README.md

---
language:
  - en
license: cc-by-sa-4.0
tags:
  - retrieval
  - text-retrieval
  - beir
  - entity-retrieval
  - dbpedia
  - benchmark
  - open-domain
pretty_name: BEIR DBPedia-Entity (retrieval)
size_categories: "100<n<1K"
task_categories:
  - text-retrieval
---

# BEIR DBPedia-Entity (`orgrctera/beir_dbpedia_entity`)

## Overview

**DBpedia-Entity v2** is a standard test collection for **entity-oriented search** over the [DBpedia](https://www.dbpedia.org/) knowledge base: given a short **information need** expressed in natural language, systems must retrieve **DBpedia entities** (articles) that satisfy that need. The collection unifies queries from several benchmarks (e.g. SemSearch, INEX, QALD entity search tasks) with **graded relevance judgments** collected under consistent guidelines.

**BEIR** (*Benchmarking IR*) repackages **DBpedia-Entity** as one of its heterogeneous **zero-shot retrieval** tasks. In the BEIR setting, each query is evaluated against a fixed **corpus** of DBpedia passages (title + abstract text per entity); models are scored with standard IR metrics after ranking corpus documents using the official **qrels**.

This Hub release (`orgrctera/beir_dbpedia_entity`) exposes the BEIR-style split as a **tabular retrieval dataset**: one row per **query**, with **gold relevant entity IDs** (and relevance grades) in `expected_output`. The underlying **467** queries from DBpedia-Entity v2 appear here as **`dev`** (67) and **`test`** (400) partitions, matching the BEIR train/dev/test convention for this benchmark.

## Task: retrieval (DBPedia-Entity)

- **Task type:** **Retrieval** — ad hoc **entity retrieval** over the DBpedia-Entity **corpus** distributed with BEIR (dense / sparse / hybrid retrievers, rerankers, or full RAG stacks).
- **Input:** A natural-language **query** (`input`) describing the sought entities or information need.
- **Supervision:** `expected_output` is a JSON string of `{ "id": "<corpus-doc-id>", "score": <grade> }` entries. Grades follow the **DBpedia-Entity v2** qrels (typically **1** and **2** for different relevance levels; treat as graded labels when computing nDCG, or binarize for binary Recall depending on your protocol).

Full benchmark evaluation requires indexing the **BEIR DBpedia-Entity corpus** (document text keyed by `_id`) and ranking with the same **query / qrels** splits. See [UKPLab/beir](https://github.com/UKPLab/beir) and [ir-datasets: `beir/dbpedia-entity`](https://ir-datasets.com/beir.html#beir/dbpedia-entity) for the canonical corpus + evaluation layout.

## Background

### DBpedia-Entity v2

[Hasibi et al. (SIGIR 2017)](https://doi.org/10.1145/3077136.3080751) introduced **DBpedia-Entity v2**, an updated test collection built on a **2015-10** DBpedia snapshot. Compared to earlier versions, v2 uses a **unified candidate pool** across retrieval models and **uniform crowdsourced assessments**, yielding graded judgments for a large set of **query–entity pairs** and making the collection a widely used benchmark for **entity search** and related semantic retrieval research.

- **Resources:** [iai-group/DBpedia-Entity on GitHub](https://github.com/iai-group/DBpedia-Entity) · [Dataset overview](https://iai-group.github.io/DBpedia-Entity/collections)

### BEIR

[Thakur et al. (2021)](https://arxiv.org/abs/2104.08663) curated **BEIR**: **18** public datasets spanning **nine** retrieval task families (including **entity retrieval**). DBpedia-Entity is the BEIR benchmark for **retrieving structured encyclopedic entities** from a large KB-derived corpus—complementary to news, biomedical, or QA-style collections in the same suite.

- **Code / data hub:** [UKPLab/beir](https://github.com/UKPLab/beir) · [BeIR/dbpedia-entity on Hugging Face](https://huggingface.co/datasets/BeIR/dbpedia-entity)

### Relation to ir-datasets / `beir/dbpedia-entity`

The same corpus, queries, and qrels are documented in [ir-datasets](https://ir-datasets.com/beir.html#beir/dbpedia-entity) (`beir/dbpedia-entity`, plus `dev` and `test` topic subsets). This dataset is a **row-oriented export** aligned with other CTERA benchmark releases: one row per query with string `expected_output` for tooling compatibility.

