Pipeline

VersoVector combines supervised and unsupervised NLP workflows.

For a visual explanation of how the modeling components connect across notebooks, see Model Topology.

The goal is to produce an integrated analytical view of poetic text: predicted tags, semantic neighbors, topics, clusters, and projection metadata.

High-level flow

1. Cleaning pipeline

The cleaning stage prepares the corpus for modeling.

Typical responsibilities:

• normalize source columns;
• clean poem text;
• generate stable identifiers;
• prepare processed datasets;
• preserve metadata needed for later interpretation.

Output examples:

data/poems_processed.csv

2. Feature pipeline

The feature pipeline transforms poems into numeric representations.

The project uses classic NLP representations such as:

• CountVectorizer;
• TfidfVectorizer;
• custom dictionary-like features;
• sparse feature matrices;
• normalization steps.

The feature pipeline is shared by supervised and unsupervised modeling stages.

3. Supervised branch

The supervised branch predicts multilabel tags.

It is used to assign interpretable emotional or thematic labels to poems.

Typical outputs:

• predicted tags;
• classifier metrics;
• multilabel binarizer artifacts;
• supervised model artifact.

4. Unsupervised branch

The unsupervised branch discovers structure without relying only on labels.

It may generate:

• cosine similarity neighbors;
• Pearson/correlation-based neighbors;
• LDA topics;
• KMeans clusters;
• Gaussian Mixture clusters;
• DBSCAN or Agglomerative clustering experiments;
• UMAP or t-SNE projections.

5. Integration

The integration stage combines supervised and unsupervised outputs into a single analytical table.

Typical integrated fields:

• poem identifier;
• metadata;
• predicted tags;
• dominant topic;
• topic terms;
• cluster assignments;
• nearest semantic neighbors;
• projection coordinates.

6. Visualizations

The final visualization stage helps interpret the model outputs.

Useful visualizations include:

• cluster maps;
• topic summaries;
• projection plots;
• tag distributions;
• nearest-neighbor examples;
• integrated result tables.