---

name: text-wordclouds description: "Use AutoViz text-column cleaning and wordcloud behavior for NLP-style tabular fields." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Text Wordclouds

Use this sub-skill when the user asks about AutoViz behavior for string columns, discrete text columns, NLP variables, wordcloud plots, text cleanup, emoji handling, stopwords, or NLTK downloads.

Use this when

• The prompt names wordcloud, AutoViz_NLP, draw_word_clouds, nltk, textblob, emoji, stopwords, contractions, or text cleanup.
• AutoViz detects string variables and attempts to generate wordclouds.
• The user needs to understand why a text column did or did not get a wordcloud.
• The user is debugging missing NLTK data, missing wordcloud, or text preprocessing side effects.
• The user wants to know how AutoViz distinguishes categorical strings from NLP-style strings.

Core flow

1. Start from the DataFrame and identify candidate string columns.
2. Let AutoViz classify variables first; only discrete/NLP string variables reach the wordcloud branch.
3. For each selected string column, AutoViz cleans text and calls wordcloud generation.
4. If a target column is present and the problem type is classification, wordclouds may be drawn per target class.
5. Use static chart formats for deterministic automated runs; interactive behavior is covered by the EDA sub-skill.
6. When the user only needs text preprocessing, explain the cleanup helpers without forcing a full AutoViz run.

Read these references

• references/workflows.md: text-cleaning and wordcloud-generation behavior.
• references/troubleshooting.md: NLTK, wordcloud, and string-column classification issues.
• ../../references/api-reference.md: signatures for text helper functions and draw_word_clouds.
• ../eda-visualization/references/workflows.md: how text columns fit into the full AutoViz run.
• ../../references/install-and-compatibility.md: package-version checks if the text branch breaks at import time.

Use these scripts

• Run scripts/wordcloud_smoke.py to verify that the wordcloud package and AutoViz text helper path can run on a safe in-memory DataFrame.
• If the text branch is failing inside a larger EDA run, use the EDA smoke script first and then return here.

Important behavior

• AutoViz may call nltk.download('popular') when discrete string variables are present.
• Short or low-cardinality strings may be classified as categorical/boolean rather than NLP.
• Very high-cardinality or long free-text columns are more likely to trigger wordcloud behavior.
• draw_word_clouds depends on matplotlib plus the wordcloud package.
• The text helpers also perform cleanup such as URL, HTML, emoji, contraction, and punctuation handling.

Common text helpers

• clean_steps
• clean_text
• remove_URL
• remove_html
• remove_emoji
• expandContractions
• remove_stopwords
• split_into_lemmas
• draw_wordcloud_from_dataframe

Cross-routing

• If the user primarily wants full-table EDA plots, route to ../eda-visualization/SKILL.md.
• If the user asks whether a text column should be cleaned or repaired before modeling, route to ../data-quality-fixes/SKILL.md and return here for visualization-specific text behavior.
• Do not rely on source notebooks or the original checkout; use the bundled references and script.
• If the user needs to know whether NLTK data will be downloaded, mention it explicitly before running the workflow.

Troubleshooting reminders

• If the column never reaches the wordcloud branch, the classifier likely treated it as categorical or boolean.
• If nltk.download('popular') is a problem, prepare the corpora or choose a static EDA path that does not trigger text plotting.
• If the output is empty or tiny, the text may have been over-cleaned or reduced to stopwords.
• Use a writable save_plot_dir when saving images so the generated wordcloud is easy to find.

Escalation

If the user really wants a full EDA view of the same dataframe, route back to the EDA sub-skill after handling the text question. If the data-quality question is about missingness or cleaning rather than visualization, route to the data-quality sub-skill instead of over-focusing on wordclouds.