Tasks & Data
On This Page
- Task Overview
- Relation Types
- Input and Output Format
- Evaluation Profiles
- Datasets
- Baseline and Starter Code
Task Overview
HIPE-2026 is a shared task on person–place relation extraction from multilingual historical texts. The goal is to assess whether a relation holds between a person and a place mentioned in a document — and to classify that relation with respect to its temporal scope.
Participants are asked to build systems that, given a historical document and a set of
entities, will classify all possible (person, place) pairs into one of three
evidence-based labels for each of two relation types. The task is designed to be
tackled by generative AI systems/LLMs as well as more traditional classification approaches.
Relation Types
Two relation types are to be evaluated independently:
at– Did the person ever reside in or visit the place prior to the document’s publication?isAt– Is the person located at the place in the immediate temporal context of the document?
This design supports different downstream goals — from spatial biographies to historical event contextualization.
Input and Output Format
Each document is provided as JSON with:
- Full article text (OCR with possible errors)
- A list of person entities (de-duplicated by surface form or linked ID)
- A list of place entities
- Metadata (e.g., document language, publication date)
Participants must return a classification for each possible (person, place, relation) triple using:
TRUE: strong textual evidence supports the relationPROBABLE: plausible inference can be made from contextFALSE: no evidence or explicitly contradicted
Realistic Example from Historical Data
This example illustrates a real instance of the HIPE-2026 task using an article from the Gazette de Lausanne dated 1928-05-06. It involves multiple persons and places, various temporal scopes, and differing levels of textual evidence.
Article Context
| Original French OCR | Automatic English Translation |
|---|---|
|
Pour les enfants sinistrés de Bulgarie et de Grèce, Mgr. Stéphane, archevêque de Sofia,
vient d’adresser à l’Union internationale de secours aux enfants une dépêche, où, après avoir
rendu hommage à cette institution, il s’exprime comme suit : La solidarité humaine se manifeste le plus
sensiblement dans les heures critiques. Le peuple bulgare est sincèrement reconnaissant envers tous ceux
qui, dans son épreuve actuelle, lui ont témoigné sympathie et aide. Dieu bénisse chaque effort qui
soulagera la souffrance, surtout celle des malheureux petits.
D’autre part, l’U.I.S.E. reçoit de sa déléguée la nouvelle qu’elle a pu assurer une distribution quotidienne de pain à 3400 enfants dans les environs de Philippopoli et, dans la ville même, de pain et de thé à 2500 enfants. En outre, elle a fourni des couvertures à l’hôpital de dix baraques ouvert près de Philippopoli par le chef de la garnison de cette ville, le général Koutzeroff. D’Athènes, le Dr Doxiadès, ancien ministre, président de la Ligue patriotique d’assistance aux enfants, télégraphie à l’U.I.S.E. : Envisageant le danger auquel sont exposés les enfants de la population de Corinthe, la Ligue patriotique fait appel aux généreux sentiments de l’Union pour aider et faciliter la bonne marche de l’œuvre de secours entreprise. |
For the children affected by disasters in Bulgaria and Greece, Mgr. Stéphane, Archbishop of Sofia,
wishes to address the International Union for Child Relief with a dispatch, in which, after paying tribute
to this institution, he expresses himself as follows: Human solidarity is most significantly manifested
in critical hours. The Bulgarian people are sincerely grateful to all those who, in its current ordeal,
have shown sympathy and assistance. God bless every effort that alleviates suffering, especially that
of the unfortunate little ones.
Furthermore, the I.U.C.R. receives news from its delegate that it has been able to ensure a daily distribution of bread to 3,400 children in the vicinity of Philippopolis and, in the city itself, bread and tea to 2,500 children. In addition, it has provided blankets to the ten-barrack hospital opened near Philippopolis by the commander of the garrison of that city, General Koutzeroff. From Athens, Dr. Doxiadès, former minister and president of the Patriotic League for Child Assistance, telegraphs to the I.U.C.R.: Considering the danger to which the children of the population of Corinth are exposed, the Patriotic League appeals to the generous sentiments of the Union to help and facilitate the smooth progress of the relief work undertaken. |
Annotated Relation Table
| Person | Place | at |
isAt |
|---|---|---|---|
| Mgr. Stéphane, archevêque de Sofia | Bulgarie | TRUE | FALSE |
| Mgr. Stéphane, archevêque de Sofia | Grèce | FALSE | FALSE |
| Mgr. Stéphane, archevêque de Sofia | Philippopoli | FALSE | FALSE |
| Mgr. Stéphane, archevêque de Sofia | Athènes | FALSE | FALSE |
| Mgr. Stéphane, archevêque de Sofia | Corinthe | FALSE | FALSE |
| Chef de la garnison de cette ville, le général Koutzeroff | Bulgarie | TRUE | TRUE |
| Chef de la garnison de cette ville, le général Koutzeroff | Grèce | FALSE | FALSE |
| Chef de la garnison de cette ville, le général Koutzeroff | Philippopoli | TRUE | TRUE |
| Chef de la garnison de cette ville, le général Koutzeroff | Athènes | FALSE | FALSE |
| Chef de la garnison de cette ville, le général Koutzeroff | Corinthe | FALSE | FALSE |
| Dr. Doxiadès, ancien ministre, président de la Ligue patriotique… | Bulgarie | FALSE | FALSE |
| Dr. Doxiadès, ancien ministre, président de la Ligue patriotique… | Grèce | TRUE | TRUE |
| Dr. Doxiadès, ancien ministre, président de la Ligue patriotique… | Philippopoli | FALSE | FALSE |
| Dr. Doxiadès, ancien ministre, président de la Ligue patriotique… | Athènes | TRUE | TRUE |
| Dr. Doxiadès, ancien ministre, président de la Ligue patriotique… | Corinthe | FALSE | FALSE |
Download Example Data
Please download the Excel file below for seven more examples and specifications on the annotation scheme.
Evaluation Profiles
HIPE-2026 reports three evaluation profiles:
-
Accuracy Profile: ranking on the multilingual
impressonewspaper test data, based primarily on macro-averaged Recall (balanced accuracy). -
Generalization Profile: evaluation on the out-of-domain
surprisetest data. -
Accuracy-Efficiency Profile: ranking that combines prediction quality with model footprint metadata such as parameter count and model size.
Please find more details in the Participation Guidelines.
Datasets
The task uses two dataset families. The original participation guidelines describe these as Test Set A and Test Set B; the website and generated reports use the dataset names impresso and surprise.
Development and impresso Data
- Historical newspaper material derived mostly from the HIPE-2022 datasets
- Public resources include French, German, English, and Luxembourgish material
- The official
impressotest ranking uses German, English, and French - Contains manually validated relation labels for pre-annotated person/place entities
- Includes metadata for temporal reasoning
surprise Data
- French literary corpus from the 16th–18th century
- Annotated for NER and enriched with relation labels
- Restricted to the
atrelation type for evaluation - Designed to evaluate domain generalization
All data is distributed under CC-BY 4.0, with public resources available through GitHub and archived releases.
Baseline and Starter Code
The public resources include:
- Input/output templates
- Scoring scripts
- A baseline system based on LLM prompting
- Evaluation reports, diagnostics, and submitted runs in the evaluation repository
Questions?
Please post to the mailing list.