---
license: mit
---


# GeoNames Semantic Primes

## Dataset Overview

We propose a dataset at the core of our semantic towers methodology which combines vectorized knowledge graph information to augment a Retrieval-and-Generation (RAG) pipeline.

## Dataset Construction

The dataset is constructed by deriving and building the semantic tower - an ensemble of primitive semantic information related to a term - of 660 category classes related to geographical locations. These locations are themselves classified into 9 higher-level categories, e.g. H for stream, lake, and sea, and R for road and railroad and are derived from the original [GeoNames](https://www.geonames.org/) dataset.

The semantic tower encompasses information gathered from Wikidata, specifically:

- label
- instance of
- subclass of
- part of
- represents
- description

This information forms the smallest subset of knowledge needed to distinguish a term from another.

## Embeddings Generation

The vector embeddings are generated using the [General Text Embeddings (GTE)](https://huggingface.co/thenlper/gte-large) large model.