Skills application solution
Skills application solution
Use this expert example to compare how you selected text features, explained the process and translated your findings into practical recommendations.

Trustpilot report: Understanding and avoiding 1-star reviews
Key terms linked with 1-star reviews
- “number there” suggested frustration around accessibility or the inability to reach support.
- “line day” pointed to long waits, queues and poor service efficiency.
- “that person” indicated negative interactions with an individual staff member or representative.
- “weekend call” highlighted out-of-hours support gaps and poor weekend responsiveness.
How the team obtained these insights
- Text preparation: Reviews were collected and lemmatised so related word forms were standardised.
- Vectorisation: A
CountVectorizerwith anngram_rangeof(1, 2)captured both single words and two-word phrases. - Filtering: A
max_dfthreshold of 0.8 removed overly common, low-value terms. - Model training: Logistic regression was trained to predict whether a review was 1-star.
- Evaluation: The team compared the model against a baseline and then used the learned coefficients to identify the most predictive terms.
- Baseline accuracy: 0.696
- Training accuracy: 0.99
- Test accuracy: 0.93
Why this worked
The combination of lemmatisation, n-grams and an interpretable linear model made it possible to trace poor review outcomes back to specific phrases, not just overall sentiment.
Recommendations for reducing 1-star reviews
- Improve accessibility: Make contact routes obvious and provide multiple support channels.
- Reduce waiting time: Streamline service processes, improve queue handling and communicate delays clearly.
- Strengthen customer interactions: Train staff in empathy, product knowledge and issue resolution.
- Extend out-of-hours support: Improve weekend support coverage and self-service resources.
Tips for applying this skill in your role
- Experiment with multiple NLP representations because the best choice depends on the task.
- Balance richer transformations with the extra compute they require.
- If a deployed language model depends on a fixed vocabulary, refresh it periodically so it keeps up with new phrases and behaviours.