Eraynet-nirig

Eraynet-nirig is a baseline semantic retrieval model for Somali lexical search. It supports Somali, English, and Italian queries over a structured Somali dictionary dataset.

Overview

This model uses sentence embeddings to perform semantic search across lexical entries containing:

• abbreviation
• Somali term
• Italian gloss
• English gloss
• domain metadata

It is designed as a baseline retrieval system for Somali language technology, terminology search, and dictionary lookup.

Model Details

• Model: paraphrase-multilingual-MiniLM-L12-v2 (fine-tuned/used for embedding)
• Embedding Dimension: 384
• Training Data: 73 structured Somali lexical entries
• Languages: Somali, English, Italian

Features

• Exact and semantic lexical retrieval
• Multilingual query support: Somali, English, Italian
• Similarity scoring (cosine similarity)
• Confidence labels: high (≥0.7), medium (≥0.5), low (<0.5)
• Top-k results (default: 5)
• Domain-aware search

Files

• build_embeddings.py: builds vector embeddings from the lexical dataset
• search.py: runs semantic search with confidence scoring
• ai_model/embeddings.npy: stored embeddings (NumPy format)
• ai_model/search_data.csv: structured lexical entries

Example Usage

from search import search

# Search for a term
results = search("medicine")
print(results)

Example output:

rank    somali   english   italian   domain    similarity_score confidence_label
1       Daawo    medicine  medicina  medicine  0.8542            high
2       ...      ...       ...       ...       ...              ...

Example Queries

• medicine → Daawo
• politics → siyaasad
• botany → Botani

Installation

pip install -r requirements.txt

Requirements

• sentence-transformers
• pandas
• numpy
• scikit-learn
• fastapi (optional, for API)
• uvicorn (optional, for API)

Intended Use

This model is suitable for:

• Somali dictionary search
• terminology lookup
• NLP preprocessing support
• lexical search in multilingual Somali applications

Limitations

• Small baseline dataset (73 entries)
• Not a generative model
• Not a translation model
• Similarity scores are embedding-based, not calibrated probabilities
• Confidence labels are based on similarity thresholds, not statistical certainty

Future Work

• expand dataset size with cleaned dictionary entries
• add part-of-speech tagging
• add richer domain annotations
• support example sentence retrieval
• train a Somali-English translation model from parallel sentence pairs

Version

• v0.1: Current baseline - 73 entries, multilingual semantic retrieval

Citation

If you use this model, please cite:

Eraynet-nirig: Somali Multilingual Lexical Retrieval Baseline (2026)