5_NLP.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import nltk\n",
    "import numpy as np\n",
    "import tensorflow as tf"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# NLP Tasks"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Text classification: Sentiment analysis, Spam filters, Author classification...\n",
    "\n",
    "Information retrieval, Question answering, Reading comprehension\n",
    "\n",
    "Summarization\n",
    "\n",
    "Translation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Как да представим текст като вектор от числа?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['Втората световна война е глобална война, продължила между 1939 и 1945 година.',\n",
       " 'Голямото мнозинство от страните в света, включително всички Велики сили, в един или друг момент се включват във войната, присъединявайки се към един от двата противопоставящи се военни съюза – Съюзниците и Оста.',\n",
       " 'Достига се до състояние на тотална война, в която пряко участват над 100 милиона души от повече от 30 страни.',\n",
       " 'Основните участници хвърлят целия си стопански, промишлен и научен потенциал във военното усилие, размивайки границите между граждански и военни ресурси.']"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Еми оказва се че тъпия подход е сравнително ефективен:\n",
    "# Bag of words\n",
    "\n",
    "example_text = [\n",
    "    \"Втората световна война е глобална война, продължила между 1939 и 1945 година.\",\n",
    "    \"Голямото мнозинство от страните в света, включително всички Велики сили, в един или друг момент се включват във войната, присъединявайки се към един от двата противопоставящи се военни съюза – Съюзниците и Оста.\",\n",
    "    \"Достига се до състояние на тотална война, в която пряко участват над 100 милиона души от повече от 30 страни.\",\n",
    "    \"Основните участници хвърлят целия си стопански, промишлен и научен потенциал във военното усилие, размивайки границите между граждански и военни ресурси.\"\n",
    "]\n",
    "example_text"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['втората световна война е глобална война, продължила между 1939 и 1945 година.',\n",
       " 'голямото мнозинство от страните в света, включително всички велики сили, в един или друг момент се включват във войната, присъединявайки се към един от двата противопоставящи се военни съюза – съюзниците и оста.',\n",
       " 'достига се до състояние на тотална война, в която пряко участват над 100 милиона души от повече от 30 страни.',\n",
       " 'основните участници хвърлят целия си стопански, промишлен и научен потенциал във военното усилие, размивайки границите между граждански и военни ресурси.']"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "example_text = [x.lower() for x in example_text]\n",
    "example_text"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[['втората',\n",
       "  'световна',\n",
       "  'война',\n",
       "  'е',\n",
       "  'глобална',\n",
       "  'война',\n",
       "  ',',\n",
       "  'продължила',\n",
       "  'между',\n",
       "  '1939',\n",
       "  'и',\n",
       "  '1945',\n",
       "  'година',\n",
       "  '.'],\n",
       " ['голямото',\n",
       "  'мнозинство',\n",
       "  'от',\n",
       "  'страните',\n",
       "  'в',\n",
       "  'света',\n",
       "  ',',\n",
       "  'включително',\n",
       "  'всички',\n",
       "  'велики',\n",
       "  'сили',\n",
       "  ',',\n",
       "  'в',\n",
       "  'един',\n",
       "  'или',\n",
       "  'друг',\n",
       "  'момент',\n",
       "  'се',\n",
       "  'включват',\n",
       "  'във',\n",
       "  'войната',\n",
       "  ',',\n",
       "  'присъединявайки',\n",
       "  'се',\n",
       "  'към',\n",
       "  'един',\n",
       "  'от',\n",
       "  'двата',\n",
       "  'противопоставящи',\n",
       "  'се',\n",
       "  'военни',\n",
       "  'съюза',\n",
       "  '–',\n",
       "  'съюзниците',\n",
       "  'и',\n",