## Data fields

| Column | Type | Description |
|--------|------|-------------|
| `id` | `string` | Unique row identifier (UUID). |
| `input` | `string` | The **query** text (information need). |
| `expected_output` | `string` | JSON array of `{ "id": "<doc_id>", "score": <int> }` for judged **relevant** corpus entities (qrels-style). |
| `metadata.query_id` | `string` | Source query id (e.g. SemSearch / INEX style ids such as `SemSearch_ES-81`, `INEX_XER-147`). |
| `metadata.split` | `string` | `dev` or `test`. |

## Splits

| Split | Queries |
|-------|--------:|
| `dev` | 67 |
| `test` | 400 |
| **Total** | **467** |

## Examples

Illustrative rows (document lists truncated for readability; full qrels may contain many entities per query).

**Example 1 — `dev`**

- **`input`:** `south dakota state university`
- **`expected_output`** (excerpt):
```json
[
  {"id": "<dbpedia:2011–12_South_Dakota_State_Jackrabbits_men's_basketball_team>", "score": 1},
  {"id": "<dbpedia:Dakota_State_University>", "score": 2},
  {"id": "<dbpedia:South_Dakota_Board_of_Regents>", "score": 1}
]
```
- **`metadata.query_id`:** `SemSearch_ES-81` · **`metadata.split`:** `dev`

**Example 2 — `test`**

- **`input`:** `Chemical elements that are named after people`
- **`expected_output`** (excerpt):
```json
[
  {"id": "<dbpedia:Bohrium>", "score": 2},
  {"id": "<dbpedia:Curium>", "score": 2},
  {"id": "<dbpedia:Einsteinium>", "score": 2},
  {"id": "<dbpedia:Naming_of_elements>", "score": 1}
]
```
- **`metadata.query_id`:** `INEX_XER-147` · **`metadata.split`:** `test`

## References

### DBpedia-Entity v2 (original test collection)

> **Abstract (SIGIR 2017):** The paper presents **DBpedia-Entity v2**, a test collection for entity search built on DBpedia, with a unified pooling setup and crowdsourced relevance judgments suitable for comparing entity retrieval methods on a large query set.

- Faegheh Hasibi, Fedor Nikolaev, Chenyan Xiong, Krisztian Balog, Svein Erik Bratsberg, Alexander Kotov, James Callan. **“DBpedia-Entity v2: A Test Collection for Entity Search.”** *Proceedings of the 40th ACM SIGIR* (2017). [ACM DL](https://doi.org/10.1145/3077136.3080751) · [Author PDF](http://hasibi.com/files/sigir2017-dbpedia_entity.pdf)

### BEIR (benchmark including DBpedia-Entity)

> **Abstract (arXiv:2104.08663):** *We introduce Benchmarking-IR (BEIR), a robust and heterogeneous evaluation benchmark for information retrieval. We leverage a careful selection of 18 publicly available datasets from diverse text retrieval tasks and domains…*

- Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, Iryna Gurevych. **“BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models.”** [arXiv:2104.08663](https://arxiv.org/abs/2104.08663)

## Citation

If you use DBpedia-Entity, cite:

```bibtex
@inproceedings{Hasibi2017DBpediaEntityV2,
  author    = {Hasibi, Faegheh and Nikolaev, Fedor and Xiong, Chenyan and
               Balog, Krisztian and Bratsberg, Svein Erik and Kotov, Alexander and
               Callan, James},
  title     = {{DBpedia-Entity v2}: A Test Collection for Entity Search},
  booktitle = {Proceedings of the 40th International ACM SIGIR Conference on
               Research and Development in Information Retrieval},
  year      = {2017},
  url       = {https://doi.org/10.1145/3077136.3080751},
  doi       = {10.1145/3077136.3080751}
}
```

If you use the BEIR benchmark formulation, cite:

```bibtex
@article{Thakur2021Beir,
  title   = {{BEIR}: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
  author  = {Thakur, Nandan and Reimers, Nils and R{\"u}ckl{\'e}, Andreas and
             Srivastava, Abhishek and Gurevych, Iryna},
  journal = {arXiv preprint arXiv:2104.08663},
  year    = {2021},
  url     = {https://arxiv.org/abs/2104.08663}
}
```

## Provenance

Exported for retrieval evaluation with **DBpedia-Entity** as the BEIR sub-benchmark `dbpedia-entity`. Corpus passages are **not** duplicated per row; join `expected_output` ids to the BEIR **DBpedia-Entity corpus** for full title and abstract text when building an index.