Picking Up Where the Linguist Left Off: Mapping Morphology to Phonology through Learning the Residuals

Salam Khalifa, Abdelrahim Qaddoumi, Ellen Broselow, Owen RambowACL 2024

TL;DR

This paper tackles the challenge of converting morphological representations into spoken forms for Egyptian Arabic, a low-resource dialect. They propose a novel approach that combines linguist-crafted rules with a neural network to learn and correct the errors left by these rules, achieving better performance than either method alone in low-resource scenarios.

What Problem Does This Paper Address?

Accurately mapping the underlying morphological structure of a word to its spoken form is crucial for many natural language processing tasks, especially for morphologically rich languages like Arabic. While Modern Standard Arabic (MSA) has well-established resources and relatively transparent orthography, Arabic dialects, such as Egyptian Arabic, are considered low-resource. They often lack standardized orthography and the mapping from morphology to phonology is less straightforward. Existing grapheme-to-phoneme models require large datasets, which are unavailable for most dialects. This paper aims to bridge this gap by developing a method to generate phonological representations for dialectal Arabic, specifically Egyptian Arabic, using a combination of linguistic knowledge and machine learning.

Methods

The core of the proposed method is a hybrid approach that leverages both linguistic expertise and neural networks. First, a linguist provides a set of hand-crafted morphophonological rules designed to map a word's morphological representation (MR) to its spoken form (SF). These rules are applied to the input MR. However, these rules are not perfect and leave residual errors. To address these inaccuracies, a character-level neural transformer model is trained to learn these residuals. Essentially, the transformer learns to correct the output of the linguist's rules, effectively acting as a refinement step. This approach is inspired by prior work that trains models to correct rule-based predictions.

Datasets

The primary dataset used in this work is the Egyptian Colloquial Arabic Lexicon (ECAL). ECAL provides spoken forms and morphological analyses. To obtain fine-grained, tokenized morphological representations (MR) required for their task, the authors utilized the CALIMAEGY morphological analyzer from the CamelTools toolkit. The ECAL data was split into TRAIN, DEV, and EVAL sets. For evaluation, they created an OOV-DEV (Out-of-Vocabulary Development) set by extracting data points from DEV that were not present in TRAIN, aiming to test the model's generalization capabilities.

Key Insights and Ideas

  • Hybrid Approach Synergy: The most significant insight is that combining linguist-crafted rules with a neural network that learns the residuals of those rules can outperform either approach individually, especially in low-resource settings.
  • Rules as a Strong Prior: Even a minimal set of linguist-derived rules provides a strong starting point, significantly improving performance over a standalone neural model when data is scarce.
  • Neural Correction of Rule Errors: The neural network's ability to learn the specific errors made by the rules is more effective than training a neural model from scratch on raw MR-SF pairs when data is limited. This suggests that the residuals are "easier" for the neural model to learn.
  • Low-Resource Efficacy: The study demonstrates that this hybrid approach can achieve reasonable accuracy even with very limited training data (around 600-1000 examples), which is a critical finding for low-resource languages.

Key Results

  • In extremely low-resource settings (fewer than 1,000 training examples), the linguist's rules alone performed best.
  • However, as the training data size increased to around 600-1,000 examples, the TRNF-RES model (the transformer trained on residuals) began to outperform both the standalone rules and the standard TRNF model.
  • At a training size of 900 examples on the OOV-DEV set, TRNF-RES achieved an accuracy of 78.6%, compared to 65.9% for the rules and 73.2% for the standard TRNF.
  • Error analysis revealed that TRNF-RES was less prone to "hallucination errors" and more effectively recovered cases missed by the rules, often making less severe errors than the standalone TRNF.

Limitations and Future Work

  • Rule Coverage: The authors acknowledge that their set of rules, while minimal, does not cover all phonological processes. The performance of TRNF-RES is dependent on the quality and coverage of the initial rules.
  • Data Size Threshold: The point at which TRNF-RES surpasses the rules is around 600-1,000 examples. Below this threshold, the rules are superior.
  • Future Work: The authors plan to explore this approach in the context of child language acquisition, where children learn from limited input. They also intend to investigate scenarios involving limited data from one dialect and linguistic knowledge from a closely related dialect to study knowledge transfer.

Summary updated

BibTeX
@inproceedings{khalifa2024picking,
  author = {Salam Khalifa and Abdelrahim Qaddoumi and Ellen Broselow and Owen Rambow},
  title = {{Picking Up Where the Linguist Left Off: Mapping Morphology to Phonology through Learning the Residuals}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure