Graph Extension Grammars (GEGs)

• Graph Extension Grammars are a formalism using graph algebras and regular tree grammars to specify and parse labeled, directed graphs with non-structural reentrancies.
• They combine binary union and unary extension operations to enable arbitrary node-sharing, crucial for modeling phenomena in semantic representations such as AMR.
• GEGs offer a tractable, polynomial-time parsing algorithm that employs memoization and profile matching to efficiently fuse graph components during derivation.

Graph Extension Grammars (GEGs) are a formalism introduced to address the expressive and computational requirements of generating and parsing languages of directed, node- and edge-labelled graphs, particularly supporting non-structural reentrancies as found in semantic representations such as Abstract Meaning Representation (AMR). A GEG consists of an algebra over graphs—incorporating specific graph operations—and a regular tree grammar that generates expressions over these operations. The resulting framework allows specification of sets of graphs with both rich structural constraints and tractable polynomial-time parsing (Björklund et al., 2021).

1. Formal Structure of Graph Extension Grammars

A GEG is fundamentally composed of two interlocked components: (i) a graph-extension algebra specifying operations over families of typed, labelled graphs, and (ii) a regular tree grammar generating terms over those operations. The graph domains are defined as follows:

Let $\Sigma_n$ be a finite set of node-labels and $\Sigma_e$ a finite set of edge-labels. For each nonnegative integer $\tau$, $\mathbb{G}_\tau$ is the set of all finite, directed, node- & edge-labelled graphs $G = (V, E, \ell, port)$ with $|port| = \tau$, where:

• $V$ is a finite set of nodes,
• $E \subseteq V \times \Sigma_e \times V$,
• $\ell \colon V \to \Sigma_n$,
• $port \in V^\tau$ is a sequence of $\Sigma_e$0 (possibly repeating) nodes called the ports.

The union $\Sigma_e$1 collects all such graphs, parameterized by type $\Sigma_e$2.

Operations in the Signature

The signature $\Sigma_e$3 comprises two key forms:

• Binary union operations $\Sigma_e$4, which yield disjoint unions (up to renaming) and concatenate ports.
• Unary extension operations $\Sigma_e$5 specified by tuples

$\Sigma_e$6

where $\Sigma_e$7 is a graph of type $\Sigma_e$8; $\Sigma_e$9 is a sequence of $\tau$0 dock-nodes; $\tau$1 is the set of "clonable" context nodes. These are subject to: - (R1) All edge sources in $\tau$2 are from $\tau$3, - (R2) Every non-port node in $\tau$4 is a target of some edge in $\tau$5.

Extension $\tau$6 is of type $\tau$7 where $\tau$8, $\tau$9. The operation $\mathbb{G}_\tau$0 acts nondeterministically by attaching new structure and enabling node-fusion as described below.

The algebra is completed by including the constant operator for the empty graph $\mathbb{G}_\tau$1 of type 0.

Algebraic Semantics

A graph-extension algebra is a many-sorted algebra

$\mathbb{G}_\tau$2

with $\mathbb{G}_\tau$3 and the above-specified operations, producing sets of graphs. Reachability is enforced: every node of every generated graph is reachable from some port-node.

2. Regular Tree Grammar Mechanism

A GEG employs a regular, many-sorted tree grammar $\mathbb{G}_\tau$4 with finitely many nonterminals $\mathbb{G}_\tau$5, each assigned a type $\mathbb{G}_\tau$6. Productions are of the form

$\mathbb{G}_\tau$7

where $\mathbb{G}_\tau$8 has type $\mathbb{G}_\tau$9. The generated tree language $G = (V, E, \ell, port)$0 encodes valid sequences of operations.

A full GEG is then

$G = (V, E, \ell, port)$1

with associated graph language

$G = (V, E, \ell, port)$2

obtained by recursively mapping derivation trees to their semantic value in the algebra.

Evaluation of $G = (V, E, \ell, port)$3 proceeds by induction:

• Base: If $G = (V, E, \ell, port)$4, $G = (V, E, \ell, port)$5.
• Union: If $G = (V, E, \ell, port)$6, then $G = (V, E, \ell, port)$7.
• Extension: If $G = (V, E, \ell, port)$8 with $G = (V, E, \ell, port)$9 of type $|port| = \tau$0, then $|port| = \tau$1.

3. Non-Structural Reentrancy Modelling

GEGs natively support non-structural reentrancy, a key limitation of prior devices like hyperedge-replacement grammars where only context-free attachments are possible. Non-structural reentrancy refers to allowing nodes in a partially constructed graph to appear multiple times as targets of newly introduced edges, a necessity for capturing phenomena such as shared arguments or referents in AMR.

This is realized in the GEG formalism via the set $|port| = \tau$2 of clonable context-nodes within each unary extension operation $|port| = \tau$3. When $|port| = \tau$4 is applied:

• Each $|port| = \tau$5 can be cloned an arbitrary number of times (non-deterministically).
• Each clone is injectively merged ("fused") to a distinct node of $|port| = \tau$6 with the same node-label.
• This process enables one or several new outgoing edges from operations to point to the same node(s) in $|port| = \tau$7, thus creating arbitrary non-structural reentrancies.

The use of $|port| = \tau$8 thus controls and enables the expressive generativity over non-structural reentrancies—an essential feature missing from many prior graph grammar formalisms.

4. Parsing Algorithms and Complexity

The challenge of parsing a given graph $|port| = \tau$9 with respect to a GEG $V$0 is to decide whether $V$1. This is determined via the existence of a derivation tree $V$2 and an assignment of semantic values yielding exactly $V$3. GEGs provide a polynomial-time parsing algorithm, summarized as follows:

• The principal recursive routine $V$4 tests whether the induced subgraph from port-sequence $V$5 in $V$6 is derivable by nonterminal $V$7.
• Results are memoized in a table $V$8.
• Union productions and extension rules are handled with explicit splitting or matching logic.
• Matching for extension rules involves fusing dock-nodes to the graph at $V$9 and considering all ways context-clones may merge, while preserving labels and the "clone only clonable" constraint.

Data Structures and Efficiency

Key data structures include:

• Profiles $E \subseteq V \times \Sigma_e \times V$0 for nodes in $E \subseteq V \times \Sigma_e \times V$1,
• Likewise, $E \subseteq V \times \Sigma_e \times V$2 for nodes in the operation's underlying graph, to efficiently check compatibility during matching and merging.

Complexity Summary

With $E \subseteq V \times \Sigma_e \times V$3 and $E \subseteq V \times \Sigma_e \times V$4, the algorithm executes in time $E \subseteq V \times \Sigma_e \times V$5, primarily determined by the combinatorics of ports and clones. Under additional unambiguity constraints on profiles, complexity can be reduced to $E \subseteq V \times \Sigma_e \times V$6 or even linear time, as per corollaries in the primary reference (Björklund et al., 2021).

Parsing Step	Complexity	Key Limitation/Optimization
Memo table computation	$E \subseteq V \times \Sigma_e \times V$7
Union production handling	$E \subseteq V \times \Sigma_e \times V$9	Constant number per nonterminal
Extension handling (naive)	$\ell \colon V \to \Sigma_n$0	Filters via profile comparison
Overall (optimized)	$\ell \colon V \to \Sigma_n$1	Profile constraints

5. Representative Example: Reentrant Graph Generation

Consider a minimal GEG showcasing non-structural reentrancy:

• Node labels: $\ell \colon V \to \Sigma_n$2,
• Edge labels: $\ell \colon V \to \Sigma_n$3,
• Extension op $\ell \colon V \to \Sigma_n$4 of type $\ell \colon V \to \Sigma_n$5:
  • $\ell \colon V \to \Sigma_n$6,
  • $\ell \colon V \to \Sigma_n$7, $\ell \colon V \to \Sigma_n$8, $\ell \colon V \to \Sigma_n$9,
  • $port \in V^\tau$0,
  • $port \in V^\tau$1.
• The regular tree grammar uses nonterminals $port \in V^\tau$2 (type 1), $port \in V^\tau$3 (type 1), $port \in V^\tau$4 (type 0), with productions:
  • $port \in V^\tau$5
  • $port \in V^\tau$6
  • $port \in V^\tau$7

In a derivation sequence, repeated applications of $port \in V^\tau$8 followed by union yield a graph $port \in V^\tau$9 (of type 1) with a "b"-node and two parallel $\Sigma_e$00-edges targeting a single "a"-node. Here, the arbitrary cloning and merging induced by $\Sigma_e$01 enables this reentrancy structure, directly modeling the non-structural scenario within the formalism (Björklund et al., 2021).

6. Position within Generative Graph Formalisms and Related Work

GEGs make several advances within the family of generative graph formalisms:

• In comparison to hyperedge-replacement grammars, GEGs transcend the context-free model by enabling arbitrary (non-structural) node-sharing.
• The extension mechanism is inspired in part by concepts such as adaptive star grammars (notably the "cloning" device) (Björklund et al., 2021).
• The framework is directly motivated by requirements from natural language processing, particularly representations such as AMR, which routinely employ non-structural reentrancies.
• The computational tractability of parsing distinguishes GEGs among more expressive, but intractable, grammar variants.

Connections to prior work include foundational treatments in graph grammar theory [Courcelle & Engelfriet, 2012], as well as device comparisons with weighted DAG automata for semantic graphs [Chiang et al., 2018].

• Björklund, Henrik; Drewes, Frank; Jonsson, Peter: "Polynomial Graph Parsing with Non-Structural Reentrancies" (Björklund et al., 2021)
• Courcelle, B.; Engelfriet, J.: Graph Structure and MSO Logic (2012).
• Drewes, F.; et al.: Adaptive Star Grammars (2010).
• Chiang, D.; et al.: Weighted DAG Automata for Semantic Graphs (2018).