fudg_for_parsing.md

FUDG representation for parsing

Nathan Schneider & Lingpeng Kong
2013-10-25

Apart from underspecification and coreference, which we will not be predicting in the output of our parser, the FUDG formalism differs from traditional dependency parsing in a few respects:

1. Special demarcation of utterances within the input
2. Ability to exclude some input tokens from the parse
3. Ability to group tokens into multiword lexical nodes
4. Special representation of coordinate structures

All of these are specified in a GFL annotation. We propose to treat (1) and (2) as preprocessing steps at test time, so the TurboParser model assumes it sees one utterance at a time and knows which tokens are relevant.

(3) and (4) can be treated by transforming the FUDG graph into a traditional dependency tree with special edges. This process is deterministic and bidirectional, so at test time the TurboParser prediction can be converted to the FUDG-style graph. An example illustrating the transformations:

I want fresh bread and1 butter and2 jam as1 well as2 tea.

GFL:

I > want < $awa
$awa :: {$aa tea} :: {[as~1 well as~2]}
$aa :: {bread butter jam} :: {and~1 and~2}
fresh > $aa

Transformed FUDG:

I > want < as~1

well -MWE> as~1 <MWE- as~2

and~1 -CONJ> as~1
tea -CONJ> as~1

and~1 <COORD- and~2

bread -CONJ> and~1
butter -CONJ> and~1
jam -CONJ> and~1
fresh > and~1

Note that:

• MWEs are encoded as the first token heading the other tokens. Constraints: MWE edges are sensitive to original token order (the head must appear earlier); MWE edge dependents are not allowed to have their own dependents.
• Coordinators serve as heads of the coordinated phrase, removing the need for special coordinator nodes. The conjuncts attach with CONJ edges; modifiers attach with unlabeled edges. If there are multiple coordinators, the first one heads them all with COORD edges (and nothing is allowed to attach to these other coordinators).