PreviewQAExamples&BiasAnalysis.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "f71610d9",
   "metadata": {},
   "source": [
    "# For what is this notebook for?\n",
    "- Defining shortcut and anti-shortcut examples in multiple-choice QA."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7529eeb8",
   "metadata": {},
   "source": [
    "# Import & Definition"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "44a98814",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import torch\n",
    "import numpy as np\n",
    "import math\n",
    "from tqdm import tqdm\n",
    "from pathlib import Path\n",
    "from collections import Counter, defaultdict\n",
    "from squad import SquadV1Processor, SquadExample\n",
    "import spacy\n",
    "nlp = spacy.load(\"en_core_web_sm\")\n",
    "import matplotlib.pyplot as plt\n",
    "from utils import load_json, save_json\n",
    "import seaborn as sns\n",
    "sns.set()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "de14193d",
   "metadata": {},
   "outputs": [],
   "source": [
    "race_dir = os.environ['RACE_DIR']\n",
    "reclor_dir = os.environ['ReClor_DIR']\n",
    "analysis_dir = os.environ['RE_ANALYSIS_DIR']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "32db6d43",
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_mcqa_example(ex):\n",
    "    \"\"\"\n",
    "    example_id: str\n",
    "    question: str\n",
    "    contexts: List[str]\n",
    "    endings: List[str]\n",
    "    label: Optional[str]\n",
    "    \"\"\"\n",
    "    print(\"----\")\n",
    "    print(\"ID:\", ex.example_id)\n",
    "    print(\"Context 📗: \", ex.contexts[0])\n",
    "    print()\n",
    "    print(\"Question 💡:\", ex.question)\n",
    "    print(\"Options 🌿:\", ex.endings)\n",
    "    print(\"Answer 🍎:\", ex.label)\n",
    "    print(\"----\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "32d76955",
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_exqa_example(ex):\n",
    "    print(\"----\")\n",
    "    print(\"ID:\", ex.qas_id)\n",
    "    print(\"Context 📗: \", ex.context_text)\n",
    "    print()\n",
    "    print(\"Question 💡:\", ex.question_text)\n",
    "    print(\"Answer 🍎:\", ex.answers)\n",
    "    print(\"----\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "d4527951",
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_race_examples(ex):\n",
    "    assert isinstance(ex, dict)\n",
    "    print(\"----\")\n",
    "    if 'example_id' in ex:\n",
    "        print(\"ID:\", ex.example_id)\n",
    "    elif 'id' in ex['document']:\n",
    "        print(\"DocID:\", ex['document']['id'])\n",
    "    print(\"Context 📗: \", ex['document']['summary']['text'])\n",
    "    print()\n",
    "    print(\"Question 💡:\", ex['question']['text'])\n",
    "    print(\"Answers 🍎:\", [_a['text'] for _a in ex['answers']])\n",
    "    print(\"----\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "ed99184a",
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_analysis_result(key, name, task_type='ex-qa'):\n",
    "    analysis_task_dir = Path(analysis_dir) / task_type\n",
    "    analysis_file = analysis_task_dir / f\"{key}_{name}.json\"\n",
    "    analysis_result = load_json(analysis_file)\n",
    "    return analysis_result"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "70ab3065",
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_offset_id2w(doc):\n",
    "    # from char index to word index\n",
    "    id2w = {}\n",
    "    tm1 = 0\n",
    "    for i, t in enumerate(doc):\n",
    "        for j in range(tm1, t.idx):\n",
    "            # t.idx: char-level start position of t (token)\n",
    "            id2w[j] = i - 1\n",
    "        tm1 = t.idx\n",
    "    for j in range(t.idx, t.idx+len(t)+1):\n",
    "        id2w[j] = i\n",
    "    return id2w\n",
    "\n",
    "def get_common_ngram(sent1, sent2, n):\n",
    "    \"\"\"\n",
    "    sent1, sent2 (str): lower cased tokens joined with spaces\n",
    "    n (int): n-gram\n",
    "    \"\"\"\n",
    "    if len(sent1.split()) < n or len(sent2.split()) < n:\n",
    "        return 0\n",
    "    m = 0\n",
    "    tokens = sent2.split()\n",
    "    for i in range(len(tokens) - n + 1):\n",
    "        ngram = tokens[i:i+n]\n",
    "        assert len(ngram) == n\n",
    "        ngram = ' '.join(ngram)\n",
    "        if ngram in sent1:\n",
    "            m += 1\n",
    "    return m"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b1dbf465",
   "metadata": {},
   "source": [
    "# EX-QA"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "3011b4eb",
   "metadata": {},
   "outputs": [],
   "source": [
    "DATASET_DIR = Path('dataset')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "7ae99494",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 442/442 [00:41<00:00, 10.60it/s]\n"
