💻 computer science

A framework for hallucinations explainability in text summarization

This paper proposes a novel framework that introduces a Hallucination Severity Score (HSS) based on semantic instability and SHAP attribution analysis within an enhanced SMILE pipeline to effectively detect, quantify, and explain hallucinations in LLM-based text summarization.

Original authors: Omnia Abd Al-Azeem Hussieny, Samah El-Tantawy, Doaa Shawky, Ehab Y. Abdel Maksoud

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Omnia Abd Al-Azeem Hussieny, Samah El-Tantawy, Doaa Shawky, Ehab Y. Abdel Maksoud

Original paper licensed under CC BY 4.0 (https://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

In the modern landscape of artificial intelligence, large language models have become powerful engines for writing, answering questions, and summarizing complex information. These systems are trained on vast amounts of text, learning to predict the next word in a sentence with remarkable fluency. However, this fluency comes with a hidden flaw: the models sometimes invent facts. In the field of natural language processing, this phenomenon is known as hallucination. It occurs when a system generates text that sounds confident and grammatically correct but is either completely made up or directly contradicts the source material it was supposed to summarize. While these errors might seem minor in a casual chat, they become dangerous in high-stakes environments like healthcare, legal analysis, or scientific research, where accuracy is non-negotiable. The core challenge for researchers has not just been to spot these lies, but to measure how severe they are and understand why they happen, moving beyond simple checks for word overlap to a deeper investigation of the model's internal logic.

A team of researchers has developed a new method to tackle this problem, creating a framework that acts like a stress test for AI summaries. Instead of simply asking if a summary is factually correct, they asked a different question: how stable is the summary when the input is slightly tweaked? Their approach relies on the idea that a trustworthy summarizer should produce a consistent result even if the source text is changed in small, meaningful ways. If the summary changes drastically or starts inventing new details after a minor edit, it suggests the model is unstable and prone to hallucinating. The researchers built a system that takes a document, makes subtle, controlled changes to the words without altering the core meaning, and then watches how the AI's output reacts. By measuring the distance between the original summary and the new ones, they can calculate a specific score that indicates the severity of the hallucination risk.

To conduct this experiment, the team worked with a collection of thirty-four documents spanning diverse topics, including news articles, scientific papers, legal texts, and product reviews. They used a sophisticated summarization model to generate a baseline summary for each document. Then, they introduced a series of perturbations, which are small, deliberate modifications to the source text. These changes were carefully designed to swap words for their closest synonyms while ensuring that names of people, places, and specific numbers remained untouched to preserve the facts. For each document, they created multiple versions of the text, each slightly different from the original, and asked the AI to summarize each one. The researchers then compared the resulting summaries to see how much they drifted from the original. They found that when the AI was forced to summarize a slightly altered text, the summaries from models prone to hallucination would shift wildly, while reliable models stayed steady.

The researchers quantified this drift using a mathematical concept that measures the difference between two groups of data, essentially calculating how far apart the meaning of the summaries had moved. They discovered that this measure of instability was not just a random fluctuation; it was tightly linked to how much the AI was lying. To make this measure useful across different types of documents, they normalized the scores, creating a final severity score that could be compared fairly whether the text was about medicine or movies. The most striking finding emerged when they looked at how the AI explained its own choices. They used a technique that highlights which words in the input were most important for generating the output. They found that when a summary was hallucinating, the importance of the input words would jump around unpredictably. In other words, the model's internal reasoning became unstable right at the moment it started to lie.

The study revealed a strong connection between this instability in the model's reasoning and the severity of the hallucinations. When the researchers compared their new severity score against other common ways of checking for errors, they found that the instability of the model's internal explanations was the most reliable predictor of a hallucination. Traditional methods that simply check if words match or if sentences sound logical were less effective at catching these deep-seated errors. The team also tested their method across different levels of difficulty. They found that as they made the changes to the text more significant, the severity score increased, proving that their system was sensitive enough to detect even subtle shifts in reliability. Furthermore, the score varied naturally across different domains; complex fields like law and science showed higher instability scores than simpler topics like social media posts, reflecting the greater difficulty and higher risk of error in those areas.

