title_feature

title_feature is a fine-tuned multi-label sequence classification model used to extract attributes and features embedded within job posting titles (such as employment type).

Basic Usage

Because this model acts as a multi-label feature extractor rather than a single-class classifier, passing its logit outputs through a sigmoid function with a confidence threshold (e.g., > 0.98) is recommended for pulling out multi-label attributes.

import torch
import numpy as np
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "loyoladatamining/title_feature"
model = AutoModelForSequenceClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Sample input title from a job posting
title = "Remote Senior Software Engineer Full-Time"

# Tokenize and infer
inputs = tokenizer(
    title, 
    truncation=True, 
    max_length=32, 
    padding="max_length", 
    return_tensors="pt"
)

with torch.no_grad():
    outputs = model(inputs.input_ids)
    logits = outputs.logits

# Apply thresholding for multi-label retrieval
for l in logits:
    probabilities = torch.sigmoid(l.cpu()).numpy()
    active_indices = np.where(probabilities > 0.9)[0]
    
    # Map index integers back to the feature labels
    extracted_features = sorted([model.config.id2label[idx] for idx in active_indices])
    
    # Optional handling for when "none" is predicted along with others
    if "none" in extracted_features and len(extracted_features) > 1:
        extracted_features = [x for x in extracted_features if x != "none"]
        
    print(f"Extracted Features: {';'.join(extracted_features)}")

Output Format

Following the above example, extracted_features would then read:

FT;Rem

This denotes a full-time position (FT) and remote work eligible (Rem).

Citation

If you find this model useful in your work, please consider citing:

@article{meisenbacher2025extracting,
  title={Extracting O* NET Features from the NLx Corpus to Build Public Use Aggregate Labor Market Data},
  author={Meisenbacher, Stephen and Nestorov, Svetlozar and Norlander, Peter},
  year={2025}
}