SIRIS-Lab/impuls-wikidata-kb

---
license: apache-2.0
language:
- ca
- es
- en
- it
task_categories:
- feature-extraction
- text-generation
tags:
- knowledge-base
- wikidata
- multilingual
- R&D
- query-expansion
- semantic-search
- catalan
- spanish
- italian
- AINA
size_categories:
- 1K<n<10K
---

# IMPULS R&D Knowledge Base

A multilingual knowledge base of 4,265 R&D concepts derived from Wikidata, designed for query expansion in scientific and research project search systems.

## Dataset Description

This knowledge base was created as part of the **IMPULS project** (AINA Challenge 2024), a collaboration between [SIRIS Academic](https://sirisacademic.com/) and [Generalitat de Catalunya](https://web.gencat.cat/) to build a multilingual semantic search system for R&D ecosystems.

The KB contains scientific and technological concepts with:
- **Multilingual labels** in Catalan, Spanish, English, and Italian
- **Aliases/synonyms** for each language
- **Definitions** where available
- **Hierarchical relationships** (instance_of, subclass_of) linking to Wikidata

### Use Cases

- **Query Expansion**: Expand search queries with synonyms and related terms across languages
- **Multilingual Search**: Find equivalent terms across CA/ES/EN/IT
- **Concept Navigation**: Traverse hierarchical relationships for broader/narrower terms
- **Named Entity Linking**: Link mentions to Wikidata identifiers

## Dataset Structure

### Format

JSONL (JSON Lines) - one concept per line.

### Schema

```json
{
"keyword": "machine learning",
"wikidata_id": "Q2539",
"languages": {
"ca": {
"label": "aprenentatge automàtic",
"description": "branca de la intel·ligència artificial",
"also_known_as": ["aprenentatge de màquines", "ML"]
},
"es": {
"label": "aprendizaje automático",
"description": "rama de la inteligencia artificial",
"also_known_as": ["aprendizaje de máquina", "ML"]
},
"en": {
"label": "machine learning",
"description": "branch of artificial intelligence",
"also_known_as": ["ML", "statistical learning"]
},
"it": {
"label": "apprendimento automatico",
"description": "ramo dell'intelligenza artificiale",
"also_known_as": []
}
},
"instance_of": [
{"id": "Q11660", "label": "artificial intelligence"}
],
"subclass_of": [
{"id": "Q11660", "label": "artificial intelligence"},
{"id": "Q816264", "label": "computational learning theory"}
]
}
```

### Field Descriptions

| Field | Type | Description |
|-------|------|-------------|
| `keyword` | string | Primary English keyword |
| `wikidata_id` | string | Wikidata entity ID (Q-number) |
| `languages` | object | Multilingual labels, descriptions, and aliases |
| `languages.{lang}.label` | string | Primary label in that language |
| `languages.{lang}.description` | string | Short description/definition |
| `languages.{lang}.also_known_as` | array | Alternative names/synonyms |
| `instance_of` | array | Wikidata instance_of relations |
| `subclass_of` | array | Wikidata subclass_of relations (for hierarchy traversal) |

## Statistics

| Metric | Value |
|--------|-------|
| Total concepts | 4,265 |
| With Catalan labels | ~4,200 |
| With Spanish labels | ~4,250 |
| With English labels | 4,265 |
| With Italian labels | ~4,100 |
| With subclass_of relations | ~3,590 |
| Unique parent concepts | ~770 (in KB) |

### Domain Coverage

The KB focuses on R&D-relevant concepts including:
- **Technology**: AI, blockchain, IoT, robotics, quantum computing
- **Science**: biotechnology, nanotechnology, materials science
- **Health**: medical devices, diagnostics, pharmaceuticals
- **Energy**: renewables, hydrogen, energy storage
- **Environment**: climate, sustainability, circular economy
- **Industry**: manufacturing, automation, Industry 4.0

## Examples

### Example 1: Technology Concept

```json
{
"keyword": "blockchain",
"wikidata_id": "Q20514253",
"languages": {
"ca": {
"label": "cadena de blocs",
"description": "estructura de dades distribuïda",
"also_known_as": ["blockchain"]
},
"es": {
"label": "cadena de bloques",
"description": "base de datos distribuida",
"also_known_as": ["blockchain"]
},
"en": {
"label": "blockchain",
"description": "distributed database technology",
"also_known_as": ["block chain", "distributed ledger"]
}
},
"subclass_of": [
{"id": "Q8513", "label": "database"}
]
}
```

