TECH

Research · 2025

Online Employee Reviews

Jupyter NLP Topic modeling

Employee review sites contain the thing every HR function claims to want — candid, unprompted, longitudinal accounts of what working somewhere is actually like. The catch is that it arrives as unstructured prose written by anonymous people with an axe to grind, and turning that into a defensible finding is a research problem, not a dashboard problem.

The problem

Airlines run on shift work, tight margins and a workforce split across radically different roles — cabin crew, ground handling, engineering, corporate. Aggregate satisfaction scores flatten all of that into a single meaningless number.

The free-text reviews contain the actual signal, but they are self-selected. People who write reviews are disproportionately delighted or furious, which means the raw sentiment distribution says more about who writes reviews than about the employer.

The approach

Topic modeling first, sentiment second. Knowing that reviews are 60% negative is useless. Knowing that the negativity concentrates on rostering and management communication while pay is comparatively neutral is actionable, because those are different departments with different fixes.

Serious preprocessing, because domain vocabulary breaks general-purpose tooling. Industry jargon, abbreviations and role-specific shorthand are exactly the terms that carry the most meaning, and a generic pipeline discards them as noise.

Everything published as notebooks so the analysis can be re-run and disagreed with. An HR finding that cannot be reproduced is an opinion with a chart attached.

What was hard

Selection bias cannot be engineered away, only reasoned about. The honest treatment is to state clearly what the corpus can and cannot support: it is good evidence about which themes recur among motivated reviewers, and poor evidence about population-level satisfaction rates. Most commercial text analytics quietly ignores this distinction.

Topic models also require judgment rather than a metric. Coherence scores will happily recommend a number of topics that produces incomprehensible clusters, and the person doing the analysis has to read the output and decide whether the topics mean anything.

Where this applies

The same method works on any large free-text corpus where the decision-maker needs themes rather than a score: support tickets, exit interviews, product feedback, open survey responses.

This is what the Data → Insight Study engagement delivers — a bounded piece of research with a written finding and notebooks you can re-run, rather than a BI tool nobody opens twice.

Related work

Research

Dyadic Reaction Emotion Database

A toolkit for collecting facial expressions via webcam while a participant watches video stimuli, then segmenting the recordings per stimulus. Built so other labs can reproduce the method.

Audio

SpineSound

Companion software for three-class spine sound recognition. A small-data audio classification problem with calibrated confidence. Research use only — explicitly not a diagnostic device.

Similar problem?

Three sentences on what you are trying to do is enough for us to tell you whether this is the same shape of problem.