       "  'оста',\n",
       "  '.'],\n",
       " ['достига',\n",
       "  'се',\n",
       "  'до',\n",
       "  'състояние',\n",
       "  'на',\n",
       "  'тотална',\n",
       "  'война',\n",
       "  ',',\n",
       "  'в',\n",
       "  'която',\n",
       "  'пряко',\n",
       "  'участват',\n",
       "  'над',\n",
       "  '100',\n",
       "  'милиона',\n",
       "  'души',\n",
       "  'от',\n",
       "  'повече',\n",
       "  'от',\n",
       "  '30',\n",
       "  'страни',\n",
       "  '.'],\n",
       " ['основните',\n",
       "  'участници',\n",
       "  'хвърлят',\n",
       "  'целия',\n",
       "  'си',\n",
       "  'стопански',\n",
       "  ',',\n",
       "  'промишлен',\n",
       "  'и',\n",
       "  'научен',\n",
       "  'потенциал',\n",
       "  'във',\n",
       "  'военното',\n",
       "  'усилие',\n",
       "  ',',\n",
       "  'размивайки',\n",
       "  'границите',\n",
       "  'между',\n",
       "  'граждански',\n",
       "  'и',\n",
       "  'военни',\n",
       "  'ресурси',\n",
       "  '.']]"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "example_text = [nltk.word_tokenize(x) for x in example_text]\n",
    "example_text"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 2., 0., 1., 0.,\n",
       "        0., 1., 0., 0., 0., 1., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0.,\n",
       "        0., 0., 1., 0., 0., 1., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 1., 0., 0.],\n",
       "       [0., 0., 0., 1., 2., 0., 0., 0., 1., 1., 3., 0., 0., 1., 0., 0.,\n",
       "        1., 1., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 1., 0., 1., 0.,\n",
       "        0., 0., 0., 3., 1., 1., 0., 0., 2., 1., 1., 0., 0., 1., 0., 0.,\n",
       "        1., 2., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 1., 1., 1., 1.,\n",
       "        1., 1., 0., 1., 1., 1.],\n",
       "       [0., 0., 1., 0., 1., 1., 0., 1., 0., 0., 1., 0., 1., 0., 0., 1.,\n",
       "        0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 1., 0., 1., 0., 0.,\n",
       "        1., 1., 1., 1., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 1., 1.,\n",
       "        0., 2., 0., 0., 1., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n",
       "        0., 0., 0., 1., 0., 0.],\n",
       "       [1., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 1., 0.,\n",
       "        0., 2., 1., 0., 0., 0., 1., 0., 1., 1., 1., 0., 0., 0., 0., 1.,\n",
       "        0., 0., 0., 2., 0., 0., 1., 0., 0., 1., 0., 0., 1., 0., 0., 0.,\n",
       "        0., 0., 0., 1., 0., 0., 0., 0., 1., 1., 0., 1., 0., 0., 0., 0.,\n",
       "        0., 0., 1., 1., 1., 0.]])"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Ако ни се занимава можем да махнем препинателни знаци и стоп думи, но няма да го правим\n",
    "\n",
    "def vectorize(data):\n",
    "    all_words = set()\n",
    "    for x in data:\n",
    "        for y in x:\n",
    "            all_words.add(y)\n",
    "    word_to_num = {w:i for i, w in enumerate(all_words)}\n",
    "    ans = np.zeros((len(data), len(all_words)))\n",
    "    for i, x in enumerate(data):\n",
    "        for y in x:\n",
    "            ans[i, word_to_num[y]] += 1\n",
    "    return ans\n",
    "\n",
    "vectorize(example_text)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Недостатъци:\n",
    "# Дебело представяне (като нищо векторите ще станат с размер 50к)\n",
    "# Губи се реда на думите (което всъщност се оказва че не е чак толкова важно)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.        , 1.38629436, 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 1.38629436, 0.        , 0.69314718,\n",
       "        0.        , 0.        , 0.28768207, 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 0.        , 1.38629436, 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 1.38629436, 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        1.38629436, 0.        , 0.        , 1.38629436, 0.        ,\n",