     ]
    }
   ],
   "source": [
    "processor = SquadV1Processor()\n",
    "file = DATASET_DIR / 'ex-qa/squad/train-v1.1.json'\n",
    "squad_train_examples = processor.get_train_examples(None, filename=file, do_lower_case=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "87e6e49f",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 48/48 [00:04<00:00,  9.79it/s]\n"
     ]
    }
   ],
   "source": [
    "processor = SquadV1Processor()\n",
    "file = DATASET_DIR / 'ex-qa/squad/dev-v1.1.json'\n",
    "squad_dev_examples = processor.get_train_examples(None, filename=file, do_lower_case=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 555,
   "id": "dd03ecb5",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 1/1 [02:21<00:00, 141.71s/it]\n"
     ]
    }
   ],
   "source": [
    "processor = SquadV1Processor()\n",
    "file = DATASET_DIR / 'ex-qa/mrqa/train/NewsQA-train-from-MRQA.json'\n",
    "newsqa_train_examples = processor.get_train_examples(None, filename=file, do_lower_case=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "213d32db",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 1/1 [00:39<00:00, 39.23s/it]\n"
     ]
    }
   ],
   "source": [
    "processor = SquadV1Processor()\n",
    "file = DATASET_DIR / 'ex-qa/mrqa/dev/SearchQA-dev-from-MRQA.json'\n",
    "searchqa_dev_examples = processor.get_train_examples(None, filename=file, do_lower_case=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "f55d8c63",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 1/1 [00:19<00:00, 19.06s/it]\n"
     ]
    }
   ],
   "source": [
    "processor = SquadV1Processor()\n",
    "file = DATASET_DIR / 'ex-qa/mrqa/dev/TriviaQA-web-dev-from-MRQA.json'\n",
    "triviaqa_dev_examples = processor.get_train_examples(None, filename=file, do_lower_case=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 567,
   "id": "6bf9e92f",
   "metadata": {},
   "outputs": [],
   "source": [
    "ar = {}\n",
    "ar['answer-candidates'] = get_analysis_result('answer-candidates', 'train-v1.1')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "77993048",
   "metadata": {},
   "outputs": [],
   "source": [
    "ar = {}\n",
    "ar['answer-candidates'] = get_analysis_result('answer-candidates', 'TriviaQA-web-dev-from-MRQA')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bce1a2d9",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "7edfffb1",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "----\n",
      "ID: 5733be284776f41900661182\n",
      "Context 📗:  Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend \"Venite Ad Me Omnes\". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary.\n",
      "\n",
      "Question 💡: To whom did the Virgin Mary allegedly appear in 1858 in Lourdes France?\n",
      "Answer 🍎: [{'answer_start': 515, 'text': 'Saint Bernadette Soubirous'}]\n",
      "----\n",
      "----\n",
      "ID: 5733be284776f4190066117f\n",
      "Context 📗:  Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend \"Venite Ad Me Omnes\". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary.\n",
      "\n",
      "Question 💡: What is in front of the Notre Dame Main Building?\n",
      "Answer 🍎: [{'answer_start': 188, 'text': 'a copper statue of Christ'}]\n",
      "----\n",
      "----\n",
      "ID: 5733be284776f41900661180\n",
      "Context 📗:  Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend \"Venite Ad Me Omnes\". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary.\n",
      "\n",
      "Question 💡: The Basilica of the Sacred heart at Notre Dame is beside to which structure?\n",
      "Answer 🍎: [{'answer_start': 279, 'text': 'the Main Building'}]\n",
      "----\n",
      "----\n",
      "ID: 5733be284776f41900661181\n",
      "Context 📗:  Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend \"Venite Ad Me Omnes\". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary.\n",
      "\n",
      "Question 💡: What is the Grotto at Notre Dame?\n",
      "Answer 🍎: [{'answer_start': 381, 'text': 'a Marian place of prayer and reflection'}]\n",
      "----\n",
      "----\n",
      "ID: 5733be284776f4190066117e\n",
      "Context 📗:  Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend \"Venite Ad Me Omnes\". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary.\n",
      "\n",
      "Question 💡: What sits on top of the Main Building at Notre Dame?\n",
      "Answer 🍎: [{'answer_start': 92, 'text': 'a golden statue of the Virgin Mary'}]\n",
      "----\n"
     ]
    }
   ],
   "source": [
    "for i in range(5):\n",
    "    print_exqa_example(squad_train_examples[i])\n",
    "    _id = squad_train_examples[i].qas_id"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fe744c4b",
