Syntaxa at BAREC Shared Task 2025: BERTnParse - Fusion of BERT and Dependency Graphs for Readability Prediction
Year: 2025
Type: conference-paper
Abstract: We describe our system submission to Task 1 (Sentence-level Readability Assessment) of the BAREC Shared Task 2025 (Elmadani et al., 2025a), in the strict track.Task 1 requires predicting the readability level of an Arabic sentence on a scale from 1 (easiest) to 19 (hardest), reflecting reading difficulty.Our approach integrates contextual and syntactic information by combining pretrained BERT embeddings (Devlin et al., 2019) with a Graph Neural Network (GNN) (Zhou et al., 2021) over dependency parse trees (Kipf and Welling, 2017).Our hypothesis is that readability is influenced not only by word choice but also by syntactic complexity-especially in morphologically rich languages like Arabic (Habash, 2010).To capture both aspects, we represent each sentence as a dependency graph with BERT token embeddings as node features, and use a GNN to model the syntactic structure.Experimental results show that our syntax-aware model improves over a strong BERT baseline, highlighting the value of structural linguistic information for fine-grained readability classification. 1
Author: Ahmed Bahloul
Institution: Technical University of Munich
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037819