There's a native way to avoid that recursion.

  var snap = document.evaluate('//text()', document, null, XPathResult.ORDERED_NODE_SNAPSHOT_TYPE, null );
  for (var i=0;i<snap.snapshotLength;i++){t=snap.snapshotItem(i);t.textContent = replace(t.textContent);}

Also, the regexps would probably be better like /\belection\b/gi, to avoid eg reelection->reeating contest.

Could you possibly minify it?