Notes on Serialization

A PENMAN-serialized graph takes the form of a tree with labeled reentrancies, so in deserialization it is first parsed directly into a tree and then the pure graph is interpreted from it.

(b / bark-01
   :ARG0 (d / dog))

The above PENMAN string is parsed to the following tree:

Tree(('b', [(':instance', 'bark-01'),
            (':ARG0', ('d', [(':instance', 'dog')]))]))

The structure of a tree node is (var, branches) while the structure of a branch is (role, target). The target of a branch can be an atomic value or a tree node. This tree is then interpreted to the following graph (triples and associated layout markers):

Graph(triples=[
       ('b', ':instance', 'bark-01'),
       ('b', ':ARG0', 'd'),
       ('d', ':instance', 'dog')
      ],
      epidata={
       ('b', ':ARG0', 'd'): [Push('d')],
       ('d', ':instance', 'dog'): [POP]
      })

Serialization goes in the reverse order: from a pure graph to a tree to a string.

Allowed Graphs

The Penman library robustly allows some kinds of invalid and unconventional graphs.

Unproblematic:

# Normal
(a / a-label :ROLE (b / b-label))

# Unlabeled nodes, edges
(a :ROLE (b))
(a / a-label : (b / b-label))
(a : (b))

# Cycles
(a :ROLE (b :ROLE a))

# Distributed nodes
(a :ROLE (b :ROLE (c / c-label)) :ROLE2 (c :ATTR val))

Allowed but Unconventional

# Empty
()

# Missing edge target
(a / a-label :ROLE )

# Missing node label
(a / :ROLE (b / b-label))

# Inverted attributes
(a / a-label :ARG0-of 2)

Disallowed

# Disconnected (parseable as two separate graphs)
(a / a-label)(b / b-label)

# Missing identifiers
(a :ROLE ( / b-label))

# Misplaced label
(a :ROLE (b) / a-label)

# Multiple labels
(a / a-label / another-label)