What question did this study set out to answer?

The aim is to develop a system that efficiently classifies job-related emails and extracts relevant entities.

April 11, 20260 citationsOpen Access

ApplyFlow: ATS-Aware Job Application Email Classification and Entity Extraction

Key Points

The aim is to develop a system that efficiently classifies job-related emails and extracts relevant entities.
Developed a web-based prototype called ApplyFlow.
Implemented a two-stage rule-based NLP pipeline for email classification.
Stage 1 separates application-confirmation emails from promotional emails using heuristic patterns.
Stage 2 classifies confirmed emails into eight different status labels using a weighted keyword scoring system.
Utilized an ATS-aware named entity extraction module to identify company names and roles.
Achieved 94.3% accuracy in distinguishing application-confirmation emails from promotional emails.
Classified emails into status labels with 88% weighted-average accuracy.
Extracted company names correctly with 94% accuracy, even through third-party HR platforms.
Demonstrated consistent accuracy ranging from 86-89% over a broader evaluation corpus.

Abstract

We present ApplyFlow, a web-based prototype that automates job application trackingby connecting to a user’s Gmail inbox and classifying job-related emails into actionablestatus categories. The system implements a two-stage rule-based NLP pipeline. Stage 1distinguishes personal application-confirmation emails from job-portal promotionalnoise using 30+ heuristic patterns and an ATS domain whitelist covering 30 platforms,achieving 94.3% accuracy on a 70-email evaluation set. Stage 2 classifies confirmedemails into eight status labels (Applied, Under Review, Next Step, OA/Assessment,Interview, Rejected, Offer, Follow-up Needed) using a weighted keyword scoringsystem with 60+ regex patterns, achieving 88% weighted-average accuracy on 50manually labelled emails. A key technical contribution is an ATS-aware named entityextraction module that correctly identifies company names and roles even when emailsare routed through third-party HR platforms such as Darwinbox, Greenhouse, Lever,and Workday—a challenge not systematically addressed in prior email NLP work,achieving 94% company-name extraction accuracy. Across a broader evaluation corpusof 200 labelled emails drawn from a raw pool of approximately 600 messages (bothjob-related and non-job), the two-stage pipeline yields a consistent accuracy rangeof 86–89%. We discuss design rationale, limitations, privacy considerations, anddirections for extending the approach to transformer-based classifiers.Keywords: Email Classification; Natural Language Processing; Information Extraction;Applicant Tracking Systems; Named Entity Recognition; Rule-Based NLP; Gmail API;Job Application Tracking

AI से पूछें

Bookmark

View Full Paper