Text data is written language treated as data. It appears in reviews, surveys, chats, social media posts, blogs, and call-center conversations converted to text. These sources can reveal customer opinions, service quality, and business-process issues, but their unstructured form is not ready for numerical algorithms. section 7.2 section 7.0
Start by inspecting the text, not by selecting a model. Junk characters, URLs, abbreviations, misspellings, multiple languages, and changing vocabulary complicate cleaning. Meaning also depends on context: two differently worded questions can ask the same thing, while one word can have different meanings in different sentences. section 7.4 section 7.6
A practical path from raw language to analysis is:
- Clean selectively: remove unwanted characters or low-value words, and consider stemming or lemmatization. Keep stop words when removing them would erase needed context. section 7.6 section 7.20
- Tokenize the cleaned text, then represent it numerically. Bag-of-words counts terms but ignores order and context; TF-IDF weights terms by importance. Word2Vec and GloVe provide compact representations that retain contextual relationships. section 7.20
- Use the prepared data for tasks such as sentiment analysis, document categorization, or text clustering. Compare preprocessing and representation choices rather than assuming one will work best. section 7.0 section 7.18
Customer feedback
A product review can contain both a star rating and free-text feedback. The written response offers detail a rating alone cannot convey; before analyzing it, check whether cleaning would discard words that matter to its meaning. section 7.2 section 7.6
