How resume parsing actually works

Before software can score a CV, it has to read it. That sounds simple. It is the step that goes wrong most often, and almost nobody checks it.

What "reading a CV" actually means

A CV is designed to be looked at by a person. Columns, boxes, headings, icons. Software needs something different: a list of facts. Name here, employers here, dates here.

Turning the first thing into the second is called parsing. It happens in seven steps, and each one can go wrong.

The seven steps

1. Open the file. PDF, Word, sometimes a photo saved as a PDF.
2. Get the text out. Most PDFs already contain text. Scans do not — those are pictures, so the computer has to recognise the letters from the image.
3. Work out the reading order. A PDF does not store sentences. It stores bits of text with positions on the page. The software has to guess what follows what.
4. Find the sections. Where does work history end and education begin?
5. Pull out the facts. Employers, job titles, dates, skills.
6. Tidy them up. "Sr. SWE II" and "Senior Software Engineer" become one thing. Eleven date styles become one.
7. Save them. Put it all into the fields your system actually has.

Where each step breaks

Getting text out of scans. Blurry scans, unusual fonts and coloured backgrounds all cause mistakes. The errors are usually small — one wrong letter in an email address. That is worse than a big obvious error, because nothing flags it.

Reading order. This is the most common failure, because most modern CV templates have two columns. The software has to decide whether text on the left continues on the right or below. Get it wrong and the skills sidebar gets mixed into the job history. The result reads like nonsense, one line at a time.

Finding sections. This depends on spotting headings. It fails when someone writes "Where I've Been" instead of "Experience", or uses a picture as a heading, or has no headings at all. Then a whole work history can land in the summary field, where nothing will look for it.

Pulling out facts. Struggles with anything that is not a straight line: freelance work, two jobs at once, a promotion inside one company. You get duplicate jobs, missing jobs, or one job that swallowed three.

Tidying up is where dates go wrong, and dates matter more than any other field:

Saving them. This one fails in silence. If the software pulls out something your system has no field for, it is thrown away. No error. The data was right until the moment it vanished.

Some fields are much harder than others

Why "99% accurate" means nothing

Accurate on which field? Measured how? On whose CVs?

Pulling out an email address is nearly a solved problem and scores brilliantly. Pulling out dates from a freelancer's CV does not. A single accuracy number has mixed an easy field with a hard one, and the easy fields win because there are more of them.

The number you care about is accuracy per field, on CVs that look like the ones you actually receive. You will measure that yourself in the next lesson. It takes half an hour and is worth more than any vendor's figure.

Test any tool in ten minutes

1. Give it a two-column CV. Did it read in the right order?
2. Give it a scan. Did the name and email survive?
3. Give it a freelancer with overlapping jobs. Count the jobs it found.
4. Give it a CV with dates written the other way round. Check the total years.

Four documents will tell you more than an hour-long demo.