Back to Blog

Developing a Large-Scale Language Model to Unveil and Alleviate Gender and Age Biases in Australian Job Ads

November 15, 2024
Papers  /
Summary  /
Featured image for “Developing a Large-Scale Language Model to Unveil and Alleviate Gender and Age Biases in Australian Job Ads”

Introduction: Job Advertisements in Australia

Job ads are a crucial component of the hiring process, serving as the first point of contact between employers and potential employees. They provide essential information about job roles, responsibilities, qualifications and company culture. A well-crafted job ad attracts qualified candidates and builds a positive employer brand, while a poorly written ad can deter candidates and perpetuate bias.

In our latest journal article, Ruochen Mao, Liming Tan, Rezza Moieni and Nicole Lee examine the current situation of recruitment advertisements in Australia. This study investigates gender and age biases in Australian job advertisements, focusing on developing bias detectors using large-scale language models (LLMs). Research has found that gender-equal language and Artificial Intelligence (AI) techniques can be used to detect and reduce bias, as well as improve female candidates’ appeal.

Background: Gender and Age Discrimination in Job Ads

Gender and age biases are prevalent in job advertisements in Australia, impacting job seekers’ perceptions and opportunities. For example, job ads that use words like “strong” or “aggressive” tend to attract more male applicants, while words like “supportive”, or “collaborative” are associated with female applicants. Similarly, phrases like “recent graduate” or “dynamic and energetic” can discourage older job seekers from applying. These biases not only reinforce gender stereotypes and age discrimination, but also limit the diversity and inclusivity of the workforce.

Our previous research found that the Victorian Government has been using more gender-neutral and inclusive terminology over the years. However, gendered and age-related language in job ads continues to influence how candidates perceive job opportunities. Age bias often restricts opportunities for older job seekers, and discrimination in job ads can be both explicit and subtle. Gender-coded words also affect applicant perceptions and industry gender equality. Various methods, including lexicon-based and similarity-based approaches, are used to detect these biases.

Methodology: Develop Bias Detector using RoBERTa, ALBERT, and GPT-2

Data Collection

Data for this study were collected from a publicly available Australian job website. The researchers selected seven popular industry categories for analysis:

  • Advertising, Arts & Media (1.0%)
  • Education & Training (4.9%)
  • Healthcare & Medical (9.3&)
  • Hospitality & Tourism (13.5%)
  • Information & Communication Technology (21.1&)
  • Manufacturing, Transport & Logistics (33.5%)
  • Trades & Services (16.8%)

A total of 21,683 job ads were analysed using a program developed with Python’s BeautifulSoup library. In particular, 40 job advertisements in the IT industry were identified as being biased towards men.

Furthermore, ChatGPT was utilised to generate new job advertisement texts based on two feminist-oriented natural language generation approaches: “adhering” and “steering.” The adhering approach used neutral language, while the steering approach used language biased towards femininity to attract more female professionals.

In addition, a total of 410 distinct workers from Amazon Mechanical Turk (MTurk) participated in the study, with 151 completing the responses. And each job advertisement was rated by at least three workers to ensure diverse feedback.

Bias Detector Development

We aim to develop a method to detect subtle gender and age biases in textual descriptions and visually present these differences. To achieve this goal, this study focused on developing both gender and age bias detectors using three language models:

For gender detection, an end-to-end approach was used, while for age detection, a hybrid approach combining end-to-end and lexicon-based methods was employed. These models were fine-tuned on datasets specifically addressing gender and age biases.

The gender bias data were derived from an annotation of the “Annotated Text Corpus of Gender Bias“. Age-related stereotypical terms for both young and old individuals were initially gathered from relevant research papers, articles, and blog posts. Subsequently, annotated datasets for gender and age biases were created using a combination of crowd annotation, rule-based annotation and ChatGPT.

Two classifiers were developed—a male language classifier and a female language classifier to assess the probability of alignment with gender stereotypes and the probability of contradiction. RoBERTa, ALBERT, and GPT-2 were used to build corresponding classifiers for male language, female language, older people language, and younger people language, with RoBERTa showing the best performance in detecting biases.

Findings: Gender and Age Bias in Different Industries

The overall bias value for men was 0.64 and for women was 0.51. Gender bias was found to be significantly higher in certain industries, such as IT, Manufacturing, Transport & Logistics and Trades & Services. And there was a preference for younger job seekers in job ads across all industries.

In addition, the study examined the steering version (femininity-biased language) of job advertisements and the adhering version (neutral language). Female job seekers were most interested in the steering version, while gender bias was reduced most effectively by the adhering version. These findings suggest that using neutral language reduces gender bias and attracts both male and female job candidates.

Conclusion

We reveal significant gender and age biases in different industries, with male bias more prevalent than female bias in certain sectors like IT and manufacturing. The research also shows the majority of advertisements lean towards younger candidates. Moreover, the study developed a gender and age bias detector using LLMs and exploring a ChatGPT-based natural language generation method to reduce gender bias.

Nevertheless, the study has limitations, including dataset size, and the authors suggest exploring other text generation models in future research. Future work should also focus on reducing age bias in job advertisements.

In conclusion, this research provides valuable insights into gender and age biases and highlights the potential of LLMs for creating more inclusive job advertisements.

About the author:

Share this Post

If you'd like to get in touch and experience a one on one demo of Diversity Atlas.

Book a Demo