💬 NLP

DWTSumm: Discrete Wavelet Transform for Document Summarization

This paper proposes DWTSumm, a lightweight framework that applies Discrete Wavelet Transform to text embeddings to decompose documents into global and local semantic components, thereby improving the factual grounding, semantic fidelity, and hallucination reduction of long-document summarization in clinical and legal domains compared to standard LLM baselines.

Original authors: Rana Salama, Abdou Youssef, Mona Diab

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rana Salama, Abdou Youssef, Mona Diab

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, 500-page legal contract or a complex medical file. You need to summarize it for a busy doctor or a judge who only has five minutes to read. If you just ask a smart AI (like a Large Language Model) to "read this and tell me the main points," it often gets overwhelmed. It might forget the middle parts, make up facts (hallucinations), or miss critical details because the document is too long for its "working memory."

This paper introduces a clever new trick called DWTSumm. Think of it as a smart audio equalizer for text.

Here is how it works, broken down with simple analogies:

1. The Problem: The "Overwhelmed Chef"

Imagine an AI summarizer is a chef trying to cook a meal using a recipe that is 10,000 pages long.

The Issue: If the chef tries to read the whole thing at once, they get tired, skip pages, or accidentally add salt instead of sugar. In AI terms, this is called "losing context" or "hallucinating."
The Old Way: Previous methods tried to chop the recipe into small pieces, summarize each piece, and then glue them back together. But often, the glue didn't hold, and the flavor (meaning) got lost.

2. The Solution: The "Wavelet Equalizer"

The authors treat the document not as a story, but as a sound wave.

The Analogy: Imagine a song. It has a deep, steady bass line (the main melody) and high-pitched, fast notes (the details).
The Tool: They use a mathematical tool called the Discrete Wavelet Transform (DWT). Think of DWT as a special pair of glasses or a sound mixer that separates the song into two tracks:
1. The Bass (Approximation): This is the "big picture." It captures the overall structure, the main plot, and the global context.
2. The Treble (Detail): This is the "fine print." It captures specific dates, names, medical dosages, or legal clauses that are crucial but easy to miss.

3. How It Works: The "Smart Sifter"

Instead of feeding the whole 500-page document to the AI, the DWT acts as a smart sifter:

Decompose: It breaks the text down into the "Bass" (global structure) and "Treble" (critical details).
Compress: It throws away the "noise" (redundant words, repetitive stories) while keeping the essential frequencies. This shrinks the document size by 60–80% without losing the important stuff.
Reconstruct: It picks out the most important sentences that represent those "Bass" and "Treble" signals.
Feed the AI: Now, instead of a 500-page monster, the AI gets a tiny, super-condensed "skeleton" of the document that still contains all the vital facts.

4. The Results: Why It's Better

The paper tested this on medical records and legal court cases. Here is what happened:

Less "Making Things Up": Because the AI was given a clean, fact-checked skeleton, it stopped guessing. It was like giving the chef a pre-measured list of ingredients instead of a messy pantry. The "hallucinations" (making up facts) dropped significantly.
Better Memory: The AI remembered the "middle" of the story much better because the DWT didn't force it to skip the middle to fit everything in.
High Fidelity: In the tests, the summaries were 97% accurate to the original facts. It was like taking a high-resolution photo and shrinking it down to a thumbnail, but the thumbnail still looked exactly like the original photo, not a blurry mess.

The Big Picture Takeaway

Think of DWTSumm as a translator that speaks "Signal" instead of "Words."

It realizes that a long document is just a signal with patterns. By using math to separate the "main theme" from the "specific details," it creates a super-efficient summary. This allows the big, powerful AI models to do their best work without getting confused by the sheer size of the text.

In short: It's a way to shrink a novel down to a single page without losing the plot, the characters, or the ending, ensuring the AI tells the truth every time.

1. Problem Statement

Summarizing long, domain-specific documents (e.g., clinical records, legal opinions) using Large Language Models (LLMs) faces three critical challenges:

Context Limitations: Documents often exceed the model's context window, necessitating truncation or aggressive multi-stage pipelines that lead to information loss.
Hallucinations: LLMs tend to fill informational gaps with unsupported content, particularly when "lost-in-the-middle" phenomena occur in long sequences.
Factual Accuracy: In high-stakes domains like law and medicine, preserving specific facts and domain-specific nuances is critical, yet current hierarchical summarization methods often degrade factual consistency.

