QuranMetaphor: A Multi-Task Framework for Qur'anic Metaphor Analysis

Note: This repository is designed to ensure the reproducibility of the experiments presented in the manuscript. It implements the specific hierarchical tasks (Type, Origin, Functional Context) and the Qarīna-Aware Interaction Layer described in the paper. This module is a specialized component derived from the broader "Borhan Project" for computational rhetoric.

---

🏛️ Context (Background) : The Borhan Project (مشروع برهان)

This work is part of Borhan, a broader initiative establishing the field of Computational Rhetoric in Qur'anic Studies.

While the "Deep Rhetoric" framework (this paper) focuses on the structural modeling of metaphor using Multi-Task Learning, the larger Borhan project aims to map the full "Forest of Meaning," including tone analysis, sarcasm detection, and conceptual networks. The Qarina-Aware mechanism presented here serves as the foundational "Aesthetic Sensing" layer for these advanced applications.

1. The Vision: From Literal to Aesthetic

Current NLP models are "rhetorically blind." When algorithms process a verse like "Shall we believe as the fools believed?", they see syntax but miss the sarcasm, tone, and social layering. Borhan transforms fluid literary taste into a solid cognitive ontology, granting digital applications an "Emotional Intelligence" parallel to their linguistic capabilities.

2. The Innovation: A "Holistic Forest" of Meaning

Unlike traditional approaches that treat rhetorical devices as isolated trees, Borhan maps the "Forest of Meanings" through:

• Granular Modeling: Deconstructing a single image into >20 attributes (Type, Origin, Context, Tone).
• Pragmatic & Affective Analysis: Detecting the speech act (e.g., Deprecation vs. Mitigation) behind the metaphor.
• Conceptual Network Mapping: Linking images (e.g., Trade, Scales) to reveal the Transactional Logic of the Qur'anic worldview.
• Accommodating Semantic Complexity: A hybrid architecture that rejects binary classifications in favor of Multi-Maqasid (Multi-Intent) recognition.

3. Academic Reliability

This methodology is documented in 3 in-depth research papers currently under review at Q1 journals (SAGE, IEEE, Elsevier), ensuring that our "Qarina-Aware" algorithms meet the highest global academic standards.

---

📊 Dataset Description

The dataset provided in this repository (data/dataset_experiment.csv) is a task-specific extraction from the comprehensive Borhan Rhetorical Ontology.

While the source ontology contains granular metadata (e.g., pragmatic functions, sensory modes), this repository includes only the three structural dimensions modeled in the paper to establish a rigorous baseline:

• Total Samples: 2,649 Verses.
• Target Labels:
  1. Type: Explicit (Taṣrīḥiyya) vs. Implicit (Makniyya).
  2. Origin: Primary (Aṣliyya) vs. Derivative (Tabʿiyya).
  3. Context: Absolute, Candidate, Implied.

Note: The raw extended metadata is reserved for future generative tasks and is not required to reproduce the classification results reported in this study.

🔬 The Specific Solution: QuranMetaphor

This repository implements the Qarina-Aware Modeling (Contextual Clue Awareness) described in the Borhan methodology. It tackles the challenge of Metaphor (Istiʿāra) not as a detection task, but as a hierarchical inference problem.

Methodology

We define the problem as a set of interdependent classification tasks, translating rhetorical constraints into a Multi-Task Learning (MTL) architecture:

1. Type ($T_{type}$): Distinguishes Taṣrīḥiyya (Explicit) vs. Makniyya (Implicit).
2. Origin ($T_{origin}$): Distinguishes Primary sensory metaphors vs. Derivative associations.
3. Functional Context ($T_{context}$): Analyzes whether the metaphor is extended (Murashaha), absolute, or abstract.

Model Architecture

The model uses a hard parameter-sharing approach with a specialized Qarīna-Aware Interaction Layer—an unsupervised mechanism that mimics the human cognitive process of scanning context for "blocking indicators" (Qarīna) to resolve ambiguity.

graph TD
    subgraph Input_Layer["Input Layer"]
        A["Tokenized Verse (X)"]
        B["Candidate Span (s)"]
    end

    A --> C["Arabic Encoder (MARBERT / CamelBERT)"]
    B --> C

    C --> D["Hidden States"]

    subgraph Borhan_Core["Borhan Deep Rhetoric Core"]
        D --> E["Qarīna-Aware Interaction Layer"]
        E -->|Unsupervised Context Fusion| F["Refined Rhetorical Embeddings"]
    end

    subgraph Multi_Task["Multi-Task Heads"]
        F --> G["Head 1: Type"]
        F --> H["Head 2: Origin"]
        F --> I["Head 3: Context"]
    end

## ⚖️ License
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.