📄 NLML — a Markup Language to Describe the Unlimited English Grammar

Summary

Jia tackles a practical gap in symbolic English parsing. Parsers such as an AGFL-generated one produce deeply nested trees whose labels (NP, subject, verb) mean something to linguists but are hard for downstream programs to use. The problem is worst when a sentence holds several verbs or subjects, as in relative clauses. The author's solution is NLML, a tag-based output format produced through AGFL's transduction mechanism. Its context-free grammar draws on the grammar textbooks by Chalker and by Hanks & Grandison. NLML explicitly labels mood, complexity, subjects, verb phrases, tense, person, number, circumstances and clause structure at three levels: sentence, clause and phrase. The paper shows how arbitrary utterances, including dialogue fragments, map onto this markup. It also presents two uses inside the CSIEC language-learning dialogue project. NLOMJ maps the markup onto Java objects, and NLDB stores NLML strings in database tables so earlier discourse can be reloaded without reparsing. The point is a directly queryable, editable and storable representation of grammar for dialogue systems.

Key points

• NLML is produced by AGFL's transduced output form, not by post-processing the usual indented or bracketed parse tree.
• The top-level expression rule lists 16 alternatives covering 11 <mood> categories:
  • full-sentence moods: statement, question, order, full exclamation;
  • fragment moods: np, adj, about, circumstances, what/how terse exclamations, subcircum.
• A max-matching priority makes full-sentence readings win, and $PENALTY weights push fragment readings down.
• The lexicon is an editable text file that stores word types, affixes and probabilities; AGFL calls the resulting probability-aware parsing hybrid parsing.
• NLOMJ models grammar elements as Java classes using inheritance and association.
• NLDB stores NLML strings in database tables split into facts, questions and fact relations. This supports context analysis for dialogue generation in CSIEC.

Interesting ideas

• I1. The parser emits a self-describing structure with named slots such as subject, verb word and mood, so a consumer can ask "what is the subject?" without searching the tree; this also resolves ambiguity when there are several clauses.
• I2. Max-matching means trying a full-sentence reading first and only falling back, under explicit penalties, to fragment readings such as a bare NP, adjective or adverbial; this suits dialogue input that is often elliptical.
• I3. Fragments receive both a mood label and their essential content word: "What a pity!" is a terse exclamation whose core is the noun "pity".
• I4. Sentence-level transformations such as negation or statement-to-question are done by editing the structured markup rather than the surface string.
• I5. Analyses are stored as serialized strings, filed by type (facts, questions, relations between facts), and rehydrated later without reparsing.

For Mova

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

• I1 (frames): Frames could be exposed as named per-clause slots tied to a specific clause node, so questions about multi-verb sentences return the right predicate's arguments.
• I2 (tree): Fragments like "Why?" or "That one you have read." could get a scored fallback analysis instead of a forced sentence parse.
• I3 (frames): Speech-act tagging could pair the act type with its head element, helping link short replies to the previous turn.
• I4 (back): Generation could implement mood and polarity changes as edits on the meaning graph before realization, making them explainable and testable.
• I5 (world): Text-world snapshots could be persisted per category so a long discourse can be reloaded and extended cheaply.