This work provides a new lens through which to view the reliability of artificial intelligence. By focusing on how a model reacts to small changes rather than just the final output, the researchers have identified a clear signal of when an AI is losing its grip on the truth. The study suggests that the best way to detect a hallucination is not to look for the lie itself, but to watch for the tremors in the model's reasoning that precede it. While the researchers tested this on a specific set of documents and a single type of AI model, the framework offers a robust path forward for building systems that can be audited and trusted. It moves the conversation from simply asking "is this true?" to "how sure can we be that this is true?", providing a quantitative way to measure the confidence we should place in machine-generated text. As these systems become more integrated into our daily lives, tools that can distinguish between a stable, factual summary and a fragile, invented one will be essential for maintaining the integrity of the information we rely on.

Technical Summary: A Framework for Hallucinations Explainability in Text Summarization

1. Problem Statement

Large Language Models (LLMs) used for abstractive text summarization frequently suffer from "hallucinations," where generated outputs are linguistically fluent but factually incorrect, unsupported, or contradictory to the source text. These errors are particularly critical in high-stakes domains such as healthcare, journalism, and scientific research.

Current evaluation approaches face two primary limitations:

Insufficient Severity Quantification: Existing factuality metrics (e.g., FactCC, SummaC) provide sentence-level judgments or binary consistency scores but lack a robust scalar measure to quantify the severity of hallucinations under semantic perturbations.
Descriptive vs. Quantitative Explainability: While explainability frameworks like SMILE (Statistical Model-Agnostic Interpretability with Local Explanations), LIME, and SHAP offer insights into token importance and model sensitivity, they remain primarily descriptive. They explain why a model behaves a certain way but do not directly provide a quantitative score for hallucination severity.

The core hypothesis of this work is that faithful summarization models remain stable under meaning-preserving perturbations, whereas hallucination-prone models exhibit significant output instability.

2. Methodology: The Enhanced SMILE Framework

The authors propose a novel Hallucination Severity Score (HSS) integrated into an enhanced SMILE pipeline. The framework transforms SMILE from a qualitative interpretability tool into a quantitative diagnostic framework through the following steps:

2.1 Semantic Perturbation Generation

The system generates controlled, meaning-preserving perturbations of the input document ( $x$ ) to create a set of perturbed inputs $\{x^{(k)}\}$ .

Token Selection: Named entities (via NER), stop words, numbers, and punctuation are excluded to preserve factual consistency. Only semantically relevant tokens are candidates for modification.
Candidate Generation: Word2Vec embeddings identify top- $k$ nearest neighbors for eligible tokens. A cosine similarity threshold ( $\tau$ ) ensures semantic proximity.
Meaning-Preserving Filter: A Sentence-BERT (SBERT) similarity constraint ( $\gamma$ ) validates that the perturbed sentence retains the original semantic meaning. If $sim_{SBERT}(s, s') < \gamma$ , the perturbation is rejected.
Strength: Perturbations modify 2–30% of tokens, categorized as low, moderate, or high strength.

2.2 Summary Generation and Representation

Model: A BART-large-CNN transformer model generates a baseline summary ( $y_0$ ) and summaries for each perturbed input ( $y_k$ ).
Embedding Distribution: Each summary is decomposed into sentences, and contextual embeddings are generated using SBERT. These form an empirical distribution ( $\mu_y$ ) representing the semantic structure of the summary.

2.3 Instability Measurement via Wasserstein Distance

The framework quantifies the semantic shift between the baseline summary and perturbed summaries using the Wasserstein distance ( $W$ ), which measures the geometric transport cost between embedding distributions.
$D_k = W(\mu_{y_0}, \mu_{y_k})$
The raw instability is the average of these distances across $K$ perturbations.

2.4 Hallucination Severity Score (HSS) and Normalization

To enable cross-domain comparison, the raw Wasserstein distances are normalized.

