CLEF HIPE-2026: Evaluating Accurate and Efficient Person-Place Relation Extraction from Multilingual Historical Texts

CLEF HIPE-2026 is an evaluation lab dedicated to advancing the extraction of person–place relations from noisy, multilingual historical texts, building on the prior HIPE-2020 and HIPE-2022 campaigns which focused on named entity recognition and linking in historical corpora. The task shifts focus toward semantic relation extraction by identifying associations between individuals and locations across multiple languages and time periods. Systems are required to classify two types of temporal relations: at, which answers whether a person has ever been at a place prior to the document's publication date, and isAt, which determines if the person was located at the place around the publication time. These classifications demand reasoning over temporal and geographical cues present in the text.

The at relation is categorized into three labels: true (explicit evidence), probable (inferred from context), and false (no or contradictory evidence), while isAt is a binary classification (+/–) indicating presence near the publication date. Conceptually, isAt=+ presupposes that the at relation is at least probable, although predictions with at=false and isAt=+ are permitted despite being epistemically inconsistent. This distinction supports abductive interpretation, where implicit or indirect cues—such as institutional roles or narrative coherence—are used to infer plausible relations even without explicit statements.

HIPE-2026 introduces a three-fold evaluation profile assessing accuracy, computational efficiency, and domain generalization. The efficiency profile rewards lightweight and scalable models considering compute cost and model size, while the generalization profile includes an unseen dataset from a different domain to evaluate robustness beyond historical newspapers. Training and test data are drawn from historical newspapers in English, German, French, and Luxembourgish, with entity pairs pre-identified for evaluation.

The task is designed to be approachable by both large language models (LLMs) and traditional classification systems, encouraging participation from diverse AI research communities. A pilot study showed that GPT-4o achieved up to 0.8 agreement with human annotators on the at relation, though performance on isAt was more variable (0.2–0.7), highlighting the challenge of temporal grounding. The lab aims to support downstream applications in knowledge graph construction, historical biography reconstruction, and spatial analysis in digital humanities.

Registration for the lab remains open until April 23, 2026, with the evaluation period scheduled for May 5–7, 2026, and results to be discussed at the CLEF conference in Jena, Germany, from September 21–24, 2026. All datasets will be publicly released to promote transparency and continued research in historical text processing

CLEF HIPE-2026: Evaluating Accurate and Efficient Person-Place Relation Extraction from Multilingual Historical Texts

HIPE-2026 is a proposed shared task within the Conference and Labs of the Evaluation Forum (CLEF) that focuses on the complex challenge of extracting semantic relations between named entities in multilingual historical documents. Specifically, this lab targets the identification of Person-Place relations, distinguishing between two primary relation types: the physical presence of a person at a location (at) and the affiliation or residence of a person to a location (isAt). The task requires participants to process noisy, digitized historical texts across multiple languages, addressing the specific difficulties posed by Optical Character Recognition (OCR) errors, archaic vocabulary, and inconsistent orthography typical of diachronic corpora.

The key contribution of this material is the establishment of a standardized benchmark dataset and evaluation framework designed to advance Information Extraction (IE) capabilities in the digital humanities domain. Unlike previous iterations that may have focused primarily on Named Entity Recognition (NER), HIPE-2026 emphasizes the downstream step of relation extraction, requiring systems to understand not just who and where, but how they are connected. Furthermore, the lab explicitly incorporates an evaluation of efficiency alongside accuracy, encouraging the development of lightweight models capable of processing large-scale archival data without prohibitive computational costs.

This research matters significantly because unlocking structured data from historical texts is crucial for prosopographical studies, historical network analysis, and the creation of rich knowledge graphs for cultural heritage. By tackling the dual challenges of linguistic noise and semantic complexity, HIPE-2026 aims to bridge the gap between state-of-the-art NLP techniques—which are often trained on clean, modern data—and the messy reality of historical archives. Success in this domain will facilitate more sophisticated querying and visualization of historical migration, social structures, and biographical data across linguistic boundaries.

# Summary: CLEF HIPE-2026 – Evaluating Person-Place Relation Extraction from Multilingual Historical Texts

The CLEF HIPE-2026 lab introduces a benchmark for evaluating person-place relation extraction in noisy, multilingual historical texts, with a focus on two key relation types: "at" (indicating a person's presence at a location) and "isAt" (indicating a location's existence or relevance to a person). This task is particularly challenging due to the linguistic variability, historical terminology, and structural noise inherent in historical documents, which complicates standard NLP pipelines. The lab provides a shared task framework, including annotated datasets in multiple languages, evaluation metrics, and a competitive leaderboard to push advancements in domain-specific information extraction from historical texts.

The key contributions of HIPE-2026 include: 1. Multilingual Historical Dataset – A curated collection of annotated historical texts covering multiple languages, enabling cross-lingual comparisons and robustness testing. 2. Noisy Text Handling – Techniques and benchmarks for addressing OCR errors, archaic language, and ambiguous references, which are pervasive in historical documents. 3. Relation Extraction Challenges – A focus on temporal and spatial disambiguation, where relations may be implicit or context-dependent, requiring advanced models (e.g., transformer-based systems with historical language adaptations). 4. Baseline Models & Evaluation – Baseline systems (e.g., fine-tuned BERT, relation extraction models) and metrics (precision, recall, F1) to standardize performance assessment.

This work is significant because historical person-place relations are critical for digital humanities, genealogical research, and cultural heritage preservation. By addressing the unique challenges of historical texts—such as linguistic drift, OCR artifacts, and sparse training data—HIPE-2026 accelerates progress in historical NLP, enabling more accurate extraction of biographical and geospatial information from vast, underutilized archives. The lab also fosters collaboration between NLP researchers and digital humanists, bridging gaps between computational techniques and humanities-driven applications.

Source: [arXiv:2602.17663](https://arxiv.org/abs/2602.17663)