Existing solutions, such as hierarchical summarization or retrieval-augmented generation (RAG), often struggle to balance length reduction with the preservation of both global structure and fine-grained local details.

2. Methodology

The authors propose DWTSumm, a framework that treats text as a semantic signal and applies the Discrete Wavelet Transform (DWT)—a signal processing technique—to decompose and compress document representations.

Core Mechanism

Signal Representation: The input document is converted into a sequential vector representation using sentence-level embeddings (e.g., BioClinical ModernBERT for medical, Legal-English-Longformer for legal).
Multiresolution Decomposition: The DWT is applied hierarchically to these embeddings. At each level, the signal is convolved with low-pass and high-pass filters to produce:
- Approximation Coefficients ($cA$): Capture the global semantic structure and narrative flow (low-frequency components).
- Detail Coefficients ($cD$): Capture localized, domain-specific information and salient variations (high-frequency components).
Coefficient Mapping: The coefficients are mapped back to representative sentences via nearest-neighbor retrieval in the embedding space. This creates a "semantic skeleton" ($cA$) and a set of "salient detail sentences" ($cD$).
Generation Strategies: The framework operates in two modes:
- DWT-Only: The reconstructed coefficients serve directly as a compressed summary.
- Hybrid (DWT + LLM): The decomposed representation is fed into an LLM (e.g., GPT-4o) as a prompt to guide abstractive generation, ensuring the model is grounded in both global context and critical facts.

Key Design Choice: Unlike CNNs where filters are learned, DWT uses analytically defined filters (specifically Daubechies wavelets in this study), providing deterministic interpretability and stability without the need for training.

3. Key Contributions

Novel Framework: The first systematic introduction of DWT-based compression for LLM-based summarization, treating text as a signal to be decomposed rather than just a sequence of tokens.
Preservation of Structure: Demonstrates that DWT effectively separates global context from local details, allowing for substantial input length reduction (often 50–75%) without losing critical domain-specific semantics.
Hallucination Reduction: Shows that DWT acts as a "semantic denoising" mechanism, filtering redundant content and strengthening factual grounding, thereby reducing LLM hallucinations.
Generalizability: Proves the method is robust across different embedding models (e.g., MedCPT, BioLinkBERT) and domains (Clinical and Legal).

4. Experimental Results

The framework was evaluated on MultiClinSum (clinical) and CaseSumm (legal) datasets, comparing DWT-only and DWT+GPT configurations against a GPT-4o baseline.

Clinical Summarization (MultiClinSum)

Compression: Achieved >60% compression ratio.
Performance:
- BERTScore: Improved by >2% over the baseline.
- Semantic Fidelity: Increased from 91.60% (Baseline) to 95.06% (DWT+GPT).
- Factual Consistency: Improved from 4.0 to 5.0 (on a 1-5 scale).
- METEOR: Significant improvement (32.60 $\to$ 46.00), indicating better preservation of nuanced clinical semantics.

Legal Summarization (CaseSumm)

Performance:
- Fidelity: Reached 98.11% with the DWT+GPT model (vs. 92.80% baseline).
- METEOR: Improved from 30.80 to 45.9, suggesting superior capture of legal semantic nuances.
- Factuality: Achieved perfect scores (5.0) in both DWT configurations.
Observation: Even the DWT-only model outperformed the zero-shot GPT-4o baseline, proving that multiresolution decomposition alone preserves critical legal details.

Robustness Analysis

Embedding Agnosticism: The framework maintained high Fidelity (up to 98.66%) and stable BERTScore across four different clinical embedding models, confirming it captures core semantic structure rather than overfitting to a specific representation.
Decomposition Levels: Increasing decomposition levels (up to Level 3) allowed for 87.5% compression while maintaining or slightly improving factual consistency.

5. Significance and Conclusion

Lightweight & Interpretable: DWTSumm offers a preprocessing mechanism that is computationally efficient and interpretable, decoupling semantic structuring from language generation.
Denoising Effect: The results suggest DWT functions as a semantic filter, removing redundancy while anchoring the summary in factual evidence, which is crucial for high-stakes domains.
Future Impact: This approach provides a generalizable solution for long-document summarization that mitigates the "lost-in-the-middle" problem and reduces hallucinations without requiring complex, multi-stage training pipelines.

Limitations: The method relies on the quality of the underlying embeddings; weaker embeddings may reduce DWT's efficacy. Additionally, the fixed wavelet family and the approximation involved in mapping coefficients back to text may occasionally affect discourse coherence, though this is mitigated by downstream LLM synthesis.