Extracting Job Roles in Job Ads: A Journey with Generative AI

Mahnaz Namazi18 Jan 2024

Problem

OLX job categories had no field for the job role. The role sat inside ad titles and descriptions, so a search for a profession could not be matched cleanly to the relevant ads. The team needed a role taxonomy per job sub-category and a role for every ad, old and new. Search keyword analysis showed why: about 60% of top searched keywords in Jobs name a specific profession.

Example from the article: in the Polish "Agriculture and Gardening" category, keywords such as "tractor driver job" and "tractor operator" were grouped under the role "Tractor driver", inside the parent "Machine Operator".

Solution in brief

An LLM does two jobs. It first builds a role tree for each sub-category from roles extracted from sampled ads and from profession-related search keywords. Then it assigns each ad to a node of that tree. The extracted roles go to the search index, which uses them at the retrieval stage.

Questions this case answers
  • How to turn free-text listings into values of a controlled vocabulary so that search can match user queries to them, when the vocabulary itself does not exist yet
  • How to build a per-category hierarchical taxonomy from a sample of items and the most frequent user queries without manual labeling
  • How to add a job-role field to job ads when the taxonomy has no roles and the role is buried in titles and descriptions?
  • How to build a normalized job-role taxonomy per category from search keywords and ad text with an LLM?

Tools and methods

  • Prosus AI Assistant (LLM API) — extracts roles, labels keywords and generates the taxonomy trees
  • Prompt engineering — task prompts with a category, an output format and the keyword and role lists
  • Amazon Kinesis — stream through which a new service sends extracted roles to the search team
  • LangChain — used to chain LLM calls and specify outputs
Continue reading — it's free

Google or a one-time email code. No password.