       "        0.        , 0.        , 0.        , 1.38629436, 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ],\n",
       "       [0.        , 0.        , 0.        , 1.38629436, 1.38629436,\n",
       "        0.        , 0.        , 0.        , 1.38629436, 1.38629436,\n",
       "        2.07944154, 0.        , 0.        , 1.38629436, 0.        ,\n",
       "        0.        , 1.38629436, 0.28768207, 0.        , 0.        ,\n",
       "        1.38629436, 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 1.38629436, 0.        ,\n",
       "        1.38629436, 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 1.38629436, 0.        , 0.        ,\n",
       "        2.77258872, 0.69314718, 1.38629436, 0.        , 0.        ,\n",
       "        1.38629436, 0.        , 0.        , 1.38629436, 1.38629436,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        1.38629436, 0.        , 0.        , 0.        , 0.        ,\n",
       "        1.38629436, 1.38629436, 1.38629436, 1.38629436, 1.38629436,\n",
       "        1.38629436, 0.        , 0.        , 0.69314718, 1.38629436],\n",
       "       [0.        , 0.        , 1.38629436, 0.        , 0.69314718,\n",
       "        1.38629436, 0.        , 1.38629436, 0.        , 0.        ,\n",
       "        0.69314718, 0.        , 0.69314718, 0.        , 0.        ,\n",
       "        1.38629436, 0.        , 0.        , 0.        , 1.38629436,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 1.38629436, 0.        , 1.38629436,\n",
       "        0.        , 0.        , 1.38629436, 1.38629436, 1.38629436,\n",
       "        0.        , 0.        , 0.        , 0.        , 1.38629436,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 1.38629436, 0.        , 1.38629436,\n",
       "        0.        , 0.        , 1.38629436, 0.        , 1.38629436,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ],\n",
       "       [1.38629436, 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.69314718,\n",
       "        0.        , 0.        , 0.57536414, 1.38629436, 0.        ,\n",
       "        0.        , 0.        , 1.38629436, 0.        , 1.38629436,\n",
       "        1.38629436, 1.38629436, 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 0.        , 0.        , 0.        ,\n",
       "        0.        , 0.        , 0.        , 1.38629436, 0.        ,\n",
       "        0.        , 0.69314718, 0.        , 0.        , 1.38629436,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 1.38629436, 0.        , 1.38629436,\n",
       "        0.        , 0.        , 0.        , 0.        , 0.        ,\n",
       "        0.        , 1.38629436, 0.        , 0.69314718, 0.        ]])"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Едно често използвано подобрение - TFIDF\n",
    "# По-редките думи са по-значими\n",
    "\n",
    "def tfidf(data):\n",
    "    all_words = set()\n",
    "    for x in data:\n",
    "        for y in x:\n",
    "            all_words.add(y)\n",
    "    word_to_num = {w:i for i, w in enumerate(all_words)}\n",
    "    \n",
    "    ans = np.zeros((len(data), len(all_words)))\n",
    "    df = np.zeros(len(all_words))\n",
    "    for i, x in enumerate(data):\n",
    "        for y in x:\n",
    "            num = word_to_num[y]\n",
    "            ans[i, num] += 1\n",
    "            if ans[i, num]==1:\n",
    "                df[num] += 1\n",
    "    ans = ans * np.log(len(data) / df)\n",
    "    return ans\n",
    "\n",
    "tfidf(example_text)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Естествено, по-полезно е с повече документи"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>text</th>\n",
       "      <th>author</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>id26305</th>\n",
       "      <td>This process, however, afforded me no means of...</td>\n",
       "      <td>EAP</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id17569</th>\n",
       "      <td>It never once occurred to me that the fumbling...</td>\n",
