Enhancing Semantic Question Similarity in Arabic Using Hybrid Models and Knowledge Graph Alignment

PeerJ Computer Science  ·  2026

Ammar Al-Edhari; Mohsen Kahani;

HIGHLIGHT

Existing semantic similarity approaches have inherent trade-offs: KGs capture deep meaning but lack adaptability; corpus -based methods, such as word embeddings, rely on statistical patterns without deep semantic understanding; and deep learning models efficiently perform but lack interpretability. In addressing these limitations, this study proposes a hybrid SQS model that integrates knowledge graph embedding (KGE) with AraBERT embeddings to enrich semantic representation.

RELATED

Papers      Patents     Grants     Venues     Experts    Cite    Share    Save

ABSTRACT

Semantic question similarity (SQS) plays a crucial role in enhancing question answering (QA) systems, particularly for low-resource languages, such as Arabic, which present challenges due to complex morphology, semantic ambiguity, and limited linguistic resources. Despite prior efforts, knowledge graphs (KGs), which are structured networks representing entities and relationships, remain underutilized in improving semantic understanding and reducing ambiguity. Existing semantic similarity approaches have inherent trade-offs: KGs capture deep meaning but lack adaptability; corpus -based methods, such as word embeddings, rely on statistical patterns without deep semantic understanding; and deep learning models efficiently perform but lack interpretability. In addressing these limitations, this study proposes a hybrid SQS model that integrates knowledge graph embedding (KGE) with AraBERT embeddings to enrich semantic representation. The model utilizes a bidirectional long short-term memory network with an attention mechanism to predict question similarity, leveraging the strengths of each approach while mitigating their individual shortcomings. Furthermore, we align LughaNet (a variant of Arabic WordNet) with Wikidata using the similarity flooding (SF) algorithm to enhance KG-based semantic representations, creating a comprehensive and large-scale Arabic knowledge graph. Experimental evaluation on an augmented Mawdoo3 dataset demonstrates that our approach achieves an F1-score of 93.88%, outperforming state-of-the-art benchmarks, including GPT-based models. These findings confirm the effectiveness of hybrid models in capturing structured and contextual semantic relationships, advancing Arabic natural language processing (NLP) research. Additionally, this study introduces a large enriched Arabic knowledge graph, providing a valuable linguistic resource that lays the foundation for future advancements in low-resource language processing, with direct applications in automated QA systems, intelligent search, and recommendation engines.

 Paraphrase     Questionfy     Wikify     Simplify     Summarize