> For the complete documentation index, see [llms.txt](https://docs.tabiya.org/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.tabiya.org/ethiopia/localized-taxonomy.md).
# Localized Taxonomy
A job seeker in Addis Ababa might describe their experience in Amharic, in informal and colloquial terms, drawing on work that never carried a formal title. On the other hand, employers may post a vacancy that potentially describes similar work but with a different vocabulary. Good labor-market matching has to recognize the full universe of skills that work can be described in. Building on the experience in South Africa and Kenya, Tabiya supported the E-LMIS team at MoLS to map noisy, inconsistent occupation labels onto a clean, standardized occupational framework for Ethiopia, aligned with international standards through localization of the Inclusive Livelihoods Taxonomy .
The outcome of this effort is a “gold standard” of roughly 9,000 canonical occupational positions distilled from more than 130,000 raw historical entries, drawn from ESCO, ISCO, ILO, and the Ethiopian Civil Service, each mapped to skill tags and competency vectors. The sections below walk through how that was done, from raw data to a finished catalog.
What did Tabiya and the E-LMIS team set out to build?
The E-LMIS team compiled a dataset of raw occupations sourced from several sources that it wanted to anchor the national Ethiopian taxonomy on. The teams processed these raw vacancies into a clean, standardized catalog through a multistep taxonomy localization process. The main tool used for this effort was the Tabiya's [Livelihoods Classifier,](/our-tech-stack/livelihoods-classifier.md) which reads free-text occupation labels and maps them to standardized occupations and skills on Tabiya's ESCO-based Inclusive Livelihoods Taxonomy.
Step 1: Sourcing the vacancy dataset for taxonomy localization
The work drew on two separate streams, kept apart during collection so that operational, user-generated data would not contaminate the authoritative reference data.
1. **E-LMIS operational dataset:** roughly 131,000 raw occupation records that users had entered through different workflows, containing duplicate entries, typographical inconsistencies, departmental and company-specific names, experience-level prefixes, and multiple naming variations for the same occupation. An initial round of cleaning and deduplication reduced these to about 9,000 unique occupation labels.
2. **Reference dataset of \~9,700 records from authoritative sources**: roughly 700 ISCO-aligned global occupations from the International Labour Organization (ILO), about 1,000 occupational descriptors and skill structures from the O\*NET Resource Center, and around 8,000 localized public-sector occupation structures from the Ethiopian Civil Service.
The two streams were merged into an intermediate pool of about 18,000 records. A further pass removed duplicate titles, structural overlaps (the same role written, ordered, or phrased differently), invalid records, and semantically redundant labels leaving roughly 8,000 cleaned and standardized occupation titles to use for taxonomy alignment and recommendation processing.
Step 2: Pre-processing and cleaning the vacancy dataset before classification
Before alignment, a Large Language Model (LLM)-assisted normalization layer was introduced to rewrite noisy, user-entered titles into the core occupational concept they represented. This step removed organization and workplace names (for example, “Administrator at Terra LAB” became “Administrator”), stripped spoken-language descriptors (“Afan Oromo and Tigrigna Translator” became “Translator”), consolidated variations of the same role (“Administrative Assistant Officer” became “Admin Assistant”), separated titles that bundled two jobs into one (“Accountants and Auditors” became separate “Accountant” and “Auditor” entries), and removed seniority or experience prefixes (“Junior Accountant” became “Accountant”). The resulting dataset trimmed noise and retained consistent occupation names for similar clusters of work.
Step 3: Mapping locally derived occupation titles to a standard taxonomy
The cleaned titles were first mapped using the Classifier on to the Tabiya (ESCO 1.1.1) v2.0.1-rc.1 taxonomy base. A named entity linking (NEL) step matched each cleaned title to a taxonomy occupation using semantic similariy and assigning a match only when confidence passed a 0.80 similarity threshold. The threshold was selected after several rounds of iterative human review of Classifier outputs. This step left about 2,000 low-similarity titles below the cutoff. For high similarity, non-exact matches, validated local occupation titles were added to the taxonomy as alternative labels, improving occupational searchability and localization.
*Note: Amharic labels were excluded from this synchronization cycle to keep formatting consistent across taxonomy structures.*
Step 4: Human review and expanding the taxonomy - pending post-pilot
As of July 2026, the taxonomy localization process is pending a targeted manual review of the roughly 2,000 unmatched/low match score titles, ensuring full dataset coverage. This manual review process will include additions to the occupations' repository beyond simply mapping to existing occupations, and explore methodologies to validate the skills associated with localized Ethiopian occupations.
The taxonomy is the foundational layer that will power the end-to-end jobseeker and vacancy pipeline on the national employment portal. It currently standardizes Ethiopia's labor market data at the source and powers skills-based matching, so that the resulting data can add up to a timely picture of labor supply and demand that the government can use for planning and policy.
---
# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.
## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.
Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:
```
GET https://docs.tabiya.org/ethiopia/localized-taxonomy.md?ask=&goal=
```
`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.
Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.