Technical Summary: Character Iconicity vs. Arbitrariness: An Arabic NLP Perspective

Problem Statement

Linguistic theory has long debated whether the relationship between orthographic form and meaning is arbitrary or iconic (systematic). In the context of Arabic script, this debate is testable because many letters share a common base shape (rasm) and are distinguished solely by dot placement. While prior research established that removing dots (creating "dotless" Arabic) preserves NLP performance, it remained unclear whether this success relies on the preservation of the original, historically motivated rasm groupings (iconicity) or if models can adapt to any stable, reduced symbol system, regardless of the mapping's logic (arbitrariness).

The core question addressed is: Does the success of dotless Arabic in NLP depend on preserving the visual/orthographic structure of the original rasm groupings, or can arbitrary, random remappings of characters to the same reduced set achieve comparable performance?

Methodology

The authors propose a controlled experimental framework to test the necessity of traditional orthographic structure by comparing standard dotted Arabic, standard dotless Arabic, and randomly remapped Arabic.

1. Mapping Generation:

   • The study utilizes the 19 undotted rasms as a fixed output symbol set.
   • Instead of the standard dotless mapping (where specific letters like H, H, and H map to a single rasm), the authors generated 2,000 random character mappings.
   • These random mappings preserve the group sizes of the original rasms but arbitrarily assign different letters to each group.
   • From the 2,000 permutations, four representative mappings were selected based on entropy metrics: Highest Character Entropy (HCE), Lowest Character Entropy (LCE), Highest Word Entropy (HWE), and Lowest Word Entropy (LWE).

2. Tokenization Strategies:
   To ensure robustness, experiments were conducted across five tokenization schemes:

   • Word Tokenizer
   • Character Tokenizer
   • Farasa Morphological Tokenizer
   • Disjoint-Letters Tokenizer
   • SentencePiece Byte Pair Encoding (BPE)

3. Tasks and Evaluation:
   The representations were evaluated across five distinct NLP tasks:

   • Language Modeling: Causal Transformer models evaluated on perplexity, vocabulary size, and training efficiency.
   • Text Classification: Poem meter classification, Sentiment Analysis (LABR dataset), and News Topic Classification (Sanad dataset).
   • Sequence Labeling: Part-of-Speech (POS) tagging and Named Entity Recognition (NER).
   • Machine Translation: Arabic-English translation with restoration of the original script post-decoding.
   • Text Restoration: An RNN-based model was trained to reconstruct the original dotted script from remapped text to assess reversibility.

4. Statistical Analysis:
   The study analyzed Zipf's Law (rank-frequency distribution) and Heap's Law (vocabulary growth) to determine if random mappings preserved the fundamental statistical regularities of the language.

Key Results

1. Arbitrariness vs. Iconicity

The primary finding is that preserving original character distinctions or traditional rasm-based groupings is not necessary for strong NLP performance.

• Random remappings (both high and low entropy) achieved competitive performance across all tasks compared to standard dotted and dotless baselines.
• In several instances, random mappings outperformed or matched the standard dotless baseline, suggesting that models rely more on stable distributional structures than on the visual iconicity of letter forms.

2. Efficiency Gains

Arbitrary remappings, particularly those with lower entropy (LCE and LWE), offered significant computational advantages:

• Vocabulary Reduction: Random mappings reduced vocabulary sizes by up to 50% compared to standard dotted text.
• Model Size and Training Cost: Reduced vocabularies led to smaller model sizes and significantly shorter training times (e.g., LWE reduced training time by ~40% in language modeling compared to dotted text) with only modest increases in perplexity.
• OOV Rates: Lower entropy mappings (LCE, LWE) achieved the lowest Out-of-Vocabulary (OOV) rates (approx. 5.04%), suggesting improved generalization to unseen word forms.

3. Task-Specific Performance

• Language Modeling: While standard dotted text achieved the lowest perplexity, random mappings maintained competitive scores. Character-level tokenization showed nearly uniform perplexity across all representations, indicating minimal impact from remapping at that granularity.
• Classification (Sentiment/Topic/Poem): Random mappings maintained high accuracy. For example, in news topic classification, accuracy remained stable (0.973–0.981) across all mappings despite substantial vocabulary compression. In poem meter classification (a task sensitive to phonology), high-entropy random mappings (HCE, HWE) achieved accuracy (0.955) very close to the dotted baseline (0.960).
• Sequence Labeling (POS/NER): Syntactic and semantic information remained effectively captured. HWE and HCE mappings maintained virtually identical accuracy to the baseline, while LCE and LWE showed only moderate drops (1–2.7%) despite larger vocabulary reductions.
• Restoration: The original dotted script could be reliably reconstructed from remapped text. The restoration model achieved low error rates (e.g., CER = 0.0235 for the highly compressed LCE mapping), confirming the reversibility of the process.

4. Statistical Regularities

Analysis of Zipf's and Heap's laws revealed that all mappings, including random ones, preserved the power-law distributions of the language. While low-entropy mappings showed steeper decay in the long tail (indicating higher compression), the fundamental statistical structure required for NLP remained intact.

Significance and Claims

The paper claims that from a computational NLP perspective, Arabic character form-function relationships are largely arbitrary. The success of dotless Arabic is not contingent on the preservation of historically motivated visual groupings but rather on the stability of the mapping itself.

The authors assert that:

1. Visual Iconicity is Non-Essential for NLP: Models do not require the "correct" visual grouping of letters to learn language structure; they can adapt to arbitrary, consistent remappings.
2. Efficiency-Performance Trade-off: Random remapping strategies offer a viable path to reduce vocabulary size, model complexity, and training costs without significantly compromising accuracy.
3. Reversibility: The compression achieved through arbitrary mapping does not result in irreversible information loss, as the original script can be accurately restored via learned models.

The study concludes that the "iconicity" of Arabic script is not a functional necessity for machine learning models, which appear to prioritize stable distributional patterns over orthographic design principles.