Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Cosine Similarity: Taming Semantic Drift and Antonym Intrusion in a 15-Million Node Turkish Synonym Graph

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

TL;DR

This work addresses the instability of distributional similarity for synonym detection in morphologically rich Turkish, where antonyms and polysemy confound embeddings. It presents a scalable, disambiguation-focused pipeline combining a large-scale LLM-augmented dataset, a specialized three-way discriminator, and a novel soft-to-hard clustering approach that uses topological voting to curb semantic drift. By processing ~15 million terms, evaluating ~520 million candidate relations, and yielding ~2.9 million high-precision synonym clusters, the method delivers a robust Turkish synonym graph suitable for high-precision search and generation. The results show that language-specific, supervised discrimination and drift-aware clustering can outperform multilingual baselines, offering a practical resource and methodology for morphologically rich languages with sparse lexical resources.

Abstract

Neural embeddings have a notorious blind spot: they can't reliably tell synonyms apart from antonyms. Consequently, increasing similarity thresholds often fails to prevent opposites from being grouped together. We've built a large-scale semantic clustering system specifically designed to tackle this problem head on. Our pipeline chews through 15 million lexical items, evaluates a massive 520 million potential relationships, and ultimately generates 2.9 million high-precision semantic clusters. The system makes three primary contributions. First, we introduce a labeled dataset of 843,000 concept pairs spanning synonymy, antonymy, and co-hyponymy, constructed via Gemini 2.5-Flash LLM augmentation and verified using human-curated dictionary resources. Second, we propose a specialized three-way semantic relation discriminator that achieves 90% macro-F1, enabling robust disambiguation beyond raw embedding similarity. Third, we introduce a novel soft-to-hard clustering algorithm that mitigates semantic drift preventing erroneous transitive chains (e.g., hot -> spicy -> pain -> depression) while simultaneously resolving polysemy. Our approach employs a topology-aware two-stage expansion-pruning procedure with topological voting, ensuring that each term is assigned to exactly one semantically coherent cluster. The resulting resource enables high-precision semantic search and retrieval-augmented generation, particularly for morphologically rich and low-resource languages where existing synonym databases remain sparse.

Beyond Cosine Similarity: Taming Semantic Drift and Antonym Intrusion in a 15-Million Node Turkish Synonym Graph

TL;DR

This work addresses the instability of distributional similarity for synonym detection in morphologically rich Turkish, where antonyms and polysemy confound embeddings. It presents a scalable, disambiguation-focused pipeline combining a large-scale LLM-augmented dataset, a specialized three-way discriminator, and a novel soft-to-hard clustering approach that uses topological voting to curb semantic drift. By processing ~15 million terms, evaluating ~520 million candidate relations, and yielding ~2.9 million high-precision synonym clusters, the method delivers a robust Turkish synonym graph suitable for high-precision search and generation. The results show that language-specific, supervised discrimination and drift-aware clustering can outperform multilingual baselines, offering a practical resource and methodology for morphologically rich languages with sparse lexical resources.

Abstract

Neural embeddings have a notorious blind spot: they can't reliably tell synonyms apart from antonyms. Consequently, increasing similarity thresholds often fails to prevent opposites from being grouped together. We've built a large-scale semantic clustering system specifically designed to tackle this problem head on. Our pipeline chews through 15 million lexical items, evaluates a massive 520 million potential relationships, and ultimately generates 2.9 million high-precision semantic clusters. The system makes three primary contributions. First, we introduce a labeled dataset of 843,000 concept pairs spanning synonymy, antonymy, and co-hyponymy, constructed via Gemini 2.5-Flash LLM augmentation and verified using human-curated dictionary resources. Second, we propose a specialized three-way semantic relation discriminator that achieves 90% macro-F1, enabling robust disambiguation beyond raw embedding similarity. Third, we introduce a novel soft-to-hard clustering algorithm that mitigates semantic drift preventing erroneous transitive chains (e.g., hot -> spicy -> pain -> depression) while simultaneously resolving polysemy. Our approach employs a topology-aware two-stage expansion-pruning procedure with topological voting, ensuring that each term is assigned to exactly one semantically coherent cluster. The resulting resource enables high-precision semantic search and retrieval-augmented generation, particularly for morphologically rich and low-resource languages where existing synonym databases remain sparse.
Paper Structure (57 sections, 5 equations, 3 figures, 4 tables, 1 algorithm)

This paper contains 57 sections, 5 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. A Large-Scale Disambiguated Training Dataset
  3. A Specialized Multi-Class Discriminator
  4. A Novel Disambiguation-Aware Clustering Algorithm
  5. Related Work
  6. Constraint-Based Approaches to Antonymy
  7. Graph-Based Clustering and Semantic Drift
  8. Positioning of the Present Work
  9. Methodology
  10. Phase 1: Embedding Model Training and Initial Data Generation
  11. Embedding Model Specialization
  12. Concept-Term List Expansion
  13. LLM-Augmented Dataset Generation
  14. Dictionary Integration
  15. Phase 2: Large-Scale Candidate Generation via GPU-Accelerated Vector Search
  16. ...and 42 more sections

Figures (3)

  • Figure 1: Classification model training architecture and weighted loss calculation flow. Word pairs are tokenized and processed through a transformer encoder, with class weights computed from inverse frequencies to address label imbalance.
  • Figure 2: Training dynamics of the semantic relation classifier. Subfigures (a) and (b) demonstrate the convergence of cross-entropy loss on training and validation sets, respectively. Subfigure (c) shows the corresponding rise in F1-Macro scores, where the turkish-e5-large variant (green line) demonstrates superior efficiency, reaching a peak score of 0.90 significantly faster than competing baselines.
  • Figure 3: Per-class F1 scores during training. Co-hyponym classification converges quickly to 0.94, while synonym detection shows more variation across models, stabilizing at 0.83 for the best model.