Introduction
The Persian language features a unique grammatical construct known as the Ezafe, an unstressed morpheme (pronounced -e or -ye) that serves as a structural linker between a head noun and its modifiers. Despite its critical role in defining phrase boundaries and syntactic relationships, the Ezafe is rarely represented in standard Persian orthography. This omission introduces significant ambiguity for computational models, impacting downstream tasks such as Part-of-Speech (POS) tagging and syntactic parsing. This paper explores the application of modern Transformer architectures to recognize Ezafe markers and evaluates their utility in improving POS tagging.
Methodology
The authors frame Ezafe recognition as a sequence labeling task, evaluating a range of models from traditional statistical baselines to state-of-the-art neural architectures:
- Conditional Random Fields (CRF): Utilizing hand-crafted features.
- Recurrent Neural Networks: Bidirectional LSTMs (BLSTM) and BLSTM+CNNs, the latter incorporating character-level embeddings to better capture Persian morphology.
- Transformers: Multilingual BERT and XLM-RoBERTa.
Experiments were conducted using the Bijankhan corpus, a large-scale collection of 10 million Persian tokens. To ensure evaluation accuracy, the authors manually re-annotated a subset of the test data to mitigate existing labeling noise in the corpus.
Ezafe Recognition Performance
The experimental results demonstrate a clear advantage for Transformer-based models. XLM-RoBERTa achieved a new state-of-the-art F1-score of 98.14%, outperforming previous CRF-based methods by a significant margin.
The inclusion of a CNN layer in the BLSTM architecture also yielded improvements, highlighting the importance of morphological information in identifying Ezafe markers.
Impact on Part-of-Speech Tagging
A key contribution of this study is the investigation into whether explicit Ezafe information improves POS tagging. The authors tested three configurations: a single-task POS tagger, an "input feature" model where Ezafe predictions are fed into the tagger, and a multi-task model predicting both simultaneously.
The results reveal a fascinating divergence between model types. While traditional models like CRFs and BLSTMs saw performance gains when provided with explicit Ezafe information, Transformer models (BERT and XLM-RoBERTa) showed negligible or even slightly negative changes.
Discussion: Implicit Syntax
The lack of improvement in Transformers when given explicit Ezafe features suggests that these models already capture the underlying syntactic dependencies during their pre-training phase. The attention mechanisms within Transformers likely learn to represent the relationships that Ezafe markers signify, making explicit feature engineering redundant. This "implicit syntax" allows Transformers to achieve superior POS tagging performance (97.09% F1 for XLM-RoBERTa) without external linguistic markers.
Conclusion
This work establishes a new benchmark for Persian Ezafe recognition and provides evidence that modern Transformers are inherently capable of capturing complex syntactic structures in Persian. While explicit Ezafe recognition remains vital for tasks like text-to-speech synthesis, its role as a feature for high-capacity neural models in syntactic tasks may be superseded by the implicit learning capabilities of the Transformer architecture.
Summary updated