📄 Converting the Penn Treebank to Systemic Functional Grammar

Summary

Honnibal tackles the absence of computational resources for Systemic Functional Grammar (SFG): no annotated corpora, no standard machine-readable format, and no usable parsers. His solution is to derive an SFG corpus from the Penn Treebank. In the Treebank, semantically important choices such as mood, tense, voice and polarity are only implicit in tree shape. The conversion makes them explicit.

A Python script applies hand-written transformation steps:

• it flattens the phrase structure into SFG's minimal-bracketing clause constituency;
• it separates clause complexes from verbal group complexes;
• it handles ellipsis;
• it adds interpersonal and textual function labels.

Systems signalled by word order or by grammatical items can be derived from the tree. These include mood, theme, voice, tense, polarity, status and clause class. Transitivity depends on lexical choice, so it is left out.

On 100 held-out clauses, 1198 of 1200 annotated properties were correct. The outcome is more than two million words of SFG-analysed text, which links NLP to a long tradition of meaning-oriented grammar.

Key points

• The source data is the Penn Treebank, specifically the Wall Street Journal section, which was the only part available. The result is an SFG constituency corpus of over two million words.
• The conversion uses hard-coded transformation functions in Python rather than induced rules. Its acknowledged weakness is that no explicit conversion grammar comes out, so fixing systematic errors and extending the script is harder.
• The main structural steps are:
  • raising complements and adjuncts out of the VP to the clause;
  • lifting hypotactic clauses to sibling position;
  • collapsing auxiliary chains;
  • detecting verbal group complexes;
  • building new clauses for ellipsed subjects.
• Clause-rank systems are annotated for two metafunctions:
  • interpersonal: mood with Subject, Finite, Predicator, Adjunct, Complement and Vocative roles, plus clause class, status, tense and polarity;
  • textual: theme (Textual, Interpersonal and Topical Theme versus Rheme) and voice.
  • Ideational analysis (transitivity) is excluded.
• Evaluation covered 100 unseen clauses, checking constituency plus eleven function structures. Results were 99.83% of properties correct and 99% of clauses fully correct. Both errors trace back to one topicalised indirect-speech clause, where a wrong status value cascaded into a wrong mood type.

Interesting ideas

• I1. Before choosing what to annotate, sort target categories by how they are signalled (word order, grammatical items, or lexical choice); only the first two can be read deterministically off a tree, while the third needs a lexicon.
• I2. Turn generic parse nodes into typed constituent objects (clause, verbal group, other group, word) first, then run an ordered series of small, targeted restructuring functions on them.
• I3. Voice is fixed by a cheap deterministic rule: passive when the final verb is VBN and the word before it is a form of 'be'; non-finites are typed by the Predicator's tag; and when no Complement exists, the clause counts as middle.
• I4. Grammatical categories form networks with entry conditions, so a single upstream mistake (status) silently triggers a downstream one (mood type); scoring both per property and per clause exposes this.
• I5. Theme is the first Adjunct, Subject, Complement or Predicator, with conjunctions before it counted as textual and vocatives or finites as interpersonal; the paper also notes that modal and conjunctive adjuncts must not be counted as topical.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): Use this test to decide which frame features (voice, polarity, mood) prag computes by rule from the UD tree and which (process type, semantic roles) must come from VerbNet.
• I2 (tree): UD dialect conversion could type nodes the same way before applying transforms, though rules should be stored as declarative data rather than buried in code, to avoid the maintenance trap the author admits.
• I3 (frames): A transparent passive/middle detector like this lets prag swap agent and patient roles with a readable justification attached to each frame.
• I4 (learn): Evaluation should report error dependencies between pipeline layers, so that repair rules aim at the root cause rather than its symptoms.
• I5 (back): The generator could choose the clause's starting point from this scheme, so that information order in regenerated text follows what the meaning graph puts in focus.