---
language:
- en
task_categories:
- text-retrieval
size_categories:
- 10K<n<100K
dataset_info:
- config_name: documents
  features:
  - name: chunk_id
    dtype: string
  - name: chunk
    dtype: string
  - name: source_url
    dtype: string
  - name: title
    dtype: string
  - name: chunk_idx
    dtype: int64
  - name: chunk_start_char
    dtype: int64
  - name: chunk_end_char
    dtype: int64
  splits:
  - name: test
    num_bytes: 137244943
    num_examples: 70140
  download_size: 78083409
  dataset_size: 137244943
- config_name: queries
  features:
  - name: original_query
    dtype: string
  - name: query
    dtype: string
  - name: answer
    list: string
  - name: score
    list: int64
  - name: source_url
    list: string
  - name: frag_start_char
    list: int64
  - name: frag_end_char
    list: int64
  - name: n_gold
    dtype: int64
  splits:
  - name: test
    num_bytes: 211299459
    num_examples: 10000
  download_size: 95617096
  dataset_size: 211299459
configs:
- config_name: documents
  data_files:
  - split: test
    path: documents/test-*
- config_name: queries
  data_files:
  - split: test
    path: queries/test-*
---

# ZeShEL — Entity-Linking Chunk-level Retrieval Eval

Contextualized chunk-level (query2chunk) retrieval eval for **entity linking**, built from
the **ZeShEL** (Zero-Shot Entity Linking) dataset
([Logeswaran et al., ACL 2019](https://arxiv.org/abs/1906.07348); via
[`naist-nlp/zeshel`](https://huggingface.co/datasets/naist-nlp/zeshel)).

Reformatted into the schema used across the
[Chunk-level Retrieval Eval](https://huggingface.co/collections/bowang0911/chunk-level-retrieval-eval)
collection.

## Task

Entity linking as retrieval: given a **mention in context**, retrieve the correct
**entity's description** from a corpus of entity descriptions. Unlike the QA benchmarks
in this collection, the query is a mention (marked inline with `[START_ENT] … [END_ENT]`)
and the gold "chunk" is the linked entity's description.

Because entity linking's target is an *entity* (represented by a whole description), each
entity description is kept as **one chunk** — so every mention resolves to **exactly one
gold chunk**. This is the natural EL unit and avoids arbitrary sub-entity chunking.

## Scope

Restricted to ZeShEL's **held-out test domains** (the standard zero-shot eval): `star_trek`,
`forgotten_realms`, `lego`, `yugioh`. Source: Wikia/Fandom.

| | count |
| --- | --- |
| entity descriptions (corpus) | 70,140 |
| query mentions (test) | 10,000 |
| test domains | 4 |

## Configs

### `documents` (one chunk per entity)
| field | type | notes |
| --- | --- | --- |
| `chunk_id` | string | ZeShEL entity id |
| `chunk` | string | entity description text |
| `source_url` | string | entity id (identity of the entity) |
| `title` | string | entity name |
| `chunk_idx` | int64 | always 0 (one chunk per entity) |
| `chunk_start_char` / `chunk_end_char` | int64 | `0` .. `len(description)` |

### `queries`
| field | type | notes |
| --- | --- | --- |
| `original_query` | string | raw mention context (unmarked) |
| `query` | string | context with the mention wrapped in `[START_ENT] … [END_ENT]` |
| `answer` | list[string] | gold entity name(s) |
| `score` | list[int64] | `1` per gold |
| `source_url` | list[string] | gold entity id(s) |
| `frag_start_char` / `frag_end_char` | list[int64] | span of the gold entity description (`0 .. len`) |
| `n_gold` | int64 | number of gold entities (almost always 1) |

## Gold matching

A chunk is gold if it shares `source_url` (entity id) with a gold fragment and overlaps its
`[frag_start_char, frag_end_char)`. Since each entity is one chunk and the gold fragment is
the whole description, every mention maps to **exactly one** gold chunk (verified:
10,000 / 10,000).

## Provenance

Built from `naist-nlp/zeshel` `dictionary` (entity descriptions → corpus, filtered to the 4
test domains) and `data/test` (mentions → queries). Records containing multiple mention
spans are expanded to one query per span (10,000 total, matching ZeShEL's canonical test
set). Mentions are marked using the dataset's char offsets.

## Citation

```bibtex
@inproceedings{logeswaran2019zeshel,
    title  = "Zero-Shot Entity Linking by Reading Entity Descriptions",
    author = "Logeswaran, Lajanugen and Chang, Ming-Wei and Lee, Kenton and
              Toutanova, Kristina and Devlin, Jacob and Lee, Honglak",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for
                 Computational Linguistics (ACL)",
    year   = "2019",
    url    = "https://arxiv.org/abs/1906.07348"
}
```