JoTA: Aligning Multilingual Job Taxonomies through Word Embeddings (Student Abstract)

Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario MezzanzanicaAAAI 2022 · DOI

Abstract

JoTA (Job Taxonomy Alignment) is a framework designed to automate the mapping between disparate job and skill taxonomies, such as national classification systems and international standards. By integrating hierarchical semantic similarity with neural network classification, JoTA achieves significant improvements in alignment accuracy, reaching 80% in top-5 recommendations. This system serves as a robust tool for reducing the manual workload of domain experts in labor market analysis.

Introduction

In the modern labor market, taxonomies are critical for organizing jobs, skills, and qualifications. However, these systems are rarely universal; different countries and organizations maintain distinct "lexical taxonomies"—hierarchical structures where concepts are linked by "is-a" relationships.

When researchers or policymakers need to compare different labor markets—for instance, mapping the Italian national taxonomy (CP) to the European standard (ESCO)—they face a complex mapping problem. Manually linking specific job titles across systems is labor-intensive and requires deep domain expertise. JoTA addresses this by automating the alignment process using distributed semantics, leveraging the principle that words appearing in similar contexts share similar meanings.

The JoTA Framework

The JoTA workflow is structured into three distinct phases designed to move from raw data to expert-validated alignments.

The three-step JoTA workflow for taxonomy alignment, from corpus generation to expert validation.

1. Embedding Generation and Evaluation

The process begins by using FastText to create vector representations (embeddings) of taxonomic terms. To ensure high-quality embeddings that reflect the specific needs of taxonomy work, the framework employs a metric called Hierarchical Semantic Similarity (HSS). This metric evaluates how well the word vectors preserve the structural hierarchy of the taxonomy, ensuring that the embeddings respect the organizational logic of the job market rather than just general linguistic patterns.

2. The Alignment Method

JoTA utilizes a "Blended" approach that combines two complementary strategies:

  • Hierarchical Approach: This calculates the cosine similarity between the leaf terms of the origin taxonomy and the concepts in the destination taxonomy.
  • Classification Approach: A 2-layer neural network is trained to treat alignment as a multi-class classification problem, predicting the most likely destination concept based on input word embeddings.
  • Blending: The results from both methods are merged. The system prioritizes classification results—which generally prove more accurate—while using hierarchical suggestions to fill gaps.

3. Expert Validation

The system generates a ranked list of the top $n$ suggestions (typically 5). A human domain expert then reviews these suggestions to make the final selection, significantly narrowing the search space compared to manual mapping.

Key Insights and Methodology

Structure-Aware Embeddings

A core innovation of JoTA is the realization that standard word embeddings may not be sufficient for taxonomy alignment. By using HSS to optimize hyperparameters (such as the CBOW algorithm and 150-dimensional vectors), the authors ensure the vectors are "structure-aware."

Weighted Mean Reciprocal Rank (wMRR)

To better evaluate performance, the authors utilize Weighted Mean Reciprocal Rank (wMRR). In taxonomy alignment, suggesting a "parent" concept (e.g., "Computing Professional" instead of "Software Developer") is more useful than an unrelated suggestion. The wMRR provides partial credit for these hierarchical "near misses," reflecting the practical utility of the tool for experts.

Experimental Results

The framework was evaluated by mapping the Italian CP taxonomy to the European ESCO taxonomy. The results demonstrate that the hybrid approach is superior to standalone methods.

Performance comparison of the Hierarchical, Classification, and Blended approaches across accuracy and ranking metrics.

As shown in the results, the Blended approach reached 80% accuracy for top-5 recommendations. This suggests that semantic closeness and learned classification patterns capture different, complementary aspects of taxonomic relationships, leading to a more robust alignment tool.

Conclusion and Future Work

While JoTA significantly reduces the manual effort required for taxonomy mapping, it remains a "human-in-the-loop" system requiring final expert validation. Future research will focus on testing the framework's domain independence by applying it to non-labor market taxonomies (e.g., medical or biological) and expanding its multilingual capabilities to handle cross-lingual embeddings where origin and destination taxonomies are in different languages.

Summary updated

BibTeX
@inproceedings{giabelli2022jota,
  author = {Anna Giabelli and Lorenzo Malandri and Fabio Mercorio and Mario Mezzanzanica},
  title = {{JoTA: Aligning Multilingual Job Taxonomies through Word Embeddings (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21614},
}
Figure