job-title-normalizer-e5-small

The lightweight variant of job-title-normalizer-e5-base: a 118M-param sentence encoder fine-tuned to normalize messy, multilingual job titles to a canonical occupation taxonomy (ESCO + O*NET), framed as retrieval over 4,055 canonical occupations. ~2.5× smaller and faster than the e5-base variant at a modest accuracy cost — the right choice for CPU serving and bulk backfills.

Live demo: https://job-title-normalizer-525186107937.us-central1.run.app (Cloud Run; may cold-start ~1 min)

Results

Held-out test set of 15,248 real ESCO/O*NET titles; split by occupation (zero train/test occupation overlap, asserted at build time).

Slice	Method	Recall@1	Recall@5	Recall@10	MRR
Overall	BM25 (lexical)	0.095	0.163	0.183	0.124
Overall	zero-shot e5-small	0.226	0.371	0.437	0.286
Overall	this model	0.367	0.539	0.601	0.441
FR→EN	this model	0.525	0.737	0.799	0.618
DE→EN	this model	0.523	0.745	0.815	0.619

BM25 scores MRR 0.034 cross-lingually; this model reaches 0.619. For the best accuracy, use the e5-base variant (overall MRR 0.463 vs 0.441).

Training

• Base: intfloat/multilingual-e5-small (mean pooling, query:/passage: prefixes, 384-dim).
• Objective: in-batch-negatives InfoNCE (MultipleNegativesRankingLoss, scale 20), wrapped in MatryoshkaLoss (dims 384/256/128/64).
• Data: 160,240 positive pairs (synonyms, alternate titles, cross-lingual label pairs) from ESCO (EN/FR/DE) + O*NET alternate titles; ≤50 pairs per occupation.
• Setup: 2 epochs, batch 128, lr 2e-5, warmup 10%, max_seq_len 64, fp16, NoDuplicatesDataLoader. Trained on an RTX 4060 Laptop (8 GB) in ~8 min.

How to use

The e5 family needs asymmetric prefixes: titles as query: …, canonical labels as passage: …. Forgetting them silently degrades accuracy.

from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

model = SentenceTransformer("Misbahuddin/job-title-normalizer-e5-small")

canonicals = [
    "passage: software developer",
    "passage: data scientist",
    "passage: nurse responsible for general care",
]
query = "query: Krankenpfleger"              # German → English canonical

q = model.encode(query, normalize_embeddings=True)
c = model.encode(canonicals, normalize_embeddings=True)
scores = cos_sim(q, c)[0]
print(canonicals[int(scores.argmax())], float(scores.max()))

Vectors are L2-normalized (cosine == dot product) — use FAISS IndexFlatIP for production. Matryoshka truncation to 256/128/64 dims works after re-normalizing. Calibrate an abstention threshold: out-of-taxonomy queries return low-score nearest neighbours rather than failing.

Intended use & limitations

• In scope: normalizing titles from resumes, postings, CRM/ATS/HRIS records to occupation IDs; semantic occupation search; FR/DE→EN cross-lingual lookup.
• Not a hiring/screening/compensation decision system. Keep a human in the loop for consequential uses.
• Coverage bounded by ESCO + O*NET; en/fr/de verified only. Short-text regime (≤64 tokens).

Acknowledgements

• ESCO — © European Union; reused under Commission Decision 2011/833/EU. https://esco.ec.europa.eu/
• O*NET — by the National Center for O*NET Development for USDOL/ETA; CC BY 4.0. https://www.onetcenter.org/

ONET® is a trademark of USDOL/ETA. This model was produced using ONET data but is not endorsed by USDOL/ETA. ESCO is a service of the European Commission; this model is not endorsed by the Commission.