MiqraBERT: Regression-Based Sentence-BERT Finetuning for Biblical Hebrew Parallel Detection

2026-06-17 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

MiqraBERT is a Sentence-BERT model specifically finetuned from AlephBERT to detect verse-level semantic similarity in Biblical Hebrew. This model addresses limitations of traditional lexical overlap methods, which struggle with paraphrased or syntactically reworked textual reuse. Trained on 1,650 labeled verse and half-verse pairs, including 825 true parallels and 825 randomly sampled negatives, MiqraBERT uses cosine-similarity regression to create an embedding space where parallel verses cluster. Evaluation using Wasserstein distance and the overlap coefficient across ten random seeds shows a 2.7-fold improvement in distributional separation over the pre-trained baseline, reducing ambiguous overlap from approximately 24% to about 6%. While achieving 87.1% recall@10 for narrative synoptic parallels, its performance for poetic parallels remains below 9%, limiting its reliable application to narrative textual reuse. MiqraBERT is publicly available on Hugging Face.

Key takeaway

For NLP Engineers developing textual reuse detection systems for ancient or low-resource languages, you should recognize that lexical overlap methods are often inadequate. MiqraBERT demonstrates that finetuning a BERT-based model with regression for semantic similarity can significantly improve detection, particularly for narrative texts. You should consider adopting this methodology, carefully evaluating genre-specific performance, and explore the publicly available MiqraBERT for Biblical Hebrew applications to enhance your analytical capabilities.

Key insights

Finetuning a BERT model with regression for semantic similarity effectively detects textual reuse in ancient languages beyond lexical overlap.

Principles

• Lexical overlap methods fail with paraphrase.
• Regression finetunes embeddings for similarity.
• Genre significantly impacts detection accuracy.

Method

Finetune a pre-trained encoder using cosine-similarity regression on balanced true/negative parallel pairs to learn a semantic embedding space.

In practice

• Apply MiqraBERT for Biblical Hebrew narrative reuse.
• Tailor NLP models to specific text genres.
• Use distributional metrics for semantic evaluation.

Topics

• MiqraBERT
• Sentence-BERT
• Biblical Hebrew
• Textual Reuse Detection
• Semantic Similarity
• Regression Finetuning

Best for: NLP Engineer, AI Scientist, Research Scientist

Related on AIssential

• IHUBERT: Vector-Based Semantic Deduplication and Domain-Balanced Pretraining for Persian Resources — Semantic deduplication and domain-balanced pretraining significantly enhance Persian PLM performance across diverse NLU tasks.
• Intro to NLP with spaCy (1): Detecting programming languages — spaCy's linguistic features enable robust rule-based NLP systems beyond simple string matching.
• Token-Level Pun Location Using Multi-Layer BERT with Mixture of Experts — A two-stage BERT-MoE model effectively locates puns in Portuguese by specializing expert components.
• Resume Text Classifier using Classical NLP — Classical NLP and Logistic Regression can effectively classify resume text into job categories.
• A BART-based approach with hierarchical strategy for Vietnamese abstractive multi-document summarization — A BART-based hierarchical strategy with golden summary-driven shortening improves Vietnamese multi-document abstractive summarization, achieving 0.2468 ROUGE2-F1.
• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization detection.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.