📄 Assisted Nominalization for Academic English Writing

Summary

Non-native university students often struggle with the formal style of academic English. A key feature of that style is nominalization, which recasts events as nouns. The authors propose a writing assistant that moves a sentence along Halliday's cline of metaphoricity. It turns a complex clause, such as a main clause plus an adverbial subordinate clause, into a simplex clause in which two noun phrases are linked by a verb like "cause."

The system is a three-stage rule pipeline:

• SpaCy produces a UD parse, and the clauses are located through the advcl and mark relations.
• A nominalizer rewrites each clause as a noun phrase, using conversion lists derived from NOMLEX and CATVAR.
• A generator joins the two phrases according to a discourse relation detected from keywords.

A small pilot on Wikipedia sentences found the system good at catching clauses that deserve nominalization. However, it also proposed many rewrites that a human expert rejected. It therefore suits advanced learners who can filter its suggestions. The work matters because it shifts writing support from fixing errors toward guiding style.

Key points

• Clause detection works on Universal Dependencies output from the SpaCy parser. The advcl head gives the main clause, its dependent gives the subordinate clause, and mark gives the connective.
• The conversion lists contain 7,879 verb→noun pairs, 11,369 adjective→noun pairs and 2,834 adverb→adjective pairs.
  • Verb→noun pairs come from NOMLEX NOM entries, with CATVAR as a fallback.
  • When a verb has several noun candidates, the system keeps the one that is most frequent in Google Web 1T and has a typical noun suffix.
  • All mappings are one-to-one.
• Copulas, modals and negated verbs are deliberately left untouched, because how they nominalize depends heavily on context.
• Discourse relations are identified by keyword spotting:
  • causal: because, since, so, therefore
  • concession: although, despite, even though
  • temporal: after, before
  • The user chooses the final linking verb.
• In the test, the system tried to nominalize all 33 clauses from 25 Wikipedia sentences, while the expert rewrote only 18. On those 18, the output was identical in 55.6%, needed minor edits in 16.7% and needed major edits in 27.8%.

Interesting ideas

• I1. The paper treats one meaning as a ladder of paraphrases, from separate sentences joined by a conjunction down to a single dense noun group, and positions a tool as a step along that ladder.
• I2. Running the rule in reverse is just as informative: a sentence like "her ignorance caused her death" hides two events and a causal link between them inside nouns.
• I3. When a derivational resource offers several candidates, the paper picks one using corpus frequency plus a morphological suffix check, which is a cheap and transparent tie-breaker.
• I4. Errors are scored on a graded scale: identical except for word order, a different determiner or preposition, or a different head lexeme. This separates surface slips from real lexical-choice failures.
• I5. The rules explicitly abstain on constructions whose transformation is unreliable (copulas, modals, negation) instead of guessing.

For Mova

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

• I1 (back): The text → graph → text compressor could render a single causal frame pair at a chosen point on this cline, giving controllable register from one meaning graph.
• I2 (frames): A de-nominalization step that uses NOMLEX-style noun→verb mappings could recover the hidden who-did-what frames and turn the linking verb into an explicit causal or temporal relation.
• I3 (tokens): The same heuristic could seed a derivational table in crate en, linking verbs, adjectives and their noun forms, while flagging polysemous cases for later sense-aware refinement.
• I4 (learn): Generation and paraphrase outputs could be scored on this kind of ladder, so that article noise does not hide genuine lexical errors in evaluation reports.
• I5 (learn): Self-induced repair rules could declare similar abstention zones, so each rule records where it refuses to fire and stays explainable rather than overreaching.