Large Language Models for Token-Efficient and Semantic-Preserving Opinion Summarization

Large Language Models (LLMs) are powerful tools for summarizing opinionated text, but they often struggle with large, redundant, and imbalanced datasets. Processing entire corpora is computationally expensive and can lead to biased summaries that favor majority viewpoints. To address this, the authors introduce a framework that optimizes the input fed to the LLM rather than relying on brute-force ingestion or retrieval-augmented generation (RAG) alone.

The framework operates in four stages: data collection, multidimensional classification, stratified sampling, and facet-aware summarization. By first annotating opinions across multiple semantic dimensions—such as sentiment, emotion, and topic—the system creates a structured representation of the corpus. It then uses stratified sampling strategies to select a compact subset of opinions that maintains the original distribution of these facets, ensuring the LLM receives a balanced and representative sample.

The authors compare three distinct sampling strategies to balance semantic relevance with distributional fidelity:

• Knapsack: A fast, constraint-based approach that selects highly relevant opinions while strictly adhering to target class counts.
• Knapsack-KL: A more sophisticated variant that balances relevance with the minimization of Kullback-Leibler (KL) divergence, ensuring the subset distribution closely matches the original corpus.
• KDE (Kernel Density Estimation): A computationally intensive method that provides the highest fidelity in reconstructing the original multidimensional probability distributions.

This approach offers a practical solution for organizations and researchers who need to analyze large volumes of user-generated content efficiently. By reducing the number of tokens required for summarization, the framework lowers costs and latency while improving the quality of the output. Crucially, by forcing the LLM to process a distribution-aligned subset, the method mitigates the common problem of models ignoring minority viewpoints or nuanced feedback, resulting in summaries that more accurately reflect the diversity of the original opinion space.