       "      <td>HPL</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id11008</th>\n",
       "      <td>In his left hand was a gold snuff box, from wh...</td>\n",
       "      <td>EAP</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id27763</th>\n",
       "      <td>How lovely is spring As we looked from Windsor...</td>\n",
       "      <td>MWS</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id12958</th>\n",
       "      <td>Finding nothing else, not even gold, the Super...</td>\n",
       "      <td>HPL</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id17718</th>\n",
       "      <td>I could have fancied, while I looked at it, th...</td>\n",
       "      <td>EAP</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id08973</th>\n",
       "      <td>The lids clenched themselves together as if in...</td>\n",
       "      <td>EAP</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id05267</th>\n",
       "      <td>Mais il faut agir that is to say, a Frenchman ...</td>\n",
       "      <td>EAP</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id17513</th>\n",
       "      <td>For an item of news like this, it strikes us i...</td>\n",
       "      <td>EAP</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>id00393</th>\n",
       "      <td>He laid a gnarled claw on my shoulder, and it ...</td>\n",
       "      <td>HPL</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>19579 rows × 2 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                      text author\n",
       "id                                                               \n",
       "id26305  This process, however, afforded me no means of...    EAP\n",
       "id17569  It never once occurred to me that the fumbling...    HPL\n",
       "id11008  In his left hand was a gold snuff box, from wh...    EAP\n",
       "id27763  How lovely is spring As we looked from Windsor...    MWS\n",
       "id12958  Finding nothing else, not even gold, the Super...    HPL\n",
       "...                                                    ...    ...\n",
       "id17718  I could have fancied, while I looked at it, th...    EAP\n",
       "id08973  The lids clenched themselves together as if in...    EAP\n",
       "id05267  Mais il faut agir that is to say, a Frenchman ...    EAP\n",
       "id17513  For an item of news like this, it strikes us i...    EAP\n",
       "id00393  He laid a gnarled claw on my shoulder, and it ...    HPL\n",
       "\n",
       "[19579 rows x 2 columns]"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import pandas as pd\n",
    "pd.options.display.max_rows = 10\n",
    "\n",
    "train = pd.read_csv('./train.csv', index_col='id')\n",
    "train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<19579x25068 sparse matrix of type '<class 'numpy.float64'>'\n",
       "\twith 429602 stored elements in Compressed Sparse Row format>"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.feature_extraction.text import TfidfVectorizer\n",
    "\n",
    "vectorizer = TfidfVectorizer()\n",
    "x_all = vectorizer.fit_transform(train.text)\n",
    "x_all"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "matrix([[0., 0., 0., ..., 0., 0., 0.],\n",
       "        [0., 0., 0., ..., 0., 0., 0.],\n",
       "        [0., 0., 0., ..., 0., 0., 0.],\n",
       "        ...,\n",
       "        [0., 0., 0., ..., 0., 0., 0.],\n",
       "        [0., 0., 0., ..., 0., 0., 0.],\n",
       "        [0., 0., 0., ..., 0., 0., 0.]])"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x_all = x_all/np.sqrt(np.sum(x_all.toarray()**2, axis=-1)).reshape(-1, 1)\n",
    "x_all"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "x_train, x_val, y_train, y_val = train_test_split(x_all, train.author, test_size=0.2, random_state=42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((15663, 25068), (3916, 25068))"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x_train.shape, x_val.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "c:\\users\\iddilov\\appdata\\local\\programs\\python\\python37\\lib\\site-packages\\sklearn\\linear_model\\_logistic.py:940: ConvergenceWarning: lbfgs failed to converge (status=1):\n",
