AlvisNLP ML data model - Bibliome/alvisnlp GitHub Wiki

The data structure contains corpus contents and annotation. The data structure is passed from a module to the next one. Each module instance can access (read and write) it through a shared object.

The following figure presents an UML-like specification of the AlvisNLP/ML data structure.

• Corpus: a Corpus object represents a collection of documents. In an AlvisNLP/ML run, the corpus is a unique object passed from module to module. A Corpus object has features and documents.

• Document: a Document object represents a single document. Each document has an identifier which is unique in the corpus. A Document object has features and sections.

• Section: a Section object contains a piece of the document's text contents. Each section has a name, a contents, features, layers, and relations.

• Layer: a Layer object is an annotation container. A Layer object has a name unique in the section.

• Annotation: an Annotation object represents a span of text created by a module. Each annotation is included in at least one layer. An Annotation object has a start and end which represent the coordinates of the annotation in the section's contents, and features.

• Relation: a Relation object is a tuple container. A Relation object has a name unique in the section and features.

• Tuple: a Tuple object represents a relation between several elements in the data structure. A Tuple object has several arguments, each argument is an element (Corpus, Document, Section, Relation, but most often Annotation or Tuple) accessible through a role name. A Tuple object also has features.

• Features are key-value pairs that contain information on an element type, tag or property. Feature keys are not unique in an element, though when accessing a feature key, the last value is returned.