Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Hybrid Protocol for Large-Scale Semantic Dataset Generation in Low-Resource Languages: The Turkish Semantic Relations Corpus

Ebubekir Tosun, Mehmet Emin Buldur, Özay Ezerceli, Mahmoud ElHussieni

TL;DR

Turkish suffers from a shortage of large-scale semantic relation data, hindering NLP development for morphology-rich languages. The authors propose a scalable three-phase hybrid protocol that combines FastText embedding-based clustering, LLM-driven semantic enrichment via Gemini 2.5-Flash, and dictionary-based validation to generate a Turkish Semantic Relations Corpus with 843,000 annotated pairs across synonyms, antonyms, and co-hyponyms. The dataset achieves strong downstream performance, with 90% top-1 retrieval in embedding tasks and 90% F1-macro in relation classification, while remaining cost-efficient (~$65 for enrichment and ~16k dictionary-validated synonyms). The approach is designed to generalize to other low-resource languages and is released publicly to accelerate semantic NLP research beyond Turkish, addressing data scarcity without extensive manual annotation.

Abstract

We present a hybrid methodology for generating large-scale semantic relationship datasets in low-resource languages, demonstrated through a comprehensive Turkish semantic relations corpus. Our approach integrates three phases: (1) FastText embeddings with Agglomerative Clustering to identify semantic clusters, (2) Gemini 2.5-Flash for automated semantic relationship classification, and (3) integration with curated dictionary sources. The resulting dataset comprises 843,000 unique Turkish semantic pairs across three relationship types (synonyms, antonyms, co-hyponyms) representing a 10x scale increase over existing resources at minimal cost ($65). We validate the dataset through two downstream tasks: an embedding model achieving 90% top-1 retrieval accuracy and a classification model attaining 90% F1-macro. Our scalable protocol addresses critical data scarcity in Turkish NLP and demonstrates applicability to other low-resource languages. We publicly release the dataset and models.

A Hybrid Protocol for Large-Scale Semantic Dataset Generation in Low-Resource Languages: The Turkish Semantic Relations Corpus

TL;DR

Turkish suffers from a shortage of large-scale semantic relation data, hindering NLP development for morphology-rich languages. The authors propose a scalable three-phase hybrid protocol that combines FastText embedding-based clustering, LLM-driven semantic enrichment via Gemini 2.5-Flash, and dictionary-based validation to generate a Turkish Semantic Relations Corpus with 843,000 annotated pairs across synonyms, antonyms, and co-hyponyms. The dataset achieves strong downstream performance, with 90% top-1 retrieval in embedding tasks and 90% F1-macro in relation classification, while remaining cost-efficient (~$65 for enrichment and ~16k dictionary-validated synonyms). The approach is designed to generalize to other low-resource languages and is released publicly to accelerate semantic NLP research beyond Turkish, addressing data scarcity without extensive manual annotation.

Abstract

We present a hybrid methodology for generating large-scale semantic relationship datasets in low-resource languages, demonstrated through a comprehensive Turkish semantic relations corpus. Our approach integrates three phases: (1) FastText embeddings with Agglomerative Clustering to identify semantic clusters, (2) Gemini 2.5-Flash for automated semantic relationship classification, and (3) integration with curated dictionary sources. The resulting dataset comprises 843,000 unique Turkish semantic pairs across three relationship types (synonyms, antonyms, co-hyponyms) representing a 10x scale increase over existing resources at minimal cost ($65). We validate the dataset through two downstream tasks: an embedding model achieving 90% top-1 retrieval accuracy and a classification model attaining 90% F1-macro. Our scalable protocol addresses critical data scarcity in Turkish NLP and demonstrates applicability to other low-resource languages. We publicly release the dataset and models.
Paper Structure (49 sections, 3 equations, 4 figures, 6 tables)

This paper contains 49 sections, 3 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Context Preparation
  3. LLM-Based Semantic Enrichment
  4. Dictionary Integration
  5. Related Work
  6. Semantic Resources for Low-Resource Languages
  7. LLM-Augmented Dataset Generation
  8. Embedding-Based Clustering for Semantic Organization
  9. Antonym-Synonym Distinction in Vector Spaces
  10. Methodology
  11. Phase I: Context Preparation
  12. Initial Term Collection
  13. Embedding Generation
  14. Semantic Clustering
  15. Phase II: LLM-Based Semantic Enrichment
  16. ...and 34 more sections

Figures (4)

  • Figure 1: Overview of the three-phase hybrid protocol for semantic dataset generation. Phase I establishes semantic structure through FastText embeddings (110K terms) and agglomerative clustering (13K clusters with distance threshold 0.4). Phase II employs Gemini 2.5-Flash to classify intra-cluster relationships into three types: synonyms (strict equivalence), antonyms (semantic opposition), and co-hyponyms (thematic relatedness). Phase III integrates 16K high-precision dictionary pairs and outputs 843K annotated pairs in JSONL format.
  • Figure 2: Embedding model training progression across seven model candidates. The multilingual-e5-large (v1) variant achieves best performance with 90% retrieval accuracy.
  • Figure 3: System prompt template for LLM-based semantic relationship classification in Phase II. The prompt enforces strict categorization rules and structured JSON output to ensure consistency across all 13,000 processed clusters.
  • Figure 4: System prompt template for NER-based term augmentation in Phase I. This prompt enabled extraction of 33,000 additional domain-specific terms from legal corpora, expanding the initial lexicon from 77,000 to 110,000 entries.