      "STOP: TOTAL NO. of ITERATIONS REACHED LIMIT.\n",
      "\n",
      "Increase the number of iterations (max_iter) or scale the data as shown in:\n",
      "    https://scikit-learn.org/stable/modules/preprocessing.html\n",
      "Please also refer to the documentation for alternative solver options:\n",
      "    https://scikit-learn.org/stable/modules/linear_model.html#logistic-regression\n",
      "  extra_warning_msg=_LOGISTIC_SOLVER_CONVERGENCE_MSG)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,\n",
       "                   intercept_scaling=1, l1_ratio=None, max_iter=100,\n",
       "                   multi_class='auto', n_jobs=None, penalty='l2',\n",
       "                   random_state=None, solver='lbfgs', tol=0.0001, verbose=0,\n",
       "                   warm_start=False)"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.linear_model import LogisticRegression\n",
    "\n",
    "lr = LogisticRegression()\n",
    "lr.fit(x_train, y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train log_loss = 0.3982996394590997\n",
      "log_loss = 0.5410314214059891\n",
      "accuracy = 0.8120531154239019\n"
     ]
    }
   ],
   "source": [
    "from sklearn import metrics\n",
    "\n",
    "y_pred = lr.predict_proba(x_train)\n",
    "print(f\"train log_loss = {metrics.log_loss(y_train, y_pred)}\")\n",
    "y_pred = lr.predict_proba(x_val)\n",
    "print(f\"log_loss = {metrics.log_loss(y_val, y_pred)}\")\n",
    "y_pred = lr.predict(x_val)\n",
    "print(f\"accuracy = {metrics.accuracy_score(y_val, y_pred)}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "MultinomialNB(alpha=1.0, class_prior=None, fit_prior=True)"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.naive_bayes import MultinomialNB\n",
    "\n",
    "nb = MultinomialNB()\n",
    "nb.fit(x_train, y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train log_loss = 0.4809467539837631\n",
      "log_loss = 0.6104175660117682\n",
      "accuracy = 0.8117977528089888\n"
     ]
    }
   ],
   "source": [
    "from sklearn import metrics\n",
    "\n",
    "y_pred = nb.predict_proba(x_train)\n",
    "print(f\"train log_loss = {metrics.log_loss(y_train, y_pred)}\")\n",
    "y_pred = nb.predict_proba(x_val)\n",
    "print(f\"log_loss = {metrics.log_loss(y_val, y_pred)}\")\n",
    "y_pred = nb.predict(x_val)\n",
    "print(f\"accuracy = {metrics.accuracy_score(y_val, y_pred)}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[1., 0., 0.],\n",
       "       [0., 1., 0.],\n",
       "       [0., 1., 0.],\n",
       "       [0., 1., 0.],\n",
       "       [0., 0., 1.]])"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "to_num = {'EAP':0, 'MWS':1, 'HPL':2}\n",
    "\n",
    "y_train = np.asarray([np.eye(3)[to_num[y]] for y in y_train])\n",
    "y_val = np.asarray([np.eye(3)[to_num[y]] for y in y_val])\n",
    "y_train[:5]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "dense (Dense)                (None, 32)                802208    \n",
      "_________________________________________________________________\n",
      "dropout (Dropout)            (None, 32)                0         \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 32)                1056      \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 32)                0         \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 3)                 99        \n",
      "=================================================================\n",
      "Total params: 803,363\n",
      "Trainable params: 803,363\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = tf.keras.Sequential([\n",
    "    tf.keras.layers.Dense(32, input_shape=(x_train.shape[-1],), activation='relu'),\n",
    "    tf.keras.layers.Dropout(0.5),\n",
    "    tf.keras.layers.Dense(32, activation='relu'),\n",