### Example 2: Health Concept

```json
{
"keyword": "patient",
"wikidata_id": "Q181600",
"languages": {
"ca": {
"label": "pacient",
"description": "",
"also_known_as": []
},
"es": {
"label": "paciente",
"description": "persona que recibe tratamiento para un problema de salud",
"also_known_as": ["pacientes", "enfermo"]
},
"en": {
"label": "patient",
"description": "person who takes a medical treatment",
"also_known_as": ["patients", "medical patient", "human patient"]
}
},
"instance_of": [
{"id": "Q214339", "label": "role"}
],
"subclass_of": [
{"id": "Q12722854", "label": "sick person"},
{"id": "Q852835", "label": "customer"}
]
}
```

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

dataset = load_dataset("SIRIS-Lab/impuls-wikidata-kb")
kb = dataset["train"]

print(f"Loaded {len(kb)} concepts")
```

### Query Expansion Example

```python
def find_concept(kb, query):
"""Find concept by keyword or label."""
query_lower = query.lower()
for concept in kb:
if concept["keyword"].lower() == query_lower:
return concept
for lang in ["en", "es", "ca"]:
if concept["languages"].get(lang, {}).get("label", "").lower() == query_lower:
return concept
return None

def get_expansions(concept):
"""Get all labels and aliases for a concept."""
expansions = set()
for lang_data in concept["languages"].values():
if lang_data.get("label"):
expansions.add(lang_data["label"])
for alias in lang_data.get("also_known_as", []):
expansions.add(alias)
return expansions

# Example
concept = find_concept(kb, "machine learning")
if concept:
print(f"Wikidata ID: {concept['wikidata_id']}")
print(f"Expansions: {get_expansions(concept)}")
# Output: {'machine learning', 'ML', 'aprenentatge automàtic', 'aprendizaje automático', ...}
```

### Building a Lookup Index

```python
def build_kb_index(kb):
"""Build wikidata_id -> concept index for fast parent lookup."""
return {concept["wikidata_id"]: concept for concept in kb}

kb_index = build_kb_index(kb)

# Get parent concepts
concept = find_concept(kb, "deep learning")
for parent in concept.get("subclass_of", []):
parent_concept = kb_index.get(parent["id"])
if parent_concept:
print(f"Parent: {parent_concept['keyword']}")
```

## Data Collection

The knowledge base was built by:

1. **Seed Selection**: Identifying R&D-relevant concepts from project databases (RIS3CAT, OpenAIRE, CORDIS)
2. **Wikidata Extraction**: Querying Wikidata API for each concept's labels, aliases, and relations
3. **Multilingual Enrichment**: Ensuring coverage across CA/ES/EN/IT
4. **Hierarchy Validation**: Filtering subclass_of relations to include only parents present in the KB
5. **Quality Control**: Manual review of key domain concepts

## Integration with IMPULS

This KB is used by the [IMPULS Query Parser](https://huggingface.co/SIRIS-Lab/impuls-salamandra-7b-query-parser) for:

- **Query Expansion**: Adding multilingual synonyms to search queries
- **Cross-lingual Search**: Finding Spanish projects with Catalan queries
- **Concept Navigation**: Broadening searches via parent concepts

## Limitations

- **Domain Focus**: Optimized for R&D/scientific concepts; general vocabulary coverage is limited
- **Language Coverage**: Best coverage in English; some concepts may lack labels in other languages
- **Temporal Snapshot**: Based on Wikidata as of late 2024; may not reflect recent additions
- **Hierarchy Depth**: Only direct parents (subclass_of) are included; transitive closure not computed

## Citation

```bibtex
@misc{impuls-wikidata-kb-2024,
author = {SIRIS Academic},
title = {IMPULS R&D Knowledge Base: Multilingual Wikidata Concepts for Query Expansion},
year = {2024},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/SIRIS-Lab/impuls-wikidata-kb}}
}
```

## Acknowledgments

- **[Wikidata](https://www.wikidata.org/)** - Source knowledge graph
- **[Barcelona Supercomputing Center (BSC)](https://www.bsc.es/)** - AINA project infrastructure
- **[Generalitat de Catalunya](https://web.gencat.cat/)** - Funding and RIS3-MCAT platform

## License

Apache 2.0

## Related Resources

- **Query Parser Model**: [SIRIS-Lab/impuls-salamandra-7b-query-parser](https://huggingface.co/SIRIS-Lab/impuls-salamandra-7b-query-parser)
- **Query Parsing Dataset**: [SIRIS-Lab/impuls-query-parsing](https://huggingface.co/datasets/SIRIS-Lab/impuls-query-parsing)
- **Project Repository**: [github.com/sirisacademic/aina-impulse](https://github.com/sirisacademic/aina-impulse)