Cluster-Based Summarization System: Scalable Text Clustering and Summarization Pipeline

1. Overview

The Cluster-Based Summarization System is designed to analyze and summarize large volumes of user-generated text, such as warranty descriptions, customer feedback, or support tickets. The system uses state-of-the-art NLP techniques to:

• Cluster semantically similar entries without supervision.
• Extract meaningful keywords from each cluster.
• Summarize recurring themes using LLMs.

This allows technical and business teams to discover insights and root causes at scale—without manual review.

2. Architecture Overview

3. Technology Stack

4. Clustering Approach

4.1 UMAP for Dimensionality Reduction

Transformer embeddings are high-dimensional (384–1024D). UMAP is used to project them to ~5 dimensions for clustering. It preserves semantic neighborhoods better than PCA or t-SNE and is highly configurable (e.g., n_neighbors, min_dist).

4.2 HDBSCAN for Density-Based Clustering

HDBSCAN offers major advantages:

• No need to predefine k.
• Robust to noise/outliers.
• Supports soft clustering (probability per cluster).

Common parameters tuned include:

• min_cluster_size: minimum samples to form a cluster
• min_samples: how conservative the algorithm is with labeling core points

5. Keyword Extraction

To interpret clusters:

• TF-IDF highlights high-frequency, low-global-frequency terms.
• KeyBERT identifies phrases semantically close to the cluster embedding centroid.
• Keywords from both methods are merged to provide both statistical and semantic anchors.

These keywords also prime the LLM to better understand each cluster’s context during summarization.

6. Summarization Pipeline

A prompt is constructed for each cluster using:

1. Top Keywords (from TF-IDF + KeyBERT)
2. Representative Samples (30 texts per cluster)
   • Selected using a hybrid strategy:
     • Highest keyword density
     • Longest and shortest entries
     • Random sampling for diversity

These elements are formatted into a bullet-point prompt to avoid “wall of text” issues and ensure clarity for the LLM.

Output Fields per Cluster:

7. Evaluation & Tuning

To ensure meaningful clustering, we evaluate the output using both qualitative and quantitative metrics:

• Silhouette Score: Measures cohesion and separation of clusters. Higher scores indicate better-defined clusters.
• Number of Clusters: Helps track how granular or coarse the clustering is.
• Outlier Percentage: Monitors the proportion of data points classified as noise.

A grid search is used to iterate over combinations of n_neighbors, min_cluster_size, and min_samples, optimizing for silhouette score and interpretability.

Example log output:

n_clusters=12, silhouette=0.45, outlier_pct=12%

💡 In production settings, text embeddings are often high-dimensional and noisy. A silhouette score between 0.30–0.50 can still yield highly useful results when paired with meaningful summaries and cluster interpretation.

8. Visualization

To validate the quality of clustering and understand the data structure, we project the embedding vectors into 2D space using UMAP for visualization. Each point represents a text entry, and colors correspond to cluster assignments.

import umap
import matplotlib.pyplot as plt

umap_2d = umap.UMAP(n_components=2).fit_transform(embeddings)
plt.scatter(umap_2d[:, 0], umap_2d[:, 1], c=labels, cmap='Spectral', s=5)
plt.title("UMAP 2D Projection of Text Clusters")
plt.xlabel("UMAP-1")
plt.ylabel("UMAP-2")
plt.colorbar()
plt.show()

This plot helps assess whether clusters are well-separated and how many points were labeled as noise (outliers).

9. Output Schema

Each cluster produces a structured summary output with the following fields:

These results can be consumed by downstream pipelines, integrated into dashboards, or exported for reporting.

10. Future Extensions

This system is modular and extensible. Possible future improvements include:

• Topic Labeling: Auto-generate descriptive titles for each cluster using rules or LLMs.
• Multilingual Support: Incorporate models like distiluse-base-multilingual to support non-English datasets.
• Feedback Loop: Allow human reviewers to accept, reject, or refine summaries and feed corrections back into the pipeline.
• Time-Based Drift Detection: Compare clustering results over time to detect shifting trends or emerging issues.

11. Conclusion

The Cluster-Based Summarization System demonstrates a practical and scalable approach to understanding large-scale unstructured text data. By combining unsupervised clustering with LLM-powered summarization, it helps teams identify patterns, failure modes, or customer concerns efficiently.

Its design emphasizes:

• Interpretability via keywords and summaries
• Scalability via vectorization and density-based clustering
• Flexibility with customizable prompts and modular stages

This solution is production-ready and adaptable across domains like customer service, product feedback, and internal QA.