What do customers mention most often? Frequency analysis counts how often words or phrases occur across their reviews. It offers a faster first look at recurring topics than reading a few reviews, which may not represent the whole collection. Counts point to subjects worth investigating; they do not, by themselves, establish what customers meant. section 5.3 section 5.3.1
For a collection of reviews, use a repeatable workflow:
- Ask a generative AI assistant for Python code, naming the DataFrame, review column, and language. Check any filename it supplies against your actual file. section 5.2.2 section 5.3.4
- Tokenize the text, lowercase it, discard punctuation, remove language-appropriate stopwords, and count the remaining words. Run and inspect the generated code rather than assuming its language settings are correct. section 5.3.4
- Read the leading words alongside their counts; examine frequent phrases when individual words lack context. Translate unfamiliar terms before summarizing them. section 5.3.6
A small Portuguese review check
A direct request to count words in pasted reviews produced frequent articles and conjunctions such as “a” and “e.” Those counts reflected grammar more than customer concerns. Repeated content words including “prazo” (deadline) and “aparelho” (device) offered more useful leads about delivery and products. A follow-up request can refine an unhelpful first result. section 5.3.2
('chegou', 'antes', 'prazo') - (arrived, before, deadline)
('bem', 'antes', 'prazo') - (well, before, deadline)
('entregue', 'antes', 'prazo') - (delivered, before, deadline)
('produto', 'chegou', 'antes') - (product, arrived, before)
('entrega', 'antes', 'prazo') - (delivery, before, deadline)
('chegou', 'bem', 'antes') - (arrived, well, before)
('produto', 'entregue', 'antes') - (product, delivered, before)
('entrega', 'super', 'rápida') - (delivery, super, fast)
('antes', 'prazo', 'previsto') - (before, deadline, expected)
('produto', 'ótima', 'qualidade') - (product, great, quality)
The translated three-word combinations give a more specific lead than isolated counts: several mention arrival or delivery before the deadline. Check the underlying reviews before treating that pattern as a conclusion about all deliveries. section 5.3.6
For reporting, a word cloud can make common terms easy to scan, but keep the counts and translations available. Spelling, punctuation, capitalization, and other noise can distort text analysis; a generated tokenizer may also ignore a stated language. Validate promising themes against raw reviews. Where review scores are available, they offer a separate check on claims about satisfaction. section 5.3.6 section 5.0 section 5.3.4 section 5.2.2
