A lot of valuable information sits in free text: survey comments, reviews, support tickets, call notes.
Getting Started
- Collect the text with useful metadata: date, product, customer type, rating.
- Clean it: remove signatures, boilerplate and duplicates; handle encoding problems.
- Read a random sample yourself before any automated analysis.
Simple Techniques
- Word and phrase frequencies, after removing common stop words.
- Keyword tagging with rules for known issues ("refund", "login").
- Comparisons: which terms are more common in negative than positive reviews?
Topic Discovery
- Clustering embeddings of comments groups them by meaning, even when wording differs.
- Topic models such as LDA find recurring themes in word patterns.
Using Language Models
LLMs can classify comments into your categories, extract specific details and summarise themes. To use them responsibly:
- define categories clearly and give examples;
- validate on a hand-labelled sample and report accuracy;
- keep quotes linked to their source comments so findings can be checked;
- follow data policies for customer text.
Presenting Findings
Combine counts ("18% of complaints mention delivery delays") with representative quotes, and show trends over time.
Cautions
Vocal minorities can dominate free text; sentiment and topic tools make mistakes; and small samples of comments don't represent all customers.