of efficiency and accuracy. The study suggests that ethical considerations should guide the development and implementation of AI-enabled recruitment practices, and that diverse groups should be represented in the development and testing of hiring algorithms. The study also highlights the importance of research methodology, including machine learning and statistical analysis, in providing greater accuracy and precision in results. Ultimately, this research contributes to the ongoing advancement of knowledge in the field of bias in hiring algorithms and provides valuable insights into effective strategies for mitigating bias and promoting equity in the workforce. Future research should continue to explore these issues and further develop the proposed framework for reducing bias in hiring algorithms. In conclusion, the recruitment and hiring process requires a proper conclusion to achieve the desired outcomes. This research paper highlights the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. While automated hiring systems are being used to detect and address discrimination against protected groups, claims of bias mitigation are rarely evaluated. Fair and equitable recruitment practices are essential to promote greater diversity and inclusion in the workforce. 8. References 1. Monedero JS, Dencik L, Edwards L. What does it mean to ‘solve’ the problem of discrimination in hiring?: social, technical and legal perspectives from the UK on automated hiring systems. FAT 2020;458-468.

LATER¶61-66

The proposed framework for fostering diversity and inclusion in the workplace is necessary to ensure that all employees are treated fairly and equitably. An effective framework should consider both the technical and legal aspects of recruitment and hiring, including laws and regulations that prohibit discrimination2. It should also include measures to prevent discrimination and ensure that every individual is given equal opportunities43. This is the only way to ensure that organizations can create a workforce that is diverse in terms of gender, ethnicity, age, and other factors44. Furthermore, organizations can use the framework to ensure that the recruitment and hiring process is transparent and fair5. This would help them create a positive work environment and foster inclusion and collaboration among employees46. A successful framework should also be capable of providing support and guidance to employees during the recruitment and hiring process7. Ultimately, the proposed framework should be comprehensive enough to provide a sense of closure and completeness to the recruitment and hiring process8. This would make sure that the process is conducted in a fair manner and that all applicants have an equal chance of being selected9. A proper conclusion should also be drawn upon the completion of the recruitment and hiring process5 so that the desired outcomes can be achieved5. This paper has presented a machine learning approach to identify five major categories of bias and discriminatory J Artif Intell Mach Learn & Data Sci | Vol: 1 & Iss: 4 Jazakallah Y., 10 language in job advertisements. We prepared a list of unique biased and discriminatory terms after examining the literature on behavioural works related to bias in recruitment. This list was used to semi-automatically generate an annotated corpus by the tagging the biased language terms (using a gazetteer-based approach) in the job advertisements of the publicly available Employment Scam Aegean Dataset, EMSCAD. This annotated corpus was used to train state-of-the-art machine learning classifiers to identify five different categories of biased and discriminatory language. We utilized a combination of linguistic features and most recent state-of-the-art word embedding representations as textual features to capture the natural language semantics of biased language. These features were fed into the machine learning classifiers. The results indicate that the Random Forest classifier with FastText word embeddings achieved the best performance with tenfold cross-validation. Overall, this work presents a major contribution in the attention phase of hiring and empowering recruiters by identifying and classifying discriminatory language in job advertisements using a machine learning-based approach. The output of this tool can be used to flag biased and discriminatory language and encourage recruiters to write more inclusive job advertisements. The findings of this research paper highlight the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. The study reveals that while automated hiring systems (AHSs) are being used to detect and address discrimination against protected groups, claims of ‘bias mitigation’ are rarely scrutinized and evaluated. The study emphasizes the importance of developing fair and equitable recruitment practices to promote greater diversity and inclusion in the workforce. The proposed framework for reducing bias in hiring algorithms aims to ensure that decision-making is done fairly and transparently, and provides support and guidance to employees during the recruitment and hiring process. However, the study also acknowledges that bias mitigation comes at a cost of efficiency and accuracy. The study suggests that ethical considerations should guide the development and implementation of AI-enabled recruitment practices, and that diverse groups should be represented in the development and testing of hiring algorithms. The study also highlights the importance of research methodology, including machine learning and statistical analysis, in providing greater accuracy and precision in results. Ultimately, this research contributes to the ongoing advancement of knowledge in the field of bias in hiring algorithms and provides valuable insights into effective strategies for mitigating bias and promoting equity in the workforce. Future research should continue to explore these issues and further develop the proposed framework for reducing bias in hiring algorithms. In conclusion, the recruitment and hiring process requires a proper conclusion to achieve the desired outcomes. This research paper highlights the critical issue of bias in hiring algorithms and the need for effective strategies to mitigate it. While automated hiring systems are being used to detect and address discrimination against protected groups, claims of bias mitigation are rarely evaluated. Fair and equitable recruitment practices are essential to promote greater diversity and inclusion in the workforce. 8. References 1. Monedero JS, Dencik L, Edwards L. What does it mean to ‘solve’ the problem of discrimination in hiring?: social, technical and legal perspectives from the UK on automated hiring systems. FAT 2020;458-468. 2. Hufthammer KT, Aasheim TH, Ånneland S, Brynjulfsen H, Slavkovik M. Bias mitigation with AIF360: A comparative study. NIK 2020;1:1-12. 3. Raghavan M, Barocas S, Kleinberg J, Levy K. Mitigating bias in algorithmic hiring: evaluating claims and practices. FAT 2020; 469-481.