Technical Summary: Automatic Classification of Arabic Literature into Historical Eras

Problem Statement

The Arabic language has undergone significant semantic and lexical transformations over its approximately 1,500-year history, characterized by the emergence of new vocabulary, the obsolescence of older terms, and shifts in usage. While historians and linguists have traditionally partitioned Arabic literature into distinct eras (e.g., Pre-Islamic, Islamic, Abbasid, Ottoman, Modern), there is a notable lack of research regarding the automatic classification of Arabic texts by time period, particularly outside the domain of poetry. Existing text classification tasks in Arabic often focus on sentiment or topic, leaving temporal periodization largely unexplored. Furthermore, the boundaries between literary eras are often debated, overlapping, and not sharply defined, making the task of assigning a text to a specific historical period a complex challenge for machine learning models.

Methodology

The study employs neural network and deep learning techniques to automatically classify Arabic texts into historical eras. The research utilizes two publicly available corpora:

1. OpenITI: A large-scale corpus of Arabic literary prose (approx. 1.5 billion words) covering religious and literary texts from diverse authors.
2. APCD (Arabic Poem Comprehensive Dataset): A dataset containing approximately 1.8 million verses from 3,701 poets spanning 12 distinct eras.

Experimental Design

The experiments were conducted under two distinct protocols to assess the impact of authorial style:

• Author-Disjoint: Authors in the evaluation set are absent from the training set, forcing the model to learn period-level stylometric signals rather than author-specific styles.
• Merged-Author: Authors appear in both training and evaluation sets, allowing the model to leverage authorial style as a feature.

Class Setups

The study evaluated various classification granularities:

• Binary: Classical vs. Modern.
• Predefined Historical Eras: Based on schemes by Daif (1982), Al-Fakhoury (1951), and Ghoniem (2017), ranging from 3 to 5 eras.
• Custom Periodizations: Fixed time bins of 300, 200, and 100 years (resulting in 5, 8, and 15 classes).

Models and Preprocessing

• Models: The primary architectures tested were Fully Connected Feed-Forward Artificial Neural Networks (ANNs) and Recurrent Neural Networks (RNNs), specifically utilizing LSTM, GRU, and Bidirectional GRU (BiGRU). Convolutional Neural Networks (CNNs) and Logistic Regression served as baselines.
• Tokenization: Both word-level and character-level tokenization were explored for RNNs.
• Preprocessing: Steps included data cleaning (removing HTML, non-Arabic digits), normalization (removing diacritics and Kashida), and stop-word removal. Lemmatization (using Farasa) was applied selectively, particularly for poetry datasets, to reduce vocabulary size and mitigate overfitting.
• Evaluation: Performance was measured using Accuracy, Precision, Recall, and Macro-averaged F1-score. Statistical significance was assessed using the difference-between-two-proportions test.

Key Contributions

1. Systematic Exploration: This is the first study to systematically investigate the automatic classification of Arabic literary texts (both prose and poetry) into distinct historical eras.
2. Dual Dataset Analysis: The work evaluates models on two diverse, publicly available corpora (OpenITI for prose, APCD for poetry) covering the spectrum from the Pre-Islamic to the Modern era.
3. Authorial Style Impact: The study explicitly isolates the effect of authorial style by comparing author-disjoint and merged-author experimental protocols.
4. Granularity Comparison: It examines classification performance across varying levels of temporal resolution, from binary classification to fine-grained 15-era setups, and compares predefined historical eras against custom time bins.
5. Model Evaluation: It provides a comparative analysis of ANNs and RNNs (word and character level) for this specific temporal task in Arabic.

Results

The study reports the following key findings:

• Feasibility: Automatic era classification is feasible but challenging. Models consistently outperformed random guessing, though performance varied significantly based on the number of classes and the dataset.
• Model Performance:
  • ANNs generally outperformed RNNs across most multi-class setups on both datasets.
  • RNNs (specifically word-level) performed competitively in binary classification tasks, achieving results comparable to or exceeding previous CNN-based baselines (Orabi et al., 2020) on poetry data.
• Dataset Specifics:
  • OpenITI (Prose): Binary classification achieved an F1-score of 0.67 (author-disjoint) and 0.83 (merged). Performance dropped significantly as class count increased (e.g., 15-class F1 ≈ 0.20). The Abbasid era was the most easily identified, while Islamic and Modern eras were the most difficult.
  • APCD (Poetry): Binary classification achieved an F1-score of 0.78 (author-disjoint) and 0.79 (merged). Lemmatization provided a notable boost for poetry classification. The 12-era setup yielded lower performance (F1 ≈ 0.18–0.25), indicating the difficulty of fine-grained distinction.
• Authorial Influence: Merging authors across splits generally improved performance for prose (OpenITI), suggesting authorial style is a strong predictor of era. However, for poetry (APCD), merging authors had little effect, implying that poetic style is less dependent on specific author identity in this context or that the models rely more on period-specific features.
• Preprocessing: Removing stop words did not improve performance; retaining them yielded better results. Lemmatization improved results for the poetry dataset but was less critical for prose.
• Confusion Patterns: Misclassifications predominantly occurred between consecutive eras (e.g., Islamic vs. Abbasid, Ottoman vs. Modern), supporting the hypothesis that literary evolution is a continuum with porous boundaries rather than a sequence of discrete steps. Merging adjacent eras significantly boosted accuracy (e.g., OpenITI accuracy rose from 43.6% to 76.6% when eras were merged).

Significance and Claims

The paper modestly claims that while automatic classification of Arabic literature into historical eras is difficult due to the gradual nature of linguistic change and overlapping boundaries, it is a viable task for machine learning.

• Literary Continuum: The results suggest that Arabic literary history is better modeled as a continuum. The high confusion rates between adjacent eras indicate that sharp temporal boundaries are artificial constructs that do not always align with linguistic reality.
• Utility: Automated classification can aid in organizing literary data, tracing language change, and authenticating works attributed to specific eras. It facilitates large-scale analysis of trends, themes, and stylistic elements across time.
• Future Directions: The authors suggest that future work could involve improving lemmatization for Classical Arabic, evaluating pretrained transformers (e.g., AraBERT, MARBERT) while accounting for temporal bias, and incorporating external knowledge such as Named Entity Recognition (NER) to link text to biographical resources. They also propose investigating the impact of text-reuse detection (removing boilerplate passages) on classification performance.