   "metadata": {},
   "source": [
    "# MC-QA"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "347c928b",
   "metadata": {},
   "outputs": [],
   "source": [
    "from utils_multiple_choice import MultipleChoiceDataset, Split, processors\n",
    "from transformers.data.data_collator import default_data_collator\n",
    "from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, Subset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "eb9d1093",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoTokenizer\n",
    "tokenizer = AutoTokenizer.from_pretrained(\"roberta-base\", use_fast=False)\n",
    "\n",
    "dataset = MultipleChoiceDataset(\n",
    "        data_dir=DATASET_DIR / \"mc-qa/RACE\",\n",
    "        tokenizer=tokenizer,\n",
    "        task=\"race\",\n",
    "        max_seq_length=512,\n",
    "        overwrite_cache=False,\n",
    "        mode=Split.dev,\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "504b65fe",
   "metadata": {},
   "outputs": [],
   "source": [
    "sampler = SequentialSampler(dataset)\n",
    "dataloader = DataLoader(\n",
    "    dataset,\n",
    "    batch_size=12,\n",
    "    sampler=sampler,\n",
    "    collate_fn=default_data_collator,\n",
    "    num_workers=0,\n",
    "    pin_memory=True,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "0719f0b0",
   "metadata": {},
   "outputs": [],
   "source": [
    "for d in dataloader:\n",
    "    break"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "6980d3ef",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(torch.Size([12, 4, 512]), torch.Tensor)"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "d[\"input_ids\"].shape, type(d[\"input_ids\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 585,
   "id": "6379b32a",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "dict_keys(['labels', 'input_ids', 'attention_mask'])"
      ]
     },
     "execution_count": 585,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "d.keys()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0ad12989",
   "metadata": {},
   "source": [
    "## Load Examples"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "07b7e519",
   "metadata": {},
   "outputs": [],
   "source": [
    "from utils_multiple_choice import  RaceProcessor, ReclorProcessor"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "e7447354",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "read files: 100%|██████████| 18728/18728 [00:13<00:00, 1437.68it/s]\n",
      "read files: 100%|██████████| 6409/6409 [00:04<00:00, 1423.82it/s]\n"
     ]
    }
   ],
   "source": [
    "race_processor = RaceProcessor()\n",
    "race_train_examples = race_processor.get_train_examples(race_dir)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "2703b27b",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "read files: 100%|██████████| 1045/1045 [00:00<00:00, 1842.92it/s]\n",
      "read files: 100%|██████████| 362/362 [00:00<00:00, 1895.83it/s]\n"
     ]
    }
   ],
   "source": [
    "race_processor = RaceProcessor()\n",
    "race_test_examples = race_processor.get_test_examples(race_dir)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "id": "c69d5348",
   "metadata": {},
   "outputs": [],
   "source": [
    "reclor_processor = ReclorProcessor()\n",
    "reclor_train_examples = reclor_processor.get_train_examples(reclor_dir)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 115,
   "id": "46707791",
   "metadata": {},
   "outputs": [],
   "source": [
    "reclor_test_examples = reclor_processor.get_test_examples(reclor_dir)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 116,
   "id": "72324dda",
   "metadata": {},
   "outputs": [],
   "source": [
    "# reclor_ids = [e.example_id + e.question + ''.join(e.endings) for e in reclor_train_examples]\n",
    "reclor_ids = [e.example_id  for e in reclor_test_examples]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 117,
   "id": "71c540e6",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1000, 1000)"
      ]
     },
     "execution_count": 117,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(reclor_ids), len(set(reclor_ids))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "40a381d3",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "dict_keys(['example_id', 'question', 'contexts', 'endings', 'label'])"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "race_train_examples[0].__dict__.keys()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 410,
   "id": "772cc5d3",
   "metadata": {},
   "outputs": [],
   "source": [
    "keys = [\n",
    "    'answer-context-ngram-overlap-per-sent',\n",
    "    'answer-context-similar-sent',\n",
    "    'answer-option-ngram-overlap',\n",
    "    'question-context-similar-sent',\n",
    "]\n",
    "ar = {}\n",
