Upload 89 files

app/engine/lexical.py

@@ -462,7 +462,44 @@ def _polish_headword_upgrade(synset, source_lemma: str, cand_lemma: str) -> bool
     if not lemmas or lemmas[0] != cand_lemma:
         return False
     source_rank = _source_lemma_rank(synset, source_lemma)
-    return 1 <= source_rank <= 2 and 2 <= len(lemmas) <= 4
+    return 1 <= source_rank <= 4 and 2 <= len(lemmas) <= 5
+
+
+def _headword_candidate_for_synset(synset, token, doc) -> tuple[str, float] | None:
+    """Try the synset headword directly for polish-only verb upgrades."""
+    lemmas = _synset_single_word_lemmas(synset)
+    if not lemmas:
+        return None
+    head = lemmas[0]
+    replacement = _inflect(head, token)
+    if not replacement or replacement.lower() == token.text.lower():
+        return None
+    source = token.lemma_.lower()
+    source_surface = token.text.lower()
+    source_frequency = max(
+        zipf_frequency(source, "en"),
+        zipf_frequency(source_surface, "en"),
+    )
+    left, right = _neighbor_words(doc, token)
+    source_phrase = _phrase_zipf(source_surface, left, right)
+    if not _collocation_ok(
+        source,
+        head,
+        token,
+        left=left,
+        right=right,
+    ):
+        return None
+    candidate_frequency = zipf_frequency(head, "en")
+    rank_frequency = zipf_frequency(replacement.lower(), "en")
+    if not _frequency_ok(source_frequency, max(candidate_frequency, rank_frequency)):
+        return None
+    phrase = _phrase_zipf(replacement.lower(), left, right)
+    if phrase + 0.55 < source_phrase:
+        return None
+    if not substitution_pos_stable(doc.text, token.i, replacement):
+        return None
+    return replacement, max(candidate_frequency, rank_frequency) - source_frequency
 
 
 def _pick_verb_candidate(
@@ -499,10 +536,25 @@ def _pick_verb_candidate(
     for synset in ordered:
         score = _sense_score(synset, context, stop_words)
         picked = _candidate_for_synset(synset, token, doc)
+        headword_polish = False
+        if (
+            polish
+            and not aggressive
+            and score < weak_gloss
+            and _source_lemma_rank(synset, src_lemma) >= 1
+        ):
+            headword_pick = _headword_candidate_for_synset(synset, token, doc)
+            if headword_pick is not None:
+                picked = headword_pick
+                headword_polish = True
         if not picked:
             continue
         replacement, simplicity_gain = picked
-        if score < effective_min and simplicity_gain < gain_floor:
+        if (
+            score < effective_min
+            and simplicity_gain < gain_floor
+            and not headword_polish
+        ):
             continue
         cand_lemma = _candidate_lemma_for_replacement(synset, token, replacement)
         high_gain_ok = False
@@ -519,7 +571,7 @@ def _pick_verb_candidate(
             # the source is a secondary lemma. Aggressive/ensure stay stricter.
             high_gain_ok = (
                 polish
-                and simplicity_gain >= 0.70
+                and simplicity_gain >= 0.08
                 and _polish_headword_upgrade(synset, src_lemma, cand_lemma)
             )
             if not unglossed_ok and not high_gain_ok: