fst_scorer.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0 <s>\n",
      "5.98510933 ABOUT\n",
      "5.46656704 HE\n",
      "4.97864962 SAID\n",
      "3.17434382 </s>\n",
      "40.5448189\n"
     ]
    }
   ],
   "source": [
    "# You need Kaldi and WSJ corpus to create this FST-based LM\n",
    "# $ gunzip -c 13-32.1/wsj1/doc/lng_modl/base_lm/tcb20onp.z | arpa2fst - lm.fst\n",
    "# $ pip install openfst-python\n",
    "import openfst_python as fst\n",
    "\n",
    "\n",
    "class FSTScorer:\n",
    "    def __init__(self, input, space=\" \", unk_label=\"<UNK>\"):\n",
    "        if isinstance(input, str):\n",
    "            self.fst = fst.Fst.read(input)\n",
    "        else:\n",
    "            self.fst = input\n",
    "        self.space = space\n",
    "        self.word2id = dict()\n",
    "        self.id2word = dict()\n",
    "        for k, v in self.fst.input_symbols():\n",
    "            assert v not in self.word2id\n",
    "            self.id2word[k] = v\n",
    "            self.word2id[v] = k\n",
    "        self.unk_ilabel = self.word2id[unk_label]\n",
    "        self.unk_score = fst.Weight(self.fst.weight_type(), self.word2id[unk_label])\n",
    "\n",
    "    def arc_weight(self, score, state, dst):\n",
    "        for a in self.fst.arcs(state):\n",
    "            if a.ilabel == dst:\n",
    "                print(a.weight, self.id2word[a.ilabel])\n",
    "                return fst.times(score, a.weight), a.nextstate\n",
    "            if a.ilabel == 0:\n",
    "                return self.arc_weight(fst.times(score, a.weight), a.nextstate, dst)\n",
    "        # assert False\n",
    "        return self.arc_weight(score, state, self.unk_ilabel)\n",
    "\n",
    "    def score_isym(self, isyms):\n",
    "        state = self.fst.start()\n",
    "        ret = fst.Weight.One(self.fst.weight_type())\n",
    "        for i in isyms:\n",
    "            ret, state = self.arc_weight(ret, state, i)\n",
    "        return float(ret)\n",
    "\n",
    "    def score_str(self, s):\n",
    "        return self.score_isym((self.word2id.get(i, self.unk_ilabel) for i in s.split(self.space)))\n",
    "\n",
    "\n",
    "scorer = FSTScorer(\"./lm_tgpr_5k.fst\")\n",
    "print(scorer.score_str(\"<s> ABOUT HE SAID </s>\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}