calliope-ts 0.0.4 → 0.1.0

package/src/clio/tagfix.ts

@@ -1,104 +0,0 @@
-// tagfix.ts — Pre-parse POS-tag correction layer.
-//
-// FinNLP's en-pos tagger is structurally sound but carries a small tail of
-// SYSTEMATIC tag errors that matter enormously for verse analysis, because a
-// wrong tag flips a word's content/function status (→ its stress tier) and
-// derails the en-parse dependency tree built from the tags.  This pass runs
-// BETWEEN en-pos and en-parse (see parseDocument in parser.ts), so corrected
-// tags repair both the tagging AND the resulting dependency structure — a
-// post-hoc fix of the parse could never do that.
-//
-// Every rule below targets an error class actually observed in this repo's
-// trials; rules are deliberately narrow (anti-gaming: each must be justified
-// by the error it fixes, not by benchmark deltas).
-
-/** Zero-derived irregular past participles that en-pos tags NN/VBP after a
- *  have-auxiliary ("had quit", "has put", "have read").  Only forms whose
- *  participle is identical to the base/noun spelling — the -en/-ed forms tag
- *  fine on their own. */
-const ZERO_PARTICIPLES = new Set([
-  'quit', 'put', 'set', 'cut', 'hit', 'let', 'shut', 'cast', 'cost', 'hurt',
-  'burst', 'split', 'spread', 'bet', 'wed', 'read', 'rid', 'shed', 'thrust',
-  'slit', 'bid', 'broadcast', 'upset', 'sunburst',
-]);
-
-const HAVE_FORMS = new Set(['have', 'has', 'had', 'having', "'ve", "'d"]);
-
-/** Archaic / Early-Modern-English forms en-pos has no lexicon entries for —
- *  ubiquitous in the verse this toolkit exists to scan. */
-const ARCHAIC_TAGS: Record<string, string> = {
-  thou: 'PRP', thee: 'PRP', ye: 'PRP',
-  thy: 'PRP$', thine: 'PRP$',
-  art: 'VBP', wert: 'VBD', wast: 'VBD',
-  doth: 'VBZ', hath: 'VBZ', dost: 'VBZ', hast: 'VBZ', saith: 'VBZ',
-  didst: 'VBD', hadst: 'VBD', wouldst: 'MD', couldst: 'MD', shouldst: 'MD',
-  shalt: 'MD', wilt: 'MD', canst: 'MD', mayst: 'MD', 'mightst': 'MD',
-  wherefore: 'WRB', whither: 'WRB', whence: 'WRB',
-  hither: 'RB', thither: 'RB', yon: 'JJ', yonder: 'RB',
-  ere: 'IN', oft: 'RB', anon: 'RB',
-};
-
-/**
- * Correct a sentence's tags in place-safe fashion (returns a new array).
- * `tokens` and `tags` are the en-pos outputs, index-aligned.
- */
-export function correctTags(tokens: string[], tags: string[]): string[] {
-  const out = tags.slice();
-  for (let i = 0; i < tokens.length; i++) {
-    const w = tokens[i].toLowerCase();
-
-    // 1. The pronoun "I".  en-norm lowercases sentence-initial "I" → "i",
-    //    which en-pos then reads as a foreign word / letter name (FW).
-    if (w === 'i' && out[i] === 'FW') out[i] = 'PRP';
-
-    // 2. Archaic forms (thou/thy/doth/shalt/wherefore…): en-pos guesses
-    //    NN/JJ/FW for these, wrecking both stress class and the parse.
-    //    Guard "art": only when a pronoun precedes ("thou art"), since the
-    //    noun reading ("the art of…") is the modern default.
-    const archaic = ARCHAIC_TAGS[w];
-    if (archaic && !/^(NNP|NNPS)$/.test(out[i])) {
-      if (w === 'art') {
-        const prev = i > 0 ? tokens[i - 1].toLowerCase() : '';
-        if (prev === 'thou' || prev === 'ye' || prev === 'you') out[i] = 'VBP';
-      } else {
-        out[i] = archaic;
-      }
-    }
-
-    // 3. Perfect-tense zero participles: have-form + ("quit"/"put"/"read"…)
-    //    tagged as NN/VBP/VBD → VBN, so en-parse builds the verb chain
-    //    instead of treating the participle as a direct-object noun
-    //    ("I had quit the programming paradigm").  An intervening adverb
-    //    ("had just quit") is allowed.
-    if (ZERO_PARTICIPLES.has(w) && /^(NN|NNS|VBP|VBD|VB)$/.test(out[i])) {
-      const prev1 = i > 0 ? tokens[i - 1].toLowerCase() : '';
-      const prev2 = i > 1 ? tokens[i - 2].toLowerCase() : '';
-      const prev1IsAdv = i > 0 && /^RB/.test(out[i - 1]);
-      if (HAVE_FORMS.has(prev1) || (prev1IsAdv && HAVE_FORMS.has(prev2))) {
-        out[i] = 'VBN';
-      }
-    }
-
-    // 4. Impossible gerunds: a VBG tag on a token that does not end in
-    //    -ing/-in' cannot be a gerund/present participle — it is an en-pos
-    //    lexicon glitch.  The right tag depends on context: before a noun it
-    //    is a noun modifier ("wisdom"/VBG teeth → NN); after a subject
-    //    pronoun it is a finite verb ("as they bicycle/VBG through" → VBP,
-    //    which keeps "through" a phrasal particle in the parse).  With no
-    //    deciding context, leave the tag alone (en-parse treats VBG
-    //    verb-ishly, the safer default).
-    if (out[i] === 'VBG' && !/in[g'’]?$/.test(w)) {
-      const prevTag = i > 0 ? out[i - 1] : '';
-      const nextTag = i + 1 < tokens.length ? out[i + 1] : '';
-      if (/^NNS?$/.test(nextTag)) out[i] = 'NN';
-      else if (prevTag === 'PRP') out[i] = 'VBP';
-    }
-
-    // 5. Vocative "O" ("O wild West Wind"): en-pos gives NNP/JJ; it is an
-    //    interjection (and must not become a content word with a beat by
-    //    default).  Only the bare capital O — "o'er" etc. are handled by the
-    //    aphaeresis lexicon in stress.ts.
-    if (tokens[i] === 'O' && i + 1 < tokens.length && out[i] !== 'UH') out[i] = 'UH';
-  }
-  return out;
-}

