COVID19 Tweets

About Dataset

Install crisp-t following the Getting Started steps.

Steps 1: Download

• Create a project folder mkdir covid_tweet
• cd to the new folder cd covid_tweet/
• create the source folder mkdir crisp_source
• Download and copy the CSV file from here to the crisp_source folder.

Step 2: Import

• Import with crisp --source crisp_source --out crisp_input --unstructured text where text field has the tweet content.

Step 3: Qualitative analysis

• See the first 5 documents: crisp --inp crisp_input --print documents
• Assign topics & cluster: crisp --inp crisp_input --assign --out crisp_input
• See the assigned topics/keywords and cluster: crisp --inp crisp_input --print "documents metadata"
• Select all documents and dataframe rows belonging to cluster 0: crisp --inp crisp_input --filters cluster=0. You may save the results to a different folder using the --out option if you want to do further analysis only on this cluster.
• Select all documents tagged mask along with the corresponding dataframe rows: crisp --inp crisp_input --filters keywords=mask. You may save the results to a different folder using the --out option if you want to do further analysis only on this cluster. Notice that the output says Applied filters ['keywords=mask']; remaining documents: 62160. The total number is 179108.

Step 4: Quantitative analysis

• Run numeric analyses (requires installation with [ml] extras):

crisp --inp crisp_input --ml --regression --cart --kmeans --pca --out crisp_input

• Inspect stored metadata for confirmatory signals:

crisp --inp crisp_input --print metadata --print decision_tree_accuracy
crisp --inp crisp_input --print metadata --print decision_tree_feature_importance
crisp --inp crisp_input --print metadata --print regression_coefficients

Step 5: Triangulation example

1. Filter tweets by keyword (e.g., mask) and re-run topics to inspect theme stability:

crisp --inp crisp_input --filters keywords=mask --topics

2. Compare qualitative emphasis with quantitative predictors:

   • If avg_response_time or a proxy metric (e.g., engagement_delay if available) shows high feature importance or positive coefficient for a COVID outcome proxy, note convergence between the "delayed response" theme and numeric effect (Mettler et al., 2025).

3. Visualize for context:

crispviz --inp crisp_input --pca --out viz_out/

4. Memo: "Mask-related discourse overlaps with clusters exhibiting higher engagement delay; topic salience and decision tree splits suggest temporal dynamics in discourse response." This structural insight can be further examined via TDABM for shape-aware context (Rudkin & Dlotko, 2024).