Morgan-Tanglish-v7: Sentence Embeddings for Romanised Tamil-English Code-Mixed Text

AbstractWe introduce Morgan-Tanglish-v7, a novel 118M-parameter sentence embedding model optimized specifically for Romanised Tamil-English (Tanglish) code-mixed text, addressing a critical gap in contemporary cross-lingual semantic representations. Existing state-of-the-art multilingual and English-centric models—including all-MiniLM-L6-v2, BAAI/bge-m3, and L3Cube-IndicSBERT—demonstrate catastrophic failure modes when evaluating code-switched text, primarily due to their reliance on formal native-script corpora and a lack of exposure to non-standard Latin-script syntactic variations.  To resolve the structural and lexical divergence inherent to code-mixed communication, we deliver three major contributions to the open-source NLP ecosystem: 1. Tanglish-Corpus-185k: A highly curated dataset comprising 185,973 domain-diverse sentences. This corpus represents an 11.8x scale expansion over the previous largest benchmark and was engineered via a robust 5-gate morphological filtering pipeline. This framework bypasses standard n-gram language identifiers in favor of a specialized 1,018-token Tamil lexicon and 27 granular grammatical pattern detectors.2. Concept Cluster Training (CCT): A novel algorithmic augmentation technique designed to systematically mitigate the "lexical gap problem." By generating C(n,2) positive pairs from n maximally diverse structural expressions of identical underlying concepts, CCT enforces mathematical invariance between surface-level vocabulary and core semantic intent.3. TanglishSTS Benchmark: The first human-annotated Semantic Textual Similarity (STS) benchmark for Romanised Tamil-English, containing 325 carefully calibrated pairs validated by native speakers. Empirical evaluations demonstrate that Morgan-Tanglish-v7 achieves a Spearman rank correlation coefficient of 0.8689 on TanglishSTS. This significantly outperforms all-MiniLM-L6-v2 by +15.79 points, L3Cube-IndicSBERT by +10.47 points, and the 5x larger BAAI/bge-m3 (570M) model by +11.06 points, establishing a new state-of-the-art for low-resource, code-mixed sentence-level representation learning. All models, data pipelines, evaluation benchmarks, and interactive deployment spaces are publicly accessible at huggingface.co.