package/src/depfix.ts

@@ -1,88 +0,0 @@
-// depfix.ts — Post-parse dependency repair via DepEdit rules (the `depedits`
-// npm package, the maintainer's TypeScript port of DepEdit).
-//
-// Runs AFTER en-parse, complementing the pre-parse tag corrections in
-// tagfix.ts: tagfix repairs what the tagger got wrong before the tree is
-// built; this pass repairs systematic attachment errors en-parse makes even
-// with correct tags.  Rules are written in DepEdit's declarative format
-// (definitions ⟶ relations ⟶ actions, tab-separated) over en-parse's own
-// label space (DOBJ/NSUBJ/DEP/…), so the round-trip is lossless and every
-// rule is independently testable.
-//
-// The rule set is deliberately small and evidence-based — each rule cites the
-// observed failure it corrects.  `depedits` is ESM-only; it is loaded lazily
-// and failures degrade gracefully (the unrepaired parse is still a parse).
-
-import { createRequire } from 'module';
-
-interface FinDepNode {
-  label: string;      // dependency label, e.g. "NSUBJ", "ROOT"
-  type: string;       // phrase type, e.g. "NP", "VP"
-  parent: number;     // 0-based index of governor token; -1 for root
-}
-
-// Observed failure (probe: "I had quit the programming paradigm"): en-parse
-// attaches BOTH nouns of a noun compound to the verb as parallel objects
-// ("programming ←DOBJ← quit", "paradigm ←DOBJ← quit"), and leaves the
-// determiner dangling on the first noun as generic DEP.  The repairs:
-//   1. Two adjacent common nouns sharing a governor with the same object
-//      relation → the first is a compound modifier (AMOD) of the second.
-//   2. A determiner left as DEP on a noun that has become a modifier →
-//      re-attach it as DET to that noun's head (the true NP head).
-const CALLIOPE_DEP_FIXES = [
-  'xpos=/NNS?/&func=/DOBJ|IOBJ/;xpos=/NNS?/&func=/DOBJ|IOBJ/;xpos=/VB.*/\t#3>#1;#3>#2;#1.#2\t#2>#1;#1:func=AMOD',
-  'xpos=/DT/&func=/DEP|EXT/;xpos=/NNS?/&func=/AMOD/;xpos=/NNS?.*/\t#2>#1;#3>#2\t#3>#1;#1:func=DET',
-].join('\n');
-
-let engine: { process(conllu: string): string } | null | undefined;
-
-function loadEngine(): typeof engine {
-  if (engine !== undefined) return engine;
-  try {
-    // This package compiles to ESM, where bare `require` does not exist, and
-    // the parse path is synchronous, so dynamic import() is not an option:
-    // createRequire gives a sync loader, and since `depedits` is itself
-    // ESM-only this resolves via Node's require(esm) (≥20.17 / ≥22.12).  On
-    // older runtimes it throws and the repair pass degrades to a no-op (the
-    // unrepaired parse is still a parse).
-    const req = createRequire(import.meta.url);
-    const { DepEditEngine } = req('depedits');
-    const e = new DepEditEngine();
-    e.loadIniString(CALLIOPE_DEP_FIXES);
-    engine = e;
-  } catch {
-    engine = null;
-  }
-  return engine;
-}
-
-/**
- * Repair systematic en-parse attachment errors.  Returns a new deps array
- * (same shape as en-parse's `toArray` output); on any failure returns the
- * input unchanged.
- */
-export function applyDepFixes(tokens: string[], tags: string[], deps: FinDepNode[]): FinDepNode[] {
-  const e = loadEngine();
-  if (!e || tokens.length === 0 || deps.length !== tokens.length) return deps;
-  try {
-    const conllu = tokens.map((tok, i) => {
-      const head = deps[i].parent >= 0 ? deps[i].parent + 1 : 0;
-      const safe = tok.replace(/\s/g, '_') || '_';
-      return `${i + 1}\t${safe}\t${safe}\t_\t${tags[i] || '_'}\t_\t${head}\t${deps[i].label || 'DEP'}\t_\t_`;
-    }).join('\n') + '\n\n';
-    const out = e.process(conllu);
-    const fixed: FinDepNode[] = deps.map(d => ({ ...d }));
-    for (const row of out.split('\n')) {
-      const cols = row.split('\t');
-      if (cols.length < 10) continue;
-      const idx = parseInt(cols[0], 10) - 1;
-      if (!(idx >= 0 && idx < fixed.length)) continue;
-      const head = parseInt(cols[6], 10);
-      fixed[idx].parent = Number.isFinite(head) ? head - 1 : fixed[idx].parent;
-      if (cols[7] && cols[7] !== '_') fixed[idx].label = cols[7];
-    }
-    return fixed;
-  } catch {
-    return deps;
-  }
-}