Resume Parsing

How a resume parser works

A parser takes a file (PDF, Word, plain text, sometimes an image) and turns it into data in a few stages:

1. Text extraction. Pull raw text out of the file. Scanned PDFs and photos need optical character recognition first, which is where many errors start.
2. Layout analysis. Work out reading order, columns, sections and headings. Two-column designs, text boxes and tables confuse weaker parsers.
3. Section and entity recognition. Identify which text is a job title, an employer, a date range, a degree or a skill. Older parsers used keyword lists and rules; modern ones use machine learning and language models.
4. Normalization. Map 'Sr. CSR', 'Customer Service Rep II' and 'Customer Care Associate' to a standard job title, convert 'Jan 22 - present' to dates, and map skills to a skills taxonomy.
5. Output. Structured data (usually JSON or XML) that fills the candidate record in the applicant tracking system.

The output is what downstream tools rely on. Search, matching, duplicate detection and reporting are only as good as the parse.

Resume parsing vs resume screening vs matching

These get blurred in vendor marketing, but they are separate steps:

• Parsing extracts data. It makes no judgment about the candidate.
• [Resume screening](/glossary/resume-screening/) decides whether a candidate meets requirements and should move forward.
• [Candidate matching](/glossary/candidate-matching/) scores or ranks candidates against a job, often using parsed data plus job requirements.

The distinction matters legally. A parser that only fills fields is low-risk. Once parsed data feeds an automated score or ranking that affects who gets interviewed, you have a selection procedure that can be audited for adverse impact, and in New York City it may be an automated employment decision tool.

Where parsers go wrong

Even good parsers make predictable mistakes:

• Creative layouts. Columns, graphics, icons for phone and email, and text in headers or footers get skipped or scrambled.
• Date confusion. 'Summer 2023', seasonal roles, overlapping jobs and missing months produce wrong tenure calculations, which then make candidates look like job hoppers.
• Job title noise. Internal titles like 'Associate III' or 'Happiness Hero' parse correctly as text but tell the system nothing useful.
• Non-US formats. International CVs with photos, date of birth, different date order or non-English sections parse less accurately.
• Skills over-extraction. A line saying 'trained new hires on Salesforce' may tag the candidate with a CRM skill they used once.
• Name and contact errors. Hyphenated names, non-Latin scripts and multiple phone numbers get mangled, and then duplicate profiles appear.

For hourly roles, there is an extra problem: many good candidates do not have a resume at all. Requiring one at the application stage can cut your application completion rate badly. A short mobile application with a few questions usually works better, with resume upload optional.

How to test a parser before you buy

Do not rely on the vendor's accuracy claims. Run your own test:

1. Pull 100 real resumes from recent applicants for your main role families, including messy ones: phone photos, two-column templates, international CVs, career changers.
2. Have the parser process them, then have a recruiter check key fields: name, email, phone, most recent title and employer, dates, highest education, and five required skills.
3. Score field-level accuracy. A common benchmark is that contact fields should be nearly perfect and most recent job details right in the large majority of cases; skills tend to be the weakest field.
4. Check whether errors fall unevenly on certain groups, such as candidates with non-English names, older candidates with long histories or people with employment gaps.
5. Confirm the candidate can review and correct parsed data before submitting. That single feature fixes many errors at the source.

Data privacy notes

Parsed resumes are personal data. Under GDPR, UK GDPR and California's CCPA as amended by the CPRA, candidates have rights over their data, including access and, in many cases, deletion, and you need a lawful basis and clear notice for processing. Do not parse and store data you have no reason to keep, such as date of birth, photos or marital status on international CVs. Set a candidate data retention period and apply it to parsed fields and source files alike. In the US, federal recordkeeping rules require keeping application records for at least one year (two for federal contractors with 150 or more employees), so deletion schedules have to respect those minimums. Rules vary by state and country.