Category-Aware Normalization: A normalization constant ( $M_c$ ) is computed as the 95th percentile of Wasserstein distances for a specific document category ( $c$ ).
Final Score: The HSS is the average of the normalized distances:
$HSS(x) = \frac{1}{K} \sum_{k=1}^{K} \min\left(\frac{D_k}{M_c}, 1\right)$
Higher HSS values indicate greater semantic instability and, consequently, higher hallucination severity.

2.5 Attribution Instability Analysis

The framework correlates HSS with attribution-based instability metrics, specifically SHAP (Shapley Additive Explanations) and LIME. It measures how much the feature attribution distributions shift when the input is perturbed.

3. Experimental Setup

Dataset: 34 documents across four domains (Simple, Medium, Complex, Mixed/Extra) covering movies, literature, AI, and medical research.
Scale: 286 total perturbations generated, resulting in 320 evaluated summaries.
Validation: Statistical analysis included Pearson and Spearman correlations, 1,000 bootstrap resamples for confidence intervals, and ablation studies.
Comparators: The HSS was compared against FactCC, SummaC, Precision/Recall/F1, Attention scores, LIME, and SHAP.

4. Key Results

The experimental evaluation yielded the following findings:

Strong Alignment with SHAP: The HSS exhibited the strongest correlation with SHAP-based attribution instability ( $r \approx 0.90$ ). This indicates that summaries prone to hallucinations produce highly unstable feature attribution patterns under semantic perturbations.
Negative Correlation with Factuality: HSS showed moderate-to-strong negative correlations with factual consistency metrics (FactCC: $r \approx -0.63$ ; SummaC: $r \approx -0.55$ ), confirming that higher severity scores correspond to lower factual reliability.
Weak Correlation with Lexical/Distributional Metrics: HSS had weak correlations with raw Wasserstein distance ( $r \approx -0.33$ ) and LIME ( $r \approx 0.29$ ). This suggests that surface-level lexical overlap or simple distributional shifts are insufficient proxies for semantic hallucination.
Domain Sensitivity: HSS effectively differentiated between domains. Knowledge-intensive domains (e.g., Scientific Papers, Legal Documents) showed higher HSS values (higher instability) compared to less structured domains (e.g., Social Media, Product Reviews).
Perturbation Sensitivity: HSS values increased consistently with perturbation strength (from 0.38 at low strength to 0.62 at high strength), validating its sensitivity as a quantitative scale.

5. Contributions and Significance

The paper claims the following contributions and significance:

Novel Metric (HSS): Introduction of a statistically grounded Hallucination Severity Score that quantifies hallucination severity through Wasserstein-based semantic instability, moving beyond binary factuality checks.
Enhanced SMILE Framework: Extension of the SMILE framework to integrate contextual embedding distributions, attribution instability, and perturbation-based robustness into a unified statistical pipeline.
Attribution Instability as a Proxy: Establishment of SHAP-based attribution instability as a reliable indicator of semantic hallucination behavior, demonstrating that it outperforms surface-level lexical metrics and LIME in detecting hallucinations.
Robust Validation: Confirmation of the framework's robustness through extensive statistical validation (correlation analysis, bootstrap confidence intervals) and ablation studies, which proved that components like normalization, NER protection, and semantic filtering are essential for reliable estimation.

The authors conclude that hallucination is fundamentally associated with semantic instability and cannot be fully characterized by surface-level lexical metrics alone. The proposed framework offers a pathway for model auditing, drift monitoring, and threshold-based quality assurance in LLM-based summarization systems.

6. Limitations

The authors modestly acknowledge several limitations:

Dataset Size: The evaluation was conducted on a relatively small dataset (34 documents), which may limit generalizability.
Model Specificity: Experiments were limited to a single summarization model (BART-large-CNN); broader validation across different LLM architectures is required.
Perturbation Constraints: The constrained perturbation strategy may not capture the full diversity of real-world linguistic variations.
Lack of Human Evaluation: The study relies on automated metrics; future work requires human evaluation to align automated severity estimates with human judgment.