DWTSumm: Discrete Wavelet Transform for Document Summarization

2026-04-24 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, long

Summary

A new framework, DWTSumm, leverages the Discrete Wavelet Transform (DWT) to enhance summarization of long, domain-specific documents, particularly in clinical and legal contexts. This method treats text as a semantic signal, decomposing it into global (approximation) and local (detail) components using DWT on sentence- or word-level embeddings. This process creates compact representations that preserve overall structure and critical domain-specific details, which can then be used directly as summaries or to guide Large Language Model (LLM) generation. Experiments on clinical and legal benchmarks show DWTSumm achieves comparable ROUGE-L scores to a GPT-4o baseline, while consistently improving semantic similarity and grounding. It demonstrates gains of over 2% in BERTScore, more than 4% in Semantic Fidelity, and significant METEOR improvements, indicating preserved domain-specific semantics and factual consistency, with Fidelity reaching up to 97%.

Key takeaway

For AI Engineers and Research Scientists developing summarization solutions for long, domain-specific documents, integrating Discrete Wavelet Transform (DWT) as a preprocessing step can significantly improve factual consistency and semantic fidelity. You should consider implementing DWT to compress input texts and guide LLM generation, especially in fields like clinical or legal documentation, to mitigate hallucinations and enhance the reliability of your summaries.

Key insights

DWT-based semantic compression improves LLM summarization by preserving factual grounding and reducing hallucinations in long, domain-specific texts.

Principles

• Multiresolution analysis provides hierarchical semantic representations.
• DWT compression reduces sequence length while preserving semantic fidelity.
• Separating global and local information supports factual grounding.

Method

The DWT-based framework embeds text, applies multi-level DWT to decompose it into approximation and detail coefficients, maps these back to representative sentences, and then uses this multi-resolution representation to guide or directly form LLM summaries.

In practice

• Apply DWT to sentence embeddings for long document compression.
• Use DWT-generated representations to guide LLM abstractive generation.
• Evaluate DWT's impact on factual consistency and semantic fidelity.

Topics

• Discrete Wavelet Transform
• Document Summarization
• Large Language Models
• Clinical Summarization
• Legal Summarization

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Summarisation and Knowledge Distillation — How Agents Summarise Without Hallucinating — Robust LLM summarization requires architectural patterns to prevent hallucinations and ensure factual grounding and completeness.
• Text Summarization with Scikit-LLM — Integrate LLM-driven text summarization into scikit-learn pipelines to manage large text volumes efficiently.
• Text-as-Signal: Quantitative Semantic Scoring with Embeddings, Logprobs, and Noise Reduction — Text can be transformed into operational, continuous semantic data points for AI engineering via embeddings and logprob scoring.
• Enhancing Factuality through Consensus and Consistency in Summarization Using Minimum Bayes Risk Decoding — ConSUM enhances summary factuality by combining source consistency with candidate consensus via MBR reranking.
• Discovery of Legal Patterns in Civil Petitions via LLM-Based Fact Extraction and Density Clustering — LLM-based fact extraction significantly improves legal document clustering by reducing noise and enhancing semantic coherence.
• DySem: Uncovering Dynamic Semantic Components via Multilingual Consensus for Calculating Semantic Textual Similarity — DySem uses multilingual consensus to find dynamic, sample-specific semantic dimensions for improved STS in LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.