📄 International Standard for a Linguistic Annotation Framework

Summary

Annotated language resources (corpora with morphosyntactic, syntactic, discourse or alignment layers) tend to use incompatible formats, often shaped by whatever software produced them. This makes it hard to merge, compare and reuse them. The paper outlines the Linguistic Annotation Framework (LAF), which ISO TC37/SC4 working group WG1-1 is developing. The outline comes from a meeting of roughly 20 experts at Pont-à-Mousson in November 2002.

LAF does not impose one annotation scheme. Instead it defines an abstract data model and a machine-oriented pivot ("dump") format based on feature structure graphs. Resource producers map their own formats onto this pivot and document how they did it. The semantics of annotation labels are anchored in a Data Category Registry, formalized with RDF schemas.

The main result is a set of agreed design requirements and an architecture. The key move is separating user-facing annotation formats from the exchange and processing format. This lets legacy resources stay valid while still becoming interoperable. The paper closes with an open call for community participation.

Key points

• Meeting: the framework's principles were set at a workshop of about 20 experts held on November 21–22, 2002, in Pont-à-Mousson, France.
• Work plan: WG1-1 committed to four tasks:
  • surveying existing annotation types and schemes;
  • defining an abstract format;
  • building a mechanism for formally defined reference data categories;
  • providing guidelines and implementation mechanisms for new schemes.
• Requirements: the experts listed expressive adequacy, media independence, semantic adequacy, incrementality (including partial or ambiguous results), uniformity, openness to any linguistic theory, extensibility, human readability, explicitness and consistency.
• Data model: an annotation is treated as a link between an annotation object and points or spans in linearly ordered base data. The base data might be characters, a speech time line or pixels. The pivot model is a feature structure graph able to reference n-dimensional regions and other annotations.
• Starting points and grounding: the XML-based Generic Modeling Tool (GMT) is a candidate starting point for the pivot format, with Annotation Graphs cited as related work. Data categories are referenced by URL in the Data Category Registry and formalized with RDFS/OWL. Example mappings, for instance for XCES, are planned.

Interesting ideas

• I1. Pivot format as "byte code": users annotate in whatever format suits them, and tools convert to and from a hidden machine-only interchange layer. This is a practical way to make heterogeneous datasets interoperable without forcing annotators to change tools. It would suit a multi-project setting where corpora come from different pipelines.
• I2. Conversion duty on the producer: whoever creates a resource converts it to the common format and documents the mapping to the data model. This spreads the integration work and keeps the semantics explicit rather than reverse-engineered later.
• I3. Shared registry instead of a fixed tagset: label meanings live in a central registry with stable URLs. Scheme-specific names are mapped onto reference categories or derived from them. Labels then become comparable across projects without everyone having to adopt one tagset, which is useful for aligning, say, different POS or NER conventions.
• I4. Two ways to mark spans: inline start/end tags give hierarchical grouping, while explicit start/end addressing is needed when the base data is read-only. Keeping both in mind, and allowing reference down to the smallest unit (character, morpheme), supports stand-off layers over untouchable source texts.
• I5. Built-in incrementality: the model is designed from the start to hold partial, underspecified or alternative analyses and to merge and compare them. This fits processing chains where several tools or annotators contribute competing layers.