    "for k in keys:\n",
    "    ar[k] = get_analysis_result(k, 'reclor-train', task_type='mc-qa')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "a07bd0b2",
   "metadata": {},
   "outputs": [],
   "source": [
    "examples = race_train_examples"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "id": "be9d326b",
   "metadata": {},
   "outputs": [],
   "source": [
    "examples = race_test_examples"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 385,
   "id": "b9f8d507",
   "metadata": {},
   "outputs": [],
   "source": [
    "examples = reclor_train_examples"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 118,
   "id": "9e7651d5",
   "metadata": {},
   "outputs": [],
   "source": [
    "examples = reclor_test_examples"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6f7cebd6",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0f086834",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "id": "601db845",
   "metadata": {},
   "source": [
    "## Lexical Overlap\n",
    "### Processing data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 119,
   "id": "1ef1ac47",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 1000/1000 [00:34<00:00, 29.19it/s]\n"
     ]
    }
   ],
   "source": [
    "overlap_dict = {}\n",
    "data_name = 'reclor-test'\n",
    "\n",
    "for e in tqdm(examples):\n",
    "    if 'race' in data_name:\n",
    "        _id = e.example_id + e.question + ''.join(e.endings)\n",
    "    else:\n",
    "        _id = e.example_id\n",
    "    q = [t.text.lower() for t in nlp(e.question)]\n",
    "    c = [t.text.lower() for t in nlp(e.contexts[0])]\n",
    "    overlap = []\n",
    "    for end in e.endings:\n",
    "        o = nlp(end)\n",
    "        n = 0\n",
    "        for t in o:\n",
    "            t = t.text.lower()\n",
    "            if t in q or t in c:\n",
    "                n += 1\n",
    "        overlap.append(n / len(o))\n",
    "    overlap_dict[_id] = overlap"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "b68722f2",
   "metadata": {},
   "outputs": [],
   "source": [
    "ids = [e.example_id + e.question + ''.join(e.endings) for e in examples]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "d8790a7e",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(87866, 87865)"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(ids), len(set(ids))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5a089baa",
   "metadata": {},
   "source": [
    "### Save/Load overlap dict "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "9421da99",
   "metadata": {},
   "outputs": [],
   "source": [
    "# data_name = 'reclor-train'\n",
    "data_name = 'race-train'\n",
    "# data_name = 'race-test'\n",
    "# data_name = 'reclor-test'\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'option-overlap_{data_name}.json'\n",
    "# save_json(overlap_dict, file)\n",
    "overlap_dict = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 75,
   "id": "a06fea79",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(87865, 4934)"
      ]
     },
     "execution_count": 75,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(overlap_dict), len(examples)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "df832aa2",
   "metadata": {},
   "source": [
    "## Does Correct Option Have Max Lexical Overlap?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 76,
   "id": "56948e0a",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'race-train'"
      ]
     },
     "execution_count": 76,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data_name"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "id": "cf48db14",
   "metadata": {},
   "outputs": [],
   "source": [
    "true_is_max_lo = 0\n",
    "N = 0\n",
    "\n",
    "id2maxlo = {}\n",
    "\n",
    "id2maxlo2 = {}\n",
    "\n",
    "examples = race_train_examples\n",
    "\n",
    "for e in examples:\n",
    "    N += 1\n",
    "    if 'race' in data_name:\n",
    "        _id = e.example_id + e.question + ''.join(e.endings)\n",
    "    else:\n",
    "        _id = e.example_id\n",
    "    label = int(e.label)\n",
    "    overlap = overlap_dict[_id]\n",
    "    to = overlap[label]\n",
    "    io = []\n",
    "    for i in [0, 1, 2, 3]:\n",
    "        if i != label:\n",
    "            io.append(overlap[i])\n",
    "    m = max(io)\n",
    "    is_max_lo = 1\n",
    "    if m > to:\n",
    "        is_max_lo = 0\n",
    "    elif m == to:\n",
    "        is_max_lo = None\n",
    "    id2maxlo2[_id] = is_max_lo\n",
    "\n",
    "    is_max_lo = 1\n",
    "    if m >= to:\n",
    "        is_max_lo = 0\n",
    "    id2maxlo[_id] = is_max_lo"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2ef87403",
   "metadata": {},
   "outputs": [],
   "source": [
