DETECH 2027

Challenge

The DEfinition and Term Extraction CHallenge (DETECH) focuses on the automatic extraction of domain-specific terms and the generation of natural language definitions for specialized concepts. Following its first edition in the biomedical domain (DETECH 2026), DETECH 2027 addresses terminology in the legal domain, with particular attention to emerging legal and legal-technological concepts. The challenge investigates how natural language processing and generative artificial intelligence can support the identification and definition of specialized legal concepts.

DETECH 2027 is organized within the Generative Artificial Intelligence in Legal Translation (GenAI-LT) project, funded by the Croatian Science Foundation (No. IP-2025-02-3140). It will take place as a satellite event of Multilingual Digital Terminology Today (MDTT) 2027, in Bolzano, Italy, on June 30, 2027.

Legal terminology presents a particularly demanding testbed: legal concepts are strongly dependent on their regulatory and institutional context, and newly emerging concepts may not yet have stable or generally accepted definitions.

Rather than assuming that each legal concept has a single established definition, the challenge investigates how automatic systems can identify legally meaningful expressions and generate accurate, informative, and context-sensitive definitions grounded in legal usage.

The challenge is organized around the following tasks:

Task A – Legal Term Extraction: Identify and rank relevant single-word and multi-word legal terms from English texts concerning emerging areas of legal regulation. Systems should distinguish expressions carrying legal, normative, institutional, procedural, or regulatory significance from general vocabulary and ordinary noun phrases.
Task B – Legal Definition Generation: Produce natural language definitions for the legal concepts designated by the extracted terms, using corpus-based evidence or automatic text generation methods. Definitions should identify the concept and, where possible, include the essential characteristics that distinguish it from closely related legal concepts.

The provisional agenda of the challenge is the following:

January 22, 2027: training data release
March 12, 2027: deadline for registering to the challenge.
March 19, 2027: test data release
March 26, 2027: submission of runs
April 9, 2027: results announced
April 16, 2027: submission of reports
April 30, 2027: reviewer feedback
May 14, 2027: camera-ready report submission
June 18, 2027: Workshop registration deadline
June 30, 2027: day of the event

Participation

We welcome participation from academic, research, and industry teams working in NLP, terminology, lexicography, legal informatics, or computational law.

In particular:

Please register your team through (TBA link)
Up to five runs per subtask per team are allowed.
External resources (e.g., pre-trained models, lexicons, ontologies) are permitted but must be clearly documented.
Manual runs are accepted but will not be ranked.
Submission details will be available soon.

Evaluation

Each run will be evaluated according to the following measures:

Task A – Legal Term Extraction: Micro-F1, measuring how consistently a system detects every occurrence of a legal term across the corpus, and Type-F1, measuring how well the system identifies the distinct legal terminology represented in the corpus. Precision at k may also be used for ranked lists of term candidates.
Task B – Legal Definition Generation: BLEU and BERTScore will be used for automatic comparison with reference definitions, complemented by manual qualitative assessment of the informativeness, legal-semantic accuracy, terminological adequacy, and linguistic quality of the generated definitions.

Clarification about Task A evaluation

The evaluation treats term strings as the same type if they only differ by case (and other purely formatting-level differences such as leading/trailing whitespace). However, we do not apply lemmatization/stemming, so inflectional variants like singular vs. plural are considered different types.

Examples

data controller vs Data Controller are the same type (case-insensitive)
regulation vs regulations are different types (singular/plural are not collapsed)

For Task B, automatic metrics will be complemented by expert assessment, since more than one valid formulation may express the same legal concept. The manual evaluation will consider legal-semantic accuracy, coverage of essential characteristics, terminological adequacy, informativeness, clarity, and linguistic quality.

Report submissions

Participants must submit a technical report describing their approach, experiments, and results.
Reports must be in the CEUR-ART single-column style.

Templates:

Submission:

Papers do note have to be anonymized
The link to submit the report will be announced here

Reports will undergo a peer review and accepted papers will be published on the online open-access platform CEUR-WS. Proceedings published with CEUR-WS are indexed in Scopus.

It is mandatory that at least one author of each accepted paper registers for the challenge to present the work and have the paper published in the proceedings.

DETECH: DEfinition and Term Extraction CHallenge

Evaluation Task @ MDTT 2027

30 June 2027, Bolzano, Italy

Challenge

Participation

Evaluation

Clarification about Task A evaluation

Examples

Report submissions

Important dates

General Chairs

Conference Venue

CONTACT US