Graphs definition

The graphs we consider in Grew are defined as usually in mathematics by two sets:

• A set N of nodes
• A set E of edges

Nodes

A node is described by a an identifier (needed to refer to nodes in edges definitions) and a feature structure: a finite list of pairs (feature_name, feature_value).

In many linguistic structures, the notion of word order plays a crucial role. To take this into account, in a Grew graph, nodes are split in two disjoint subsets:

• totally ordered nodes (in general the words of some sentence)
• non ordered nodes for other layers of information encoding (examples: constituent nodes in phrase structure, nodes in AMR graphs, additional nodes encoding MWE in PARSEME graphs…)

In the node creation command add_node, the user can choose to add an unordered node or to place the new node before or after a existing one.

Edges

An edge is described by two nodes (called the source and the target of the edge) and by an edge label.

Before version 1.2, edge labels were atomic and did not have an internal structure. This was not very convenient to deal with complex edges:

• in UD, the label aux:pass is a subtype of the label aux
• in SUD, the label compl:obl@agent contains both a subtype obl and a deep feature agent (see TLT 2019)
• in Deep-sequoia, the edge suj:obj means that the final function is suj and the canonical function is obj

In all these cases, with atomic edge labels, it is not possible to deal with one part of the label independently. Since version 1.2, the implementation of edge labels has changed to tackle this problem. Edge labels are now encoded as feature structures.

In Grew graphs, an edge label is internally stored as a flat feature structure or, in other words, a finite set of couples (f_1,v_1) … (f_k,v_k) where all f_i are pairwise different. We will use the traditional notation f=v for these couples.

For backward compatibility and for ease of use in practice, a compact notation may be used for edge labels.

The correspondence between compact notation and feature structure depends on the config parameter. In version 1.4, four predefined configurations are available: ud, sud, sequoia and basic.

The symbol : (used in ud, sud and sequoia) is interpreted as a separator, the left part is given feature value 1 and the right part feature value 2.

The tables below give more examples of correspondances between compact and internal representation.

ud

Relation	Compact notation	Internal representation
Simple relation	obj	1=obj
relation with subtype	aux:pass	1=aux, 2=pass
Enhanced UD relation	E:nsubj	1=nsuj, enhanced=yes

sud

Relation	Compact notation	Internal representation
Simple relation	mod	1=mod
relation with subtype	comp:aux	1=comp, 2=aux
SUD relation with deep feature	compl:obl@agent	1=compl, 2=obl, deep=agent

sequoia

Relation	Compact notation	Internal representation
Simple relation	obj	1=obj
Deep-sequoia (both surf & deep)	suj:obj	1=suj, 2=obj
Deep-sequoia (surf only)	S:suj:obj	1=suj, 2=obj, kind=surf
Deep-sequoia (deep only)	D:suj:obj	1=suj, 2=obj, kind=deep

basic

Relation	Compact notation	Internal representation
Simple relation	obj	rel=obj

Any other feature names (except a few reserved names) can be freely used in edge label representation. But, if the internal representation does not correspond to one described in the tables above, there is not compact representation and the internal representation is used.

Reserved feature names are:

• label: the syntax e.label is a shortcut to make reference to the full feature structure. It can be used for instance to copy the edge label from one edge e to antothe edge f with the command: f.label = e.label.
• length: the syntax e.length is used to refer the distance (natural number) between two ordered nodes. The length of a relation between two consecutive nodes is 1.
• delta: the syntax e.delta is used to refer the relative position (an integer) between two ordered nodes.
• __id__: internal identifier, useful for dealing with subset of equivalent nodes in a request (see here)

Graph input formats

To describe a graph in practice, Grew offers several input formats:

• JSON
• CoNLL-U format
• AMR format: Grew is also able to read data following format used for instance in corpora freely available on AMR web page.

Graph output formats

• CoNLL-U: this is the format used by default with grew transform
• JSON: available with -json argument on the command line
  • if the output contains one graph, the CoNLL code of the graph given
  • if the output contains zero or more than two graphs, a JSON list is returned
• multi JSON: available with -multi_json argument on the command line. Each graph is written is a separate file. With grew transform … -o out.json -multi_json, files are named out__0.json, out__1.json…
• Graphviz dot: available with -dot argument on the command line