    "[]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 71,
   "id": "e0348ab4",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1020, 4934)"
      ]
     },
     "execution_count": 71,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sum(id2maxlo.values()), len(id2maxlo)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 79,
   "id": "fcbb95a7",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Counter({None: 23437, 1: 18691, 0: 45737})"
      ]
     },
     "execution_count": 79,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Counter(id2maxlo2.values())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bb565772",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 125,
   "id": "1790e58d",
   "metadata": {},
   "outputs": [],
   "source": [
    "# data_name = 'reclor-train'\n",
    "data_name = 'race-train'\n",
    "# data_name = 'race-test'\n",
    "# data_name = 'reclor-test'\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'correct-has-max-lexical-overlap_{data_name}.json'\n",
    "save_json(id2maxlo, file)\n",
    "id2maxlo = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "id": "37c3ad3f",
   "metadata": {},
   "outputs": [],
   "source": [
    "# data_name = 'reclor-train'\n",
    "# data_name = 'race-train'\n",
    "data_name = 'race-test'\n",
    "# data_name = 'reclor-test'\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'correct-has-max-lexical-overlap-with-none_{data_name}.json'\n",
    "save_json(id2maxlo2, file)\n",
    "# id2maxlo = load_json(file)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e93cdc45",
   "metadata": {},
   "source": [
    "## Vocab"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "dcd90979",
   "metadata": {},
   "source": [
    "### Tokenize and Construct Vocab"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "5af19a95",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'race-train'"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data_name"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 464,
   "id": "189a605b",
   "metadata": {},
   "outputs": [],
   "source": [
    "answers = [e.endings[int(e.label)] for e in examples]\n",
    "distractors = []\n",
    "labels = [0, 1, 2, 3]\n",
    "for e in examples:\n",
    "    for l in labels:\n",
    "        if l != int(e.label):\n",
    "            distractors.append(e.endings[l])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 465,
   "id": "0d75c508",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(87866, 263598)"
      ]
     },
     "execution_count": 465,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(answers), len(distractors)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 466,
   "id": "21b43e01",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 87866/87866 [05:17<00:00, 276.38it/s]\n",
      "100%|██████████| 263598/263598 [15:43<00:00, 279.45it/s]\n"
     ]
    }
   ],
   "source": [
    "vocab = defaultdict(dict)\n",
    "for a in tqdm(answers):\n",
    "    tokens = [t.text.lower() for t in nlp(a)]\n",
    "    for t in tokens:\n",
    "        vocab[t]['true'] = vocab[t].get('true', 0) + 1\n",
    "for d in tqdm(distractors):\n",
    "    tokens = [t.text.lower() for t in nlp(d)]\n",
    "    for t in tokens:\n",
    "        vocab[t]['false'] = vocab[t].get('false', 0) + 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 467,
   "id": "44aa4bea",
   "metadata": {},
   "outputs": [],
   "source": [
    "for v in vocab:\n",
    "    vocab[v]['prob'] = vocab[v].get('true', 0) / (vocab[v].get('true', 0) + vocab[v].get('false', 0))\n",
    "    vocab[v]['freq'] = vocab[v].get('true', 0) + vocab[v].get('false', 0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 468,
   "id": "d844dfcf",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "36104"
      ]
     },
     "execution_count": 468,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(vocab)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "id": "56bc6ca4",
   "metadata": {},
   "outputs": [],
   "source": [
    "for v in vocab:\n",
    "    p = vocab[v]['freq']\n",
    "    y = vocab[v]['prob']\n",
    "    z = (y - 0.25) / math.sqrt(0.25 * 0.75 / p)\n",
    "    vocab[v]['z'] = z"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "id": "ff32c56e",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'true': 7,\n",
       " 'false': 17,\n",
       " 'prob': 0.2916666666666667,\n",
       " 'freq': 24,\n",
       " 'z': 0.47140452079103184}"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vocab['famous']"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7e563e2b",
   "metadata": {},
   "source": [
    "### Save/Load Vocab"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "id": "b62c6abf",
   "metadata": {},
   "outputs": [],
   "source": [
