README.md

---
license: apache-2.0
language:
  - ca
  - es
  - en
task_categories:
  - text-generation
tags:
  - query-parsing
  - structured-output
  - json-generation
  - multilingual
  - catalan
  - spanish
  - R&D
  - semantic-search
  - AINA
  - semantic-parsing
size_categories:
  - n<1K
---

# IMPULS Query Parsing Dataset

A multilingual dataset for training and evaluating query parsing models that convert natural language queries into structured JSON for R&D project semantic search.

## Dataset Description

This dataset 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 dataset contains natural language queries in **Catalan, Spanish, and English** paired with their structured JSON representations, designed for training models to:
- Extract semantic search terms from natural language
- Identify structured filters (funding programme, year, location, organization type)
- Detect query language and intent

### Supported Tasks

- **Query Parsing / Semantic Parsing**: Convert natural language to structured JSON
- **Information Extraction**: Extract entities and filters from queries
- **Multilingual NLU**: Understanding queries across CA/ES/EN

## Dataset Structure

### Data Splits

| Split | Examples | Description |
|-------|----------|-------------|
| `train` | 682 | Synthetic, template-generated queries |
| `test` | 100 | Real queries from domain experts |

### Schema

Each example contains a structured JSON with the following fields:

```json
{
  "doc_type": "projects",
  "filters": {
    "programme": "Horizon 2020 | FEDER | SIFECAT | null",
    "funding_level": "string | null",
    "year": ">=2020 | 2015-2020 | null",
    "location": "Catalunya | Spain | null",
    "location_level": "region | province | country | null"
  },
  "organisations": [
    {
      "type": "university | research_center | hospital | company | null",
      "name": "UPC | CSIC | null",
      "location": "Barcelona | null",
      "location_level": "province | region | null"
    }
  ],
  "semantic_query": "intel·ligència artificial salut",
  "query_rewrite": "Human-readable interpretation of the query",
  "meta": {
    "id": "TRAIN_001",
    "source": "synthetic | expert",
    "lang": "CA | ES | EN",
    "original_query": "The original natural language query",
    "intent": "Discover | Quantify",
    "style": "Concise | Verbose | Technical",
    "components": ["Content", "Programme", "Year", "Location"],
    "resolvability": "Direct | Adapted | Partial",
    "notes": "Optional notes about interpretation"
  }
}
```

### Field Descriptions

| Field | Description |
|-------|-------------|
| `doc_type` | Document type to search (always "projects") |
| `filters.programme` | Funding programme (H2020, Horizon Europe, FEDER, SIFECAT, etc.) |
| `filters.year` | Year filter (single year, range, or comparison like ">=2020") |
| `filters.location` | Geographic filter |
| `filters.location_level` | Geographic granularity (country, region, province) |
| `organisations` | List of organization filters with type, name, and location |
| `semantic_query` | Core thematic content for vector search |
| `query_rewrite` | Human-readable interpretation |
| `meta.original_query` | The original natural language query |
| `meta.lang` | Query language (CA/ES/EN) |
| `meta.intent` | Query intent (Discover/Quantify) |
| `meta.resolvability` | How well the query maps to the schema |

## Dataset Statistics

### Language Distribution

| Language | Training | Test |
|----------|----------|------|
| Catalan (CA) | ~33% | ~33% |
| Spanish (ES) | ~33% | ~21% |
| English (EN) | ~33% | ~46% |

### Intent Distribution

| Intent | Count | Percentage |
|--------|-------|------------|
| Discover | 600 | 88.0% |
| Quantify | 82 | 12.0% |

### Resolvability Distribution

| Type | Count | Percentage | Description |
|------|-------|------------|-------------|
| Direct | 529 | 77.6% | Fully mappable to schema |
| Adapted | 15 | 2.2% | Requires interpretation |
| Partial | 138 | 20.2% | Cannot fully express (ranking, aggregation) |

### Component Distribution

| Component | Frequency |
|-----------|-----------|
| Thematic content | 92.8% |
| Organization type | 39.9% |
| Organization location | 17.7% |
| Programme (funding) | 17.6% |
| Time expressions | 10.7% |
| Project location | 10.4% |
| Year (specific) | 7.8% |
| Organization name | 7.3% |

