Blame for README.md

---
license: apache-2.0
pretty_name: "POISS: Point-of-Interest Search Suite"
task_categories:
- text-retrieval
language:
- en
- es
- fr
- pt
- de
- it
- nl
size_categories:
- 100K<n<1M
tags:
- point-of-interest
- poi-search
- geospatial
- reranking
- multilingual
- graded-relevance
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: test
    path: data/test-*.parquet
---

# POISS: A Large-Scale Multilingual Dataset for Point-of-Interest Search

POISS is the first large-scale dataset for Point-of-Interest (POI) search. Given a natural-language query and
the coordinate it was issued from, a system ranks POIs by relevance. Relevance here is
genuinely two-sided: "coffee shops near me" depends on where the user is, while "hotel caribe en
mérida yucatán" names a specific place hundreds of kilometres away. Models have to weigh
**semantics** and **geography** together.

The dataset contains **309,386 queries in 7 languages across 14 locales**, grounded in the
[Overture Maps Places](https://docs.overturemaps.org/guides/places/) corpus of more than 72
million POIs. Each query comes with a **ranked list of candidate POIs (~301) with graded
relevance labels**, so the same data supports both retrieval over the full corpus and reranking
of a candidate pool.

Fine-tuned baselines: [`amazon/poiss-bge-m3-retriever`](https://huggingface.co/amazon/poiss-bge-m3-retriever)
and [`amazon/poiss-bge-m3-reranker`](https://huggingface.co/amazon/poiss-bge-m3-reranker).

## About

The datasets and models in this repository originate from our paper accepted at the **Main Conference** of **The 2026 Conference on Empirical Methods in Natural Language Processing**.

For more details, please check out the paper [POISS: A Large-Scale Multilingual Dataset for Point-of-Interest Search](https://aclanthology.org/2026.emnlp-main.2497/).

✨ **Come visit our poster session in Budapest!** We look forward to meeting you and discussing our work.

## What you can build with it

- **Geo-aware retrieval at scale.** Retrieve from a 72M-POI index using the query text and the
  user's coordinate. Recall@{10,100,1000} and NDCG@{5,20} against graded labels.
- **Reranking with graded relevance.** Up to 20 ranked positives per query let you study
  fine-grained ordering, not just relevant/non-relevant.
- **Multilingual and cross-lingual behaviour.** English, Spanish, French, Portuguese, German,
  Italian and Dutch, each with enough test queries to report per-language numbers. Every row
  carries `language`, `locale` and `query_country`, so the breakdowns are a `group_by` away.
- **Intent analysis.** The `intent_group` column splits queries into name-anchored (`Search`,
  `Detail`) and open-ended (`Recommend`, `Things-to-do`) intents, which behave very differently:
  open-ended queries stay substantially harder for every system we measured.
- **Geography versus semantics.** Every query carries its issuing coordinate and every candidate
  its location, so distance can be modelled explicitly, combined with lexical or dense scores,
  or studied as a feature.
- **Training data for retrievers and rerankers.** The ranked, labeled candidate list per query
  is designed to be sampled from: positives, graded positives, judged negatives and an unjudged
  tail are all identified, so you can pick the positive/negative scheme your setup needs.

## Dataset at a glance

| | train | test |
|---|---|---|
| Queries | 267,368 | 42,018 |
| Candidate query–POI pairs | 80,495,030 | 12,655,117 |
| Candidates per query (avg) | 301.1 | 301.2 |
| Graded relevant POIs per query (avg) | 14.15 | 13.78 |

Across both splits the dataset references 20,960,591 distinct Overture POIs. Queries average
3.26 tokens. These counts are measured on the files in this repository; Table 1 of the paper
reports the candidate pools before de-duplication, so its per-query averages are
marginally higher.