    "data_name = 'race-train'\n",
    "# data_name = 'reclor-train'\n",
    "\n",
    "\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'option-vocab_{data_name}.json'\n",
    "# save_json(vocab, file)\n",
    "vocab = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "134a2e96",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "36104"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(vocab)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1e2140e3",
   "metadata": {},
   "source": [
    "### top k"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "7b47df3c",
   "metadata": {},
   "outputs": [],
   "source": [
    "vz = [(v, vocab[v]['z']) for v in vocab]\n",
    "sorted_vz = sorted(vz, key=lambda x: x[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "00bf9151",
   "metadata": {},
   "outputs": [],
   "source": [
    "topk = 1\n",
    "pos_word = [x[0] for x in sorted_vz[-topk:]]\n",
    "neg_word = [x[0] for x in sorted_vz[:topk]]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 134,
   "id": "e9acc3ff",
   "metadata": {},
   "outputs": [],
   "source": [
    "file = Path(analysis_dir) / 'mc-qa' / f'top1-word_{data_name}.json'\n",
    "save_json(pos_word, file)\n",
    "# pos_word = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "id": "f4ed0ce9",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "race-test\n",
      "['and']\n",
      "['only']\n"
     ]
    }
   ],
   "source": [
    "print(data_name)\n",
    "print(pos_word)\n",
    "print(neg_word)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 158,
   "id": "5b3c5012",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "race-train\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[(',', 9.320269439394076),\n",
       " ('something', 9.771763639228007),\n",
       " ('of', 9.77779021541762),\n",
       " ('objective', 10.014275524755362),\n",
       " ('might', 10.539303728279354),\n",
       " ('c', 13.529628504050919),\n",
       " ('b', 16.484449451145487),\n",
       " ('may', 20.66054087385846),\n",
       " ('above', 20.713908915129547),\n",
       " ('and', 23.645075672987726)]"
      ]
     },
     "execution_count": 158,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "print(data_name)\n",
    "sorted_vz[-10:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 162,
   "id": "9a32db8f",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "reclor-train\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[('aphids', 3.8729833462074166),\n",
       " ('player', 4.01188709901439),\n",
       " ('dinner', 4.016632088371219),\n",
       " ('stays', 4.242640687119285),\n",
       " ('not', 4.337094210761294),\n",
       " ('motive', 4.454276564173461),\n",
       " ('the', 4.948719535606995),\n",
       " ('an', 5.075982783427859),\n",
       " ('result', 5.293792820298356),\n",
       " ('a', 6.654926072139417)]"
      ]
     },
     "execution_count": 162,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "print(data_name)\n",
    "sorted_vz[-10:]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "666fbc97",
   "metadata": {},
   "source": [
    "### z threshold"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 283,
   "id": "c9df44cb",
   "metadata": {},
   "outputs": [],
   "source": [
    "thre = 2.33 # alfa = 1%\n",
    "# thre = 2.58 # alfa = 0.5%\n",
    "\n",
    "n_pos = 0\n",
    "pos_word = []\n",
    "n_neg = 0\n",
    "neg_word = []\n",
    "for v, z in vz:\n",
    "    if z >= thre:\n",
    "        # print(v, z)\n",
    "        n_pos += 1\n",
    "        pos_word.append(v)\n",
    "    if z <= -1 * thre:\n",
    "        n_neg += 1\n",
    "        neg_word.append(v)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 265,
   "id": "7f6adf4b",
   "metadata": {},
   "outputs": [],
   "source": [
    "qids = [e.example_id + e.question + ''.join(e.endings) for e in examples]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fff57f0d",
   "metadata": {},
   "source": [
    "### Tokenize options"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 139,
   "id": "d27659a7",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 1000/1000 [00:19<00:00, 51.28it/s]\n"
     ]
    }
   ],
   "source": [
    "# examples = race_train_examples\n",
    "# examples = reclor_train_examples\n",
    "examples = reclor_test_examples\n",
    "# examples = race_test_examples\n",
    "# data_name = 'reclor-train'\n",
    "data_name = 'reclor-test'\n",
    "# data_name = 'race-test'\n",
    "\n",
    "id2correct = {}\n",
    "id2incorrect = defaultdict(list)\n",
    "for e in tqdm(examples):\n",
    "    if 'race' in data_name:\n",
    "        _id = e.example_id + e.question + ''.join(e.endings)\n",
    "    else:\n",
    "        _id = e.example_id\n",
    "    for i, end in enumerate(e.endings):\n",