## Examples

### Example 1: Catalan Query (Discover)

**Original query:** `"Projectes on la UPC és coordinadora en l'àmbit de la ciberseguretat"`

```json
{
  "doc_type": "projects",
  "filters": {
    "programme": null,
    "funding_level": null,
    "year": null,
    "location": null,
    "location_level": null
  },
  "organisations": [
    {
      "type": "university",
      "name": "UPC",
      "location": null,
      "location_level": null
    }
  ],
  "semantic_query": "ciberseguretat",
  "query_rewrite": "Llista de projectes de la UPC sobre ciberseguretat",
  "meta": {
    "id": "TRAIN_488",
    "source": "synthetic",
    "lang": "CA",
    "original_query": "Projectes on la UPC és coordinadora en l'àmbit de la ciberseguretat",
    "intent": "Discover",
    "style": "Concise",
    "components": ["Scope", "Organisation Name", "Content", "Role Qualifier"],
    "resolvability": "Partial",
    "notes": "No es pot filtrar pel rol de 'coordinadora'"
  }
}
```

### Example 2: Spanish Query with Filters

**Original query:** `"proyectos de inteligencia artificial financiados por H2020 desde 2019"`

```json
{
  "doc_type": "projects",
  "filters": {
    "programme": "Horizon 2020",
    "funding_level": null,
    "year": ">=2019",
    "location": null,
    "location_level": null
  },
  "organisations": [],
  "semantic_query": "inteligencia artificial",
  "query_rewrite": "Proyectos sobre inteligencia artificial del programa H2020 desde 2019",
  "meta": {
    "lang": "ES",
    "intent": "Discover",
    "resolvability": "Direct"
  }
}
```

### Example 3: English Quantify Query

**Original query:** `"how many universities participated in quantum computing projects?"`

```json
{
  "doc_type": "projects",
  "filters": {
    "programme": null,
    "funding_level": null,
    "year": null,
    "location": null,
    "location_level": null
  },
  "organisations": [
    {
      "type": "university",
      "name": null,
      "location": null,
      "location_level": null
    }
  ],
  "semantic_query": "quantum computing",
  "query_rewrite": "Count of universities participating in quantum computing projects",
  "meta": {
    "lang": "EN",
    "intent": "Quantify",
    "resolvability": "Partial",
    "notes": "Aggregation (count) cannot be expressed in schema"
  }
}
```

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

dataset = load_dataset("SIRIS-Lab/impuls-query-parsing")

# Access splits
train_data = dataset["train"]
test_data = dataset["test"]

# Example
print(train_data[0]["meta"]["original_query"])
print(train_data[0]["semantic_query"])
```

### For Training with Transformers

```python
from datasets import load_dataset

dataset = load_dataset("SIRIS-Lab/impuls-query-parsing")

def format_for_training(example):
    # Format as instruction-following
    return {
        "instruction": "Convert this query to structured JSON for R&D project search.",
        "input": example["meta"]["original_query"],
        "output": json.dumps(example, ensure_ascii=False, indent=2)
    }

formatted = dataset.map(format_for_training)
```

## Data Collection

### Training Data
The training set was **synthetically generated** using:
- Controlled vocabularies (funding programmes, organization names, locations)
- Thematic keywords extracted from real R&D project titles and abstracts
- Domain-specific templates mirroring realistic user queries
- Balanced language distribution across Catalan, Spanish, and English

### Test Data
The test set contains **real queries from domain experts**:
- Collected from researchers and R&I policy analysts
- Elicited through structured questionnaires asking for typical information needs
- Manually annotated with gold-standard JSON structures

## Intended Use

This dataset is designed for:
- Training query parsing models for R&D project search systems
- Evaluating multilingual NLU capabilities for Catalan, Spanish, and English
- Benchmarking structured output generation from natural language
- Research on semantic parsing for specialized domains

## Limitations

- **Domain-specific**: Focused on R&D project search; may not generalize to other domains
- **Schema constraints**: Some query types (ranking, complex aggregations) cannot be fully represented
- **Synthetic training data**: Training examples are template-generated, which may limit diversity
- **Language balance**: Test set has more English queries than training distribution

## Citation

```bibtex
@misc{impuls-query-parsing-2024,
  author = {SIRIS Academic},
  title = {IMPULS Query Parsing Dataset: Multilingual Queries for R&D Semantic Search},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/SIRIS-Lab/impuls-query-parsing}}
}
```

## Acknowledgments

- **[Barcelona Supercomputing Center (BSC)](https://www.bsc.es/)** - AINA project infrastructure
- **[Generalitat de Catalunya](https://web.gencat.cat/)** - Funding and RIS3-MCAT platform
- **[AINA Project](https://projecteaina.cat/)** - AINA Challenge 2024 framework

## License

Apache 2.0

## Links

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