labse

In this project, we implemented a topic detection system on Twitter. This system reads tweets from a data stream and assigns them to one of the existing clusters or a new one. Each cluster acts as an agent, which makes the proposed approach a multi-agent system. There is also a coordinator, who monitors the whole system and coordinates the agent.

Convert LaBSE model from TF Hub to PyTorch.

AQUILIGN is a multilingual alignment and collation tool for medieval texts. It uses phrase-level segmentation and contextual alignment based on BERT models, with applications in historical linguistics, philology, and historical NLP.

A curated collection of manually aligned multilingual texts for sentence embedding, LaBSE fine-tuning, and historical text alignment.

Modular pipeline and interface for multilingual tag prediction on RKD art exhibition catalogues, built with LaBSE and Streamlit.

LeBSE — legal-domain sentence embeddings adapted from LaBSE (training + held-out benchmark harness)

English-Sinhala multilingual word embedding alignment resources

A comprehensive framework for detecting hate speech in the low-resource Telugu language using multimodal (audio and text) analysis

Evidence-first multimodal RAG framework for grounded clinical decision support under Hinglish (Hindi–English) code-switched queries, aligned with authoritative medical reports and optional radiological evidence.

High-precision bilingual text and subtitle alignment workbench using Google LaBSE neural embeddings and dynamic programming

Final Year Research & Developmental Project focused on detecting Urdu Bias in native Urdu news text.

Multilingual sentence embedding evaluation for Finnish-English course similarity detection. Bachelor's thesis, JAMK 2026. Includes dataset, notebooks, production FastAPI service, and live demo.

Multilingual news aggregator for Bangladeshi sources — LaBSE embeddings for semantic vectors, Qdrant for retrieval, a five-service Kafka pipeline, and a cascaded LLM-judge (Mistral → Claude Haiku → Claude Sonnet) for deduplication at calibrated cosine thresholds. Validated across 800+ article pairs.

Comparative evaluation of neural machine translation architectures (MarianMT, mBART-50, NLLB-200, GPT-2) across German, Spanish, and Arabic. Includes multi-metric scoring (BLEU, chrF, METEOR, BERTScore, LaBSE), cross-lingual semantic similarity analysis, LLM-as-judge evaluation via LangChain, and WMT14/OPUS-100 benchmark runs.

Multilingual alignment and collation of the De Regimine Principum in latin and vernacular

Align Parallel Sentence of 104 Languages with the help of mBERT and LaBSE

Comparative study of five sentence-embedding models (USE, SBERT, MPNet, MiniLM, LaBSE) for extractive news summarization with LexRank; LaBSE ranks best on ROUGE across the BBC News dataset.

Powering Silah's smart features!

AI-powered scoring tool for the Emotional and Classic Hayling Test

Reference implementation and synthetic demo for multilingual semantic bridge construction between local clinical knowledge graphs and standard clinical ontologies (NLP4KGC 2026).