    "        o = nlp(end)\n",
    "        if i == int(e.label):\n",
    "            id2correct[_id] = [t.text.lower() for t in o]\n",
    "        else:\n",
    "            id2incorrect[_id].append([t.text.lower() for t in o])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "id": "54f2b1e4",
   "metadata": {},
   "outputs": [],
   "source": [
    "data_name = 'race-train'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "id": "c78e7623",
   "metadata": {},
   "outputs": [],
   "source": [
    "file = Path(analysis_dir) / 'mc-qa' / f'tokenized-correct_{data_name}.json'\n",
    "# save_json(id2correct, file)\n",
    "id2correct = load_json(file)\n",
    "\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'tokenized-incorrects_{data_name}.json'\n",
    "# save_json(id2incorrect, file)\n",
    "id2incorrect = load_json(file)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e342c7e9",
   "metadata": {},
   "source": [
    "### count biased examples"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 500,
   "id": "683ee299",
   "metadata": {},
   "outputs": [],
   "source": [
    "file = Path(analysis_dir) / 'mc-qa' / f'top50-words_race-train.json'\n",
    "pos_word = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "5c6e049a",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['and'], ['only'])"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pos_word, neg_word"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "0319a862",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'race-train'"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data_name"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "489e7479",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "id": "61edbc55",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 87866/87866 [00:00<00:00, 134472.86it/s]\n"
     ]
    }
   ],
   "source": [
    "data_name = 'race-train'\n",
    "examples = race_train_examples\n",
    "\n",
    "only_correct_has_pos = {}\n",
    "only_incorrect_has_neg = {}\n",
    "\n",
    "only_correct_has_pos2 = {}\n",
    "\n",
    "for e in tqdm(examples):\n",
    "    if 'race' in data_name:\n",
    "        _id = e.example_id + e.question + ''.join(e.endings)\n",
    "    else:\n",
    "        _id = e.example_id\n",
    "    correct_has_at_least_one_pos = False\n",
    "    correct_has_at_least_one_neg = False\n",
    "    incorrect_has_at_least_one_pos = False\n",
    "    incorrect_has_at_least_one_neg = False\n",
    "    correct = id2correct[_id]\n",
    "    for t in correct:\n",
    "        if t in pos_word:\n",
    "            correct_has_at_least_one_pos = True\n",
    "        if t in neg_word:\n",
    "            correct_has_at_least_one_neg = True\n",
    "    incorrects = id2incorrect[_id]\n",
    "    for o in incorrects:\n",
    "        for t in o:\n",
    "            if t in pos_word:\n",
    "                incorrect_has_at_least_one_pos = True\n",
    "            if t in neg_word:\n",
    "                incorrect_has_at_least_one_neg = True\n",
    "    if correct_has_at_least_one_pos and (not incorrect_has_at_least_one_pos):\n",
    "        only_correct_has_pos[_id] = 1\n",
    "    else:\n",
    "        only_correct_has_pos[_id] = 0\n",
    "\n",
    "    if (not correct_has_at_least_one_neg) and incorrect_has_at_least_one_neg:\n",
    "        only_incorrect_has_neg[_id] = 1\n",
    "    else:\n",
    "        only_incorrect_has_neg[_id] = 0\n",
    "        \n",
    "    if correct_has_at_least_one_pos and (not incorrect_has_at_least_one_pos):\n",
    "        only_correct_has_pos2[_id] = 1\n",
    "    elif not correct_has_at_least_one_pos and incorrect_has_at_least_one_pos:\n",
    "        only_correct_has_pos2[_id] = 0\n",
    "    else:\n",
    "        only_correct_has_pos2[_id] = None"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3b41748a",
   "metadata": {},
   "source": [
    "### Save & Load"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 148,
   "id": "60d79562",
   "metadata": {},
   "outputs": [],
   "source": [
    "# data_name = 'reclor-train'\n",
    "# data_name = 'race-train'\n",
    "# data_name = 'race-test'\n",
    "data_name = 'reclor-test'\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'only-correct-has-top1-word_{data_name}.json'\n",
    "save_json(only_correct_has_pos, file)\n",
    "# only_correct_has_pos = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "id": "c0bc2f35",
   "metadata": {},
   "outputs": [],
   "source": [
    "# data_name = 'reclor-train'\n",
    "data_name = 'race-train'\n",
    "# data_name = 'race-test'\n",
    "# data_name = 'reclor-test'\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'only-correct-has-top1-word-with-none_{data_name}.json'\n",