## Row schema

| Field | Type | Description |
|---|---|---|
| `query_id` | string | uuid4 |
| `split` | string | `train` or `test` |
| `query_text` | string | the query |
| `user_geolocation` | struct{`lat`,`lng`: float64} | coordinate the query was issued from |
| `language` | string | ISO 639-1 code, e.g. `it` |
| `locale` | string | language–country locale, e.g. `it_IT`, `es_US` |
| `query_country` | string | ISO 3166-1 alpha-2 country, e.g. `IT` |
| `intent_group` | string | `Search`, `Detail`, `Recommend` or `Things-to-do` |
| `candidates` | list[string] | ordered Overture GERS identifiers |
| `candidate_labels` | list[string] | parallel to `candidates`, one label each |
| `relevant_candidates` | list[string] | the graded-relevant subset (≤ 20), in ranked order |
| `consolidation_removed` | list[string] | POIs identified as cross-source duplicates |
| `label_counts` | struct of 5 int64 | count per label |

### Candidate ordering and labels

`candidate_labels[i]` describes `candidates[i]`, and the labels form contiguous blocks in this
order:

| Order | Label | Meaning | avg/query (train) |
|---|---|---|---|
| 1 | `relevant` | graded positives, ranked | 14.15 |
| 2 | `semantic_positive` | judged relevant, beyond the graded top-20 | 6.45 |
| 3 | `semantic_negative_google` | reference-engine candidate judged non-relevant | 2.34 |
| 4 | `semantic_negative_poi_search_system` | retrieved candidate judged non-relevant | 11.14 |
| 5 | `unjudged_poi_search_system` | retrieve-and-rerank tail | 266.98 |

Because the order is fixed, a rank-based slice of `candidates` is also a label-based one:
`relevant_candidates` is the head of the list, and `label_counts` gives the block boundaries in
constant time.

## Loading

```python
from datasets import load_dataset

poiss = load_dataset("amazon/poiss")
row = poiss["test"][0]
print(row["query_text"], row["user_geolocation"], row["locale"], row["intent_group"])
print(row["candidates"][:5], row["candidate_labels"][:5])

# per-language or per-intent slices
italian = poiss["test"].filter(lambda r: r["language"] == "it")
open_ended = poiss["test"].filter(lambda r: r["intent_group"] == "Recommend")
```

## Recovering POI content

POISS ships Overture identifiers; POI names, addresses, coordinates and categories come from
Overture itself. `scripts/join_overture.py` does the join against a current release:

```bash
python scripts/join_overture.py --split test --release 2026-08-19.0 --out test_pois.parquet
```

Output columns: `poiss_candidate_id`, `overture_id`, `names`, `addresses`, `categories`,
`confidence`, `latitude`, `longitude`, `quality_score`. Requires `pyarrow` and `tqdm`.

POISS was built on Overture release `2026-03-18`, and Overture reassigns some GERS identifiers
when it re-conflates the corpus. `data/overture_id_map.parquet` translates 1,853,120 such
identifiers to their current equivalent, matched through the provider records the two releases
share. With the map applied, 88.5% of candidate identifiers and about 90% of the graded
positives resolve against release `2026-08-19`; the remainder are POIs Overture has since
retired, and the script lists them in its report. The map is refreshed as new Overture releases
appear, and `--id-map` accepts your own.

## Quality score

`data/poi_quality_score.parquet` maps `overture_id` to a `score` in [0, 5]: an LLM-derived
quality signal used during labeling and read by the cross-encoder as its `Score` feature. It is
not an Overture field, so it ships with the dataset.

```python
import pyarrow.parquet as pq
table = pq.read_table("data/poi_quality_score.parquet")
scores = dict(zip(table.column("overture_id").to_pylist(), table.column("score").to_pylist()))
```

The file covers the POIs for which upstream rating information was available — 30.7% of
query–candidate pairs, and 43.7% of the graded positives. Models were trained and evaluated
with the field absent where no score exists.

