Traditional Explainable AI (XAI) frameworks typically communicate through mathematical abstractions. They present users with SHAP values, attention weights, or raw saliency maps. While these outputs are technically precise, a vector of floating-point numbers or a dense matrix of statistical weights does not inherently convey meaning to a non-expert user, nor does it provide actionable insights to a domain specialist making high-stakes decisions.
To bridge this gap, a prominent research direction focuses on natural language explanation generation. This paradigm translates quantitative model attributions into structured, human-readable narratives, transforming raw statistical weights into contextualized justifications.
Decoupling Reasoning from Communication
The core intuition behind why this approach works lies in the explicit separation of the predictive/interpretability engine from the linguistic realization engine.
[Input Data] ──> [Predictive Model + XAI Tool] ──> [Quantitative Weights] ──> [NLG / LLM Layer] ──> [Natural Language Narrative]
As highlighted by Jonas Bokstaller et al. and the design of the iLLuMinaTE framework, decoupling the pipeline ensures that the final explanation remains strictly grounded in data-driven statistical weights, while simultaneously benefiting from the communicative fluency of natural language generation. The specialized machine learning model handles the analytical reasoning, while the generation layer handles human-centric communication.
The Template and Grammar Era
Early systems achieved this translation using deterministic mapping functions. The goal was to convert numerical intervals into qualitative human concepts without losing fidelity.
The authors of these early systems established structured frameworks to map data to discourse:
- Narrative Roles: Or Biran et al. categorized feature contributions into specific "narrative roles" (such as exceptional evidence or missing evidence) to structure justifications rather than just listing numbers.
- Formal Grammars: To make these descriptions precise, Azzurra Ragone et al. utilized formal Backus-Naur Form (BNF) grammars to programmatically map numerical feature importance intervals to specific linguistic modifiers, such as adverbs and adjectives that accurately reflect the intensity of a feature's impact.
- Document Planning: Pipelines designed by James Forrest et al., Mohsin Munir et al., and Devleena Das et al. demonstrated how raw sensor events or statistical time-series characteristics could be converted into multi-tier textual summaries via formal document planning and predefined templates.
While highly controllable, these rule-based methods are rigid; they struggle to scale when features interact in complex, non-linear ways or when the target audience requires deep domain context.
LLMs as Domain-Aware Translators
Modern architectures replace rigid templates with Large Language Models (LLMs) acting as intelligent translators. Instead of generating text from scratch—which risks hallucination—the LLM is provided with sparse feature masks, decision paths, or attribution weights as a conditioning prompt.
This approach is highly effective because LLMs possess vast internal domain-specific prior knowledge. They do not merely print the names of the features; they understand what those features mean in context:
- Contextual Mapping: In the InterpreTabNet architecture proposed by Jacob Yoke Hong Si et al., a neural network generates sparse, non-overlapping feature attention masks from tabular data. These discrete masks are fed directly into an LLM, which leverages its prior knowledge to translate raw statistical weights into domain-specific rationales for healthcare or finance.
- Stylistic Adaptation: The Explingo framework, introduced by Alexandra Zytek et al., pairs a quantitative explanation tool with an LLM-based "Narrator" and "Grader" pipeline. This architecture allows complex machine learning outputs to be refactored into diverse linguistic styles tailored to different audiences (e.g., casual users versus domain experts) using few-shot prompting.
- Actionable Mitigation: By anchoring the LLM's prompt with quantitative feature importance data, frameworks like XG-NID (Yasir Ali Farrukh et al.) bridge the gap between black-box model attributions and human security analysts, generating natural language reports that pair classification results with actionable mitigation strategies.
Grounding and Domain Integration
To ensure that these generated narratives are linguistically coherent and faithful to the underlying data, recent architectures integrate external knowledge structures and multi-agent systems. For instance, Xinyue Quan et al. combined multi-task learning with external linguistic knowledge bases to aggregate character-level importance scores into coherent phrases, ensuring the textual evidence aligns with human reading patterns.
In highly specialized domains, this synthesis is critical for safety and utility:
- Clinical Reasoning: Yinghao Zhu et al. (ColaCare) utilize multi-agent LLM collaboration grounded in medical guidelines to interpret numerical Electronic Health Record (EHR) models, transforming raw predictive signals into natural language clinical reviews.
- Time-Series and Anomaly Contextualization: Systems tracking complex temporal sequences—such as the civil unrest forecasting system by Sathappan Muthiah et al., the suicide risk detection framework by Koushik L et al., and the real-time stress monitoring system by Davide Gabrielli et al.—leverage LLMs to condense intricate anomaly detection outputs into brief, actionable summaries for practitioners.
- Physical and Chemical Intuition: In molecular modeling, Vansh Ramani et al. (DISSOLVR) demonstrate that feeding internal model artifacts and decision paths into an LLM yields chemically accurate rationales that bridge the gap between statistical predictions and physical intuition.
Standardizing the Narrative Pipeline
As the field shifts toward fluid text generation, evaluating the truthfulness of these explanations becomes paramount. Standardized benchmarking frameworks, such as EvalxNLP, have emerged to evaluate post-hoc explainability methods across dimensions of faithfulness, plausibility, and complexity, using LLMs to systematically verbalize technical metrics into plain English.
Ultimately, by translating abstract mathematical weights into structured prose, these systems satisfy regulatory transparency demands (Vito Walter Anelli et al.) and empower non-technical stakeholders to safely trust and act upon automated decisions (Karishma Mohiuddin et al.).
Key Foundational Work
EMBERS at 4 years: Experiences operating an Open Source Indicators Forecasting System (2016) — contributed 1 concept elements to this direction.
Human-Centric Justification of Machine Learning Predictions (2017) — contributed 1 concept elements to this direction.
The Role of Textualisation and Argumentation in Understanding the Machine Learning Process (2017) — contributed 1 concept elements to this direction.