Harnessing AI for Business Insight: Deploying LLM Summarization Pipelines

Book Chapter August 2024

Large language modelsHuman-AI interactionGPT-4 case study

Overview

This chapter proposes a business-oriented framework for building and evaluating LLM-powered text-summarization systems. It critiques single-number metrics such as ROUGE and BERTScore, arguing for multi-dimensional evaluation (relevance, factuality, fluency, coherence) and explicit consideration of human–computer-interaction factors and domain context. A detailed GPT-4 case study shows how the framework surfaces actionable insights from 1,000+ YouTube comments on a smartphone launch, using sentiment-split summarizers, topic alignment, and a rich multi-dimensional data visualization.

Read the Published Chapter (Oxford UP)
In The Oxford Handbook on the Foundations and Regulation of Generative AI (2024)

How the Work Came Together

We began with a critical review of how LLM summarization systems are typically evaluated, focusing on the limitations of reducing quality to a single score such as ROUGE or BERTScore. From that review, we developed a richer framework that treats relevance, factuality, fluency, coherence, and the surrounding human workflow as distinct parts of system quality.

To demonstrate the framework in practice, I built a two-stage structured user-generated-content pipeline powered by GPT-4. Sentiment-specific summarizers first condensed the source material, after which a comment-to-topic labeler aligned individual comments to emerging themes. The resulting dashboard made it possible to trace high-level summaries back to the structure of the underlying feedback.

What the Chapter Proposes

The chapter challenges the “myth of the single number”: automated metrics such as ROUGE and BERTScore correlate only weakly with the dimensions that determine whether a summary is genuinely useful to a business. Trust, explainability, and placement within a manager's workflow all mediate the value of the final output.

We therefore propose a context ontology for product-feedback summarization that provides a reusable set of context types and prompts for new domains. The goal is to retain the flexibility of general-purpose language models while supplying enough task-specific structure to produce nuanced, actionable insights.