## Tasks and reference results

**Full-corpus retrieval.** The search space is the entire Overture corpus; systems are scored
against the graded labels. Test split, percentages:

| System | R@10 | R@100 | R@1k | N@5 | N@20 |
|---|---|---|---|---|---|
| Distance only | 0.6 | 2.4 | 10.2 | 0.5 | 0.7 |
| BM25 | 17.8 | 29.9 | 42.8 | 16.0 | 18.0 |
| BM25 + distance | 18.4 | 31.2 | 42.8 | 17.1 | 19.1 |
| BGE-M3 zero-shot | 15.2 | 26.4 | 40.0 | 13.7 | 15.3 |
| **BGE-M3 fine-tuned** | **45.2** | **83.2** | **95.6** | 58.3 | 60.3 |
| Qwen3-Embedding-0.6B zero-shot | 22.2 | 35.3 | 49.7 | 22.6 | 24.1 |
| Qwen3-Embedding-0.6B fine-tuned | 45.0 | 82.5 | 94.7 | **58.7** | **60.7** |

Fine-tuned per-language R@100 spans 79.8 (NL) to 86.1 (DE); per-intent R@100 is 89.1 on
`Search` against 77.3 on `Recommend`. Both breakdowns can be reproduced from the `language` and
`intent_group` columns.

The test split is 19.5% US English, 15.0% UK English, 14.4% Canadian English, 8.9% Brazilian
Portuguese, 8.6% French, 8.2% Spanish (Spain), 7.6% Mexican Spanish, 7.1% Italian, 6.9% German,
with the remainder spread over Dutch, Irish English, Canadian French, Austrian German and US
Spanish. By intent it is 49.8% `Search`, 47.8% `Recommend`, 1.9% `Things-to-do`, 0.5% `Detail`.

**Candidate reranking.** A system reorders a per-query pool. The numbers below use the pool from
the reference setup — the fine-tuned retriever's top-1000 over the full corpus — which can be
rebuilt with the released retriever. Reranking the `candidates` column is also a valid task, in
an easier setting where every graded positive is present by construction.

| System | P@5 | P@20 | MRR | N@5 | N@20 |
|---|---|---|---|---|---|
| BGE-M3 retriever (bi-encoder order) | 57.6 | 35.5 | 85.5 | 61.5 | 62.6 |
| BGE-Reranker-v2-m3 zero-shot | 24.6 | 14.3 | 50.0 | 32.1 | 34.8 |
| **BGE-Reranker-v2-m3 fine-tuned** | **63.0** | **38.3** | **89.7** | **66.8** | **66.6** |
| Qwen3-Reranker-0.6B zero-shot | 38.3 | 21.3 | 68.3 | 43.8 | 44.6 |
| Qwen3-Reranker-0.6B fine-tuned | 61.8 | 37.4 | 88.8 | 65.7 | 65.7 |

For reference, the reranker baseline was trained with positives = `relevant` and negatives drawn
from `semantic_negative_poi_search_system` + `unjudged_poi_search_system`, padded to 100 per
query. Other sampling schemes, or training on a rank-based subset of the full list, are equally
available.

## License and attribution

POISS is released under the **Apache License 2.0**, covering the queries, the relevance labels
and their ordering, the identifier map, the quality scores and the scripts.
As a condition of use, you must cite our associated paper whenever you use this dataset to train a model, run evaluations, or publish research.

POI content is not redistributed here. When you join against Overture you are bound by
Overture's terms: the places theme is published under CDLA-Permissive-2.0 and Apache-2.0
depending on the source provider and contains no OpenStreetMap data. Attribute Overture Maps
Foundation and the per-source licenses accordingly; each Overture record carries its own
`sources[].license`. See [Overture licensing](https://docs.overturemaps.org/guides/places/).

The released baselines carry the license of the model they were fine-tuned from:
[`poiss-bge-m3-retriever`](https://huggingface.co/amazon/poiss-bge-m3-retriever) is MIT, from
`BAAI/bge-m3`; [`poiss-bge-m3-reranker`](https://huggingface.co/amazon/poiss-bge-m3-reranker) is
Apache-2.0, from `BAAI/bge-reranker-v2-m3`.

## Citation

```bibtex
@inproceedings{maritan-etal-2026-poiss,
    title = "{POISS}: A Large-Scale Multilingual Dataset for Point-of-Interest Search",
    author = "Maritan, Nicola  and
      Moschitti, Alessandro  and
      Borazio, Federico  and
      Zhou, Xiaokun  and
      Bai, Zhengwei",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    note = "To appear",
}
```