Mitigating Language Bias in Cross-Lingual Job Retrieval - A Recruitment Platform Perspective

Sep 13, 2026

Uploaded by Gloria

Understanding textual components of resumes and job postings is critical for improving job-matching accuracy and optimizing job search systems in online recruitment platforms. Existing approaches often analyze individual components separately, relying on multiple specialized tools and reducing overall generalizability for recruitment-related text processing. A unified sentence encoder is proposed using a multi-task dual-encoder framework to jointly learn multiple components, outperforming state-of-the-art models while remaining smaller. A new metric, Language Bias Kullback-Leibler Divergence (LBKL), evaluates bias and confirms significant bias reduction and stronger cross-lingual performance.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages6

Reading Time2 min

AccessFree to download

Abstract

Introduction

Background and Related Work

Pretraining on the Job Recruitment Domain

Mitigating Language Bias in Cross-Lingual Job Retrieval: A Recruitment Platform Perspective Napat Laosaengpha 1 , Thanit Tativannarat 1 , Attapol Rutherford2 , Ekapol Chuangsuwanich 1 1 Department of Computer Engineering, Faculty of Engineering, Chulalongkorn University 2 Department of Linguistics, Faculty of Arts, Chulalongkorn University {napatnicky,[thanit.tati](thanit.tati}@gmail.com {attapol.t)[}](thanit.tati}@gmail.com {attapol.t)[@gmail.com](thanit.tati}@gmail.com {attapol.t)[ {](thanit.tati}@gmail.com {attapol.t)[attapol.t](thanit.tati}@gmail.com {attapol.t) ,[Ekapol.C](Ekapol.C}@chula.ac.th)[}](Ekapol.C}@chula.ac.th)[@chula.ac.th](Ekapol.C}@chula.ac.th) Abstract Understanding the textual components of resumes and job postings is critical for improving job-matching accuracy and optimizing job search systems in online recruitment platforms. However, existing works primarily focus on analyzing individual components within this information, requiring multiple specialized tools to analyze each aspect. Such disjointed methods could potentially hinder overall generalizability in recruitment-related text processing. Therefore, we propose a unified sentence encoder that utilized multi-task dual-encoder framework for jointly learning multiple component into the unified sentence encoder. The results show that our method outperforms other state-of-the-art models, despite its smaller model size. Moreover, we propose a novel metric, Language Bias Kullback-Leibler Divergence (LBKL), to evaluate language bias in the encoder, demonstrating significant bias reduction and superior cross-lingual performance. Introduction The online job recruitment platforms have emerged as essential tools to streamline and accelerate the talent acquisition process. These platforms usually rely on essential details from resume and job posting to facilitate the connection between recruiters and job seekers. The resumes showcase personal overview of candidate’s capabilities including work experience, skills and expertise, whereas the job postings outline job title, specialties and responsibility for open positions. Therefore, understanding the semantic meaning of the textual information within resumes and job postings would greatly facilitate the matchmaking process on the online job recruitment platform, where the underlying process is an automatic job recommendation system (Zhao, Wang et al. 2021) . Additionally, this understanding could be further developed into analytical tools for various job-related tasks such as job mobility prediction (Zha, Sun et al. 2024) and job demand forecasting (Lu 2022) . Previous approaches to understanding this textual information as low-dimensional dense sentence representations primarily focus on a individual component within the resumes and the job posting such as job title (JT) (Decorte, Hautte et al. 2021; Laosaengpha, Tativannarat et al. 2024), Copyright © 2025, Association for the Advancement of Artificial Intelligence ([www.aaai.org](www.aaai.org)). All rights reserved. job description (JD) (Goyal, Kalra et al. 2023) and skills set (Lin et al. 2023) . However, learning sentence representations in these previous works are constrained to handling the individual component, which limit their versatility in being applied to analyze the other job-related components. Moreover, the existing works often prioritize the study on a mainstream language like English, leaving non-English language especially in low-resource ones under-explored. It might be due to the scarcity of public dataset and the specialized nature of the job recruitment domain for non-English language. Additionally, their studies mainly cover monolingual setting, which might not applied to regions where the users often alternate between languages on the platform. In this paper, we propose a multi-task learning framework to develop a bilingual sentence encoder (Thai and English) for general-purpose use in the recruitment domain by u

FAQs

Why does the paper focus on language bias in cross-lingual job retrieval?

Language bias in the encoder can limit retrieval quality when job seekers and job postings use different languages. The paper introduces a metric to quantify this bias and reduce it through a unified training framework.

What model does the paper propose for learning representations from resumes and job postings?

It proposes a unified sentence encoder trained with a multi-task dual-encoder framework that jointly learns multiple job-related components for Thai and English.

What are the three job-related training tasks used in the framework?

The tasks are (A) job title translation ranking, (B) job description–title matching, and (C) job field classification.

Sep 13, 2026

Uploaded by Gloria