    "save_json(only_correct_has_pos2, file)\n",
    "# only_correct_has_pos = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 103,
   "id": "e3e9c314",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "4934"
      ]
     },
     "execution_count": 103,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(only_correct_has_pos)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8f5dfdd7",
   "metadata": {},
   "source": [
    "## plot vocab"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "id": "45b704a7",
   "metadata": {},
   "outputs": [],
   "source": [
    "# data_name = 'race-train'\n",
    "data_name = 'reclor-train'\n",
    "\n",
    "\n",
    "file = Path(analysis_dir) / 'mc-qa' / f'option-vocab_{data_name}.json'\n",
    "# save_json(vocab, file)\n",
    "vocab = load_json(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "id": "21c263a5",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "reclor-train\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f547c5258d0>"
      ]
     },
     "execution_count": 44,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "print(data_name)\n",
    "\n",
    "Xhigh = []\n",
    "Yhigh = []\n",
    "\n",
    "Xmid = []\n",
    "Ymid = []\n",
    "\n",
    "Xlow = []\n",
    "Ylow = []\n",
    "\n",
    "for v in vocab:\n",
    "    p = vocab[v]['freq']\n",
    "    x = math.log(p) / math.log(10)\n",
    "    y = vocab[v]['prob']\n",
    "    high_boundary = 2.33 * math.sqrt(0.25 * 0.75 / 10 ** x) + 0.25\n",
    "    low_boundary = -2.33 * math.sqrt(0.25 * 0.75 / 10 ** x) + 0.25\n",
    "    if y > high_boundary:\n",
    "        Xhigh.append(x)\n",
    "        Yhigh.append(y)\n",
    "    # elif y < low_boundary:\n",
    "    elif False:\n",
    "        Xlow.append(x)\n",
    "        Ylow.append(y)\n",
    "    else:\n",
    "        Xmid.append(x)\n",
    "        Ymid.append(y)\n",
    "\n",
    "    # z = (y - 0.25) / math.sqrt(0.25 * 0.75 / p)\n",
    "    # Z[v] = z\n",
    "size = 2\n",
    "plt.scatter(Xhigh, Yhigh, marker='.', color='C0', s=size)\n",
    "plt.scatter(Xmid, Ymid, marker='.', color='gray', s=size)\n",
    "# plt.scatter(Xlow, Ylow, marker='.', color='C1', s=size)\n",
    "\n",
    "\n",
    "Xplot = [0.1 * i for i in range(55)]\n",
    "\n",
    "Yplot_upper = [ 2.33 * math.sqrt(0.25 * 0.75 / 10 ** x) + 0.25 for x in Xplot]\n",
    "Yplot_lower = [ -2.33 * math.sqrt(0.25 * 0.75 / 10 ** x) + 0.25 for x in Xplot]\n",
    "\n",
    "plt.plot(Xplot, Yplot_upper, color='C0', label='α =1%')\n",
    "# plt.plot(Xplot, Yplot_lower, color='C1')\n",
    "\n",
    "plt.xticks([0, 1, 2, 3, 4, 5], ['$10^0$', '$10^1$', '$10^2$', '$10^3$', '$10^4$', '$10^5$'], size=11)\n",
    "plt.ylim(-0.1, 1.1)\n",
    "plt.xlabel('Frequency of word $x_i$ in options', size=13)\n",
    "plt.ylabel('$p(y|x_i)$', size=13)\n",
    "plt.legend(fontsize=13)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 82,
   "id": "9cd30d99",
   "metadata": {},
   "outputs": [],
   "source": [
    "sortedZ = sorted(Z.items(), key=lambda x: x[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 92,
   "id": "dbeaba56",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[('still', 8.371584807348908),\n",
       " ('about', 8.468725144606875),\n",
       " ('different', 8.633553233971561),\n",
       " ('or', 8.825894609202274),\n",
       " ('life', 8.866112058847547),\n",
       " ('positive', 8.91132788679007),\n",
       " ('probably', 8.919508528991598),\n",
       " ('mind', 8.93941843982621),\n",
       " ('both', 9.15177209629895),\n",
       " ('some', 9.262899818139202),\n",
       " (',', 9.320269439394076),\n",
       " ('something', 9.771763639228007),\n",
       " ('of', 9.77779021541762),\n",
       " ('objective', 10.014275524755362),\n",
       " ('might', 10.539303728279354),\n",
       " ('c', 13.529628504050919),\n",
       " ('b', 16.484449451145487),\n",
       " ('may', 20.66054087385846),\n",
       " ('above', 20.713908915129547),\n",
       " ('and', 23.645075672987726)]"
      ]
     },
     "execution_count": 92,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sortedZ[-20:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "id": "070950ac",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'true': 316, 'false': 527, 'prob': 0.3748517200474496, 'freq': 843}"
      ]
     },
     "execution_count": 93,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vocab['still']"
   ]
  }
 ],
 "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": "3.6.8"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {
    "height": "calc(100% - 180px)",
    "left": "10px",
    "top": "150px",
    "width": "260px"
   },
   "toc_section_display": true,
   "toc_window_display": true
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}