Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLMs Got Rhythm? Hybrid Phonological Filtering for Greek Poetry Rhyme Detection and Generation

Stergios Chatzikyriakidis

TL;DR

Phonologically grounded tasks like rhyme demand more than raw text modeling, especially for low-resource languages such as Modern Greek. The authors present a hybrid neural-symbolic system that couples LLM-based generation with a deterministic phonological engine, enabling reliable rhyme identification and generation through a Generate-Verify-Refine loop. They provide a 40,576-rhyme corpus, achieve 100% verification accuracy for identification, and reach 73.1% valid generation with the verifier, outperforming pure LLM generation. The work offers substantial resources and demonstrates that integrating phonological rules with neural models can significantly advance phonology-aware NLP for Greek and beyond.

Abstract

Large Language Models (LLMs), despite their remarkable capabilities across NLP tasks, struggle with phonologically-grounded phenomena like rhyme detection and generation. This is even more evident in lower-resource languages such as Modern Greek. In this paper, we present a hybrid system that combines LLMs with deterministic phonological algorithms to achieve accurate rhyme identification/analysis and generation. Our approach implements a comprehensive taxonomy of Greek rhyme types, including Pure, Rich, Imperfect, Mosaic, and Identical Pre-rhyme Vowel (IDV) patterns, and employs an agentic generation pipeline with phonological verification. We evaluate multiple prompting strategies (zero-shot, few-shot, Chain-of-Thought, and RAG-augmented) across several LLMs including Claude 3.7 and 4.5, GPT-4o, Gemini 2.0 and open-weight models like Llama 3.1 8B and 70B and Mistral Large. Results reveal a significant "Reasoning Gap": while native-like models (Claude 3.7) perform intuitively (40\% accuracy in identification), reasoning-heavy models (Claude 4.5) achieve state-of-the-art performance (54\%) only when prompted with Chain-of-Thought. Most critically, pure LLM generation fails catastrophically (under 4\% valid poems), while our hybrid verification loop restores performance to 73.1\%. We release our system and a crucial, rigorously cleaned corpus of 40,000+ rhymes, derived from the Anemoskala and Interwar Poetry corpora, to support future research.

LLMs Got Rhythm? Hybrid Phonological Filtering for Greek Poetry Rhyme Detection and Generation

TL;DR

Phonologically grounded tasks like rhyme demand more than raw text modeling, especially for low-resource languages such as Modern Greek. The authors present a hybrid neural-symbolic system that couples LLM-based generation with a deterministic phonological engine, enabling reliable rhyme identification and generation through a Generate-Verify-Refine loop. They provide a 40,576-rhyme corpus, achieve 100% verification accuracy for identification, and reach 73.1% valid generation with the verifier, outperforming pure LLM generation. The work offers substantial resources and demonstrates that integrating phonological rules with neural models can significantly advance phonology-aware NLP for Greek and beyond.

Abstract

Large Language Models (LLMs), despite their remarkable capabilities across NLP tasks, struggle with phonologically-grounded phenomena like rhyme detection and generation. This is even more evident in lower-resource languages such as Modern Greek. In this paper, we present a hybrid system that combines LLMs with deterministic phonological algorithms to achieve accurate rhyme identification/analysis and generation. Our approach implements a comprehensive taxonomy of Greek rhyme types, including Pure, Rich, Imperfect, Mosaic, and Identical Pre-rhyme Vowel (IDV) patterns, and employs an agentic generation pipeline with phonological verification. We evaluate multiple prompting strategies (zero-shot, few-shot, Chain-of-Thought, and RAG-augmented) across several LLMs including Claude 3.7 and 4.5, GPT-4o, Gemini 2.0 and open-weight models like Llama 3.1 8B and 70B and Mistral Large. Results reveal a significant "Reasoning Gap": while native-like models (Claude 3.7) perform intuitively (40\% accuracy in identification), reasoning-heavy models (Claude 4.5) achieve state-of-the-art performance (54\%) only when prompted with Chain-of-Thought. Most critically, pure LLM generation fails catastrophically (under 4\% valid poems), while our hybrid verification loop restores performance to 73.1\%. We release our system and a crucial, rigorously cleaned corpus of 40,000+ rhymes, derived from the Anemoskala and Interwar Poetry corpora, to support future research.
Paper Structure (38 sections, 1 figure, 6 tables, 1 algorithm)

This paper contains 38 sections, 1 figure, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Rhyme in Modern Greek Poetry
  4. Stress-based Classification
  5. Computational Approaches to Rhyme
  6. Dataset
  7. System Architecture
  8. Phonological Engine
  9. Rhyme Classification Module
  10. LLM Integration
  11. Agentic Generation Pipeline
  12. Experimental Setup
  13. Results
  14. Rhyme Identification Results
  15. Qualitative Analysis
  16. ...and 23 more sections

Figures (1)

  • Figure 1: Hybrid system architecture. Left: Identification combines LLM predictions with Engine-generated ground truth for validation. Right: Generation uses the Engine to verify and refine LLM outputs.