Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Skill matching at scale: freelancer-project alignment for efficient multilingual candidate retrieval

Warren Jouanneau, Marc Palyart, Emma Jouffroy

TL;DR

This work tackles scalable multilingual skill matching between job proposals and freelancer profiles by introducing a structure-aware two-tower retriever built on a frozen multilingual backbone. It encodes documents at the section level, incorporates section-type awareness, and adds a document-level transformer head to produce aligned embeddings trained with contrastive losses, including InfoNCE and adjacency-based variants. The approach outperforms baselines on retrieval-quality metrics while preserving language alignment, and its production deployment in a vector-store-based pipeline yields lower latency and higher conversion for effective matches. The results demonstrate that preserving document structure and leveraging historical interactions in a multilingual setting can significantly improve scalable candidate retrieval in global marketplaces.

Abstract

Finding the perfect match between a job proposal and a set of freelancers is not an easy task to perform at scale, especially in multiple languages. In this paper, we propose a novel neural retriever architecture that tackles this problem in a multilingual setting. Our method encodes project descriptions and freelancer profiles by leveraging pre-trained multilingual language models. The latter are used as backbone for a custom transformer architecture that aims to keep the structure of the profiles and project. This model is trained with a contrastive loss on historical data. Thanks to several experiments, we show that this approach effectively captures skill matching similarity and facilitates efficient matching, outperforming traditional methods.

Skill matching at scale: freelancer-project alignment for efficient multilingual candidate retrieval

TL;DR

This work tackles scalable multilingual skill matching between job proposals and freelancer profiles by introducing a structure-aware two-tower retriever built on a frozen multilingual backbone. It encodes documents at the section level, incorporates section-type awareness, and adds a document-level transformer head to produce aligned embeddings trained with contrastive losses, including InfoNCE and adjacency-based variants. The approach outperforms baselines on retrieval-quality metrics while preserving language alignment, and its production deployment in a vector-store-based pipeline yields lower latency and higher conversion for effective matches. The results demonstrate that preserving document structure and leveraging historical interactions in a multilingual setting can significantly improve scalable candidate retrieval in global marketplaces.

Abstract

Finding the perfect match between a job proposal and a set of freelancers is not an easy task to perform at scale, especially in multiple languages. In this paper, we propose a novel neural retriever architecture that tackles this problem in a multilingual setting. Our method encodes project descriptions and freelancer profiles by leveraging pre-trained multilingual language models. The latter are used as backbone for a custom transformer architecture that aims to keep the structure of the profiles and project. This model is trained with a contrastive loss on historical data. Thanks to several experiments, we show that this approach effectively captures skill matching similarity and facilitates efficient matching, outperforming traditional methods.
Paper Structure (23 sections, 22 equations, 4 figures, 2 tables)

This paper contains 23 sections, 22 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Approach
  4. Architecture
  5. Leveraging pre-trained multilingual model : section-level context
  6. Differentiating between section : positional encoding
  7. Introducing a transformer head : document-level context
  8. From embeddings tokens to vector representation
  9. Training objective
  10. The contrastive loss approach
  11. Generalizing to n-uplet
  12. Adding weak negatives
  13. Adjacency matrix based contrastive losses
  14. Experiment
  15. Models
  16. ...and 8 more sections

Figures (4)

  • Figure 1: Illustration of matching interactions and latent space projection. (a) : Examples of positive and negative interactions between a project proposal and a set of freelancers. (b) : Illustration of documents projection and retrieval within the latent space.
  • Figure 2: Illustration of the proposed architecture for the freelancer model. Similar sections across profiles are projected into a latent space (shown in the first three squares). Positional encodings and resulting embeddings are then combined to retain section type information before entering a transformer head to process document-level context. The final step involves pooling, producing the document embedding.
  • Figure 3: Bipartite graph and sub adjacency matrices of six triplets: - two freelancer triplets ($f_{b1}, f_{b1+}, f_{b1-}$), ($f_{b2}, f_{b2+}, f_{b2-}$), - four project-freelancer triplets ($p_1, f_{b1}, f_{b1-}$), ($p_1, f_{b1+}, f_{b1-}$), ($p_2, f_{b2}, f_{b2-}$), ($p_2, f_{b2+}, f_{b2-}$)
  • Figure 4: KDE Density plots of obtained profile embeddings projected in two dimensions using T-SNE. Color encodes the family or job category associated with the profile embeddings. On the right we zoom in the web, graphic and design family.