    "    tf.keras.layers.Dropout(0.3),\n",
    "    tf.keras.layers.Dense(3, activation='softmax')\n",
    "])\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 15663 samples, validate on 3916 samples\n",
      "Epoch 1/10\n",
      "15663/15663 [==============================] - 3s 164us/sample - loss: 1.0804 - categorical_crossentropy: 1.0804 - accuracy: 0.4053 - val_loss: 1.0422 - val_categorical_crossentropy: 1.0422 - val_accuracy: 0.4053\n",
      "Epoch 2/10\n",
      "15663/15663 [==============================] - 2s 111us/sample - loss: 0.9812 - categorical_crossentropy: 0.9812 - accuracy: 0.5517 - val_loss: 0.8881 - val_categorical_crossentropy: 0.8881 - val_accuracy: 0.7104\n",
      "Epoch 3/10\n",
      "15663/15663 [==============================] - 2s 99us/sample - loss: 0.7828 - categorical_crossentropy: 0.7828 - accuracy: 0.7523 - val_loss: 0.6714 - val_categorical_crossentropy: 0.6714 - val_accuracy: 0.8184\n",
      "Epoch 4/10\n",
      "15663/15663 [==============================] - 2s 99us/sample - loss: 0.5651 - categorical_crossentropy: 0.5651 - accuracy: 0.8399 - val_loss: 0.5156 - val_categorical_crossentropy: 0.5156 - val_accuracy: 0.8335\n",
      "Epoch 5/10\n",
      "15663/15663 [==============================] - 2s 98us/sample - loss: 0.4088 - categorical_crossentropy: 0.4088 - accuracy: 0.8815 - val_loss: 0.4420 - val_categorical_crossentropy: 0.4420 - val_accuracy: 0.8389\n",
      "Epoch 6/10\n",
      "15663/15663 [==============================] - 2s 99us/sample - loss: 0.3116 - categorical_crossentropy: 0.3116 - accuracy: 0.9125 - val_loss: 0.4110 - val_categorical_crossentropy: 0.4110 - val_accuracy: 0.8407\n",
      "Epoch 7/10\n",
      "15663/15663 [==============================] - 2s 99us/sample - loss: 0.2501 - categorical_crossentropy: 0.2501 - accuracy: 0.9302 - val_loss: 0.3978 - val_categorical_crossentropy: 0.3978 - val_accuracy: 0.8414\n",
      "Epoch 8/10\n",
      "15663/15663 [==============================] - 2s 98us/sample - loss: 0.2039 - categorical_crossentropy: 0.2039 - accuracy: 0.9445 - val_loss: 0.3954 - val_categorical_crossentropy: 0.3954 - val_accuracy: 0.8407\n",
      "Epoch 9/10\n",
      "15663/15663 [==============================] - 2s 101us/sample - loss: 0.1671 - categorical_crossentropy: 0.1671 - accuracy: 0.9558 - val_loss: 0.4005 - val_categorical_crossentropy: 0.4005 - val_accuracy: 0.8422\n",
      "Epoch 10/10\n",
      "15663/15663 [==============================] - 2s 97us/sample - loss: 0.1417 - categorical_crossentropy: 0.1417 - accuracy: 0.9628 - val_loss: 0.4091 - val_categorical_crossentropy: 0.4091 - val_accuracy: 0.8396\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x1f1c178a908>"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.compile(tf.keras.optimizers.Adam(1e-3), loss='categorical_crossentropy', metrics = ['categorical_crossentropy', 'accuracy'])\n",
    "\n",
    "model.fit(x_train, y_train, batch_size=512, epochs=10, validation_data=(x_val, y_val))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Както виждаме получаваме по-добър резултат с малко по-дълбока архитектура"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Няколко думи за различен preprocessing\n",
    "# Stemming\n",
    "# Lemmatization\n",
    "# NE tagging\n",
    "# POS tagging"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Word embeddings"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[['This',\n",
       "  'process',\n",
       "  ',',\n",
       "  'however',\n",
       "  ',',\n",
       "  'afforded',\n",
       "  'me',\n",
       "  'no',\n",
       "  'means',\n",
       "  'of',\n",
       "  'ascertaining',\n",
       "  'the',\n",
       "  'dimensions',\n",
       "  'of',\n",
       "  'my',\n",
       "  'dungeon',\n",
       "  ';',\n",
       "  'as',\n",
       "  'I',\n",
       "  'might',\n",
       "  'make',\n",
       "  'its',\n",
       "  'circuit',\n",
       "  ',',\n",
       "  'and',\n",
       "  'return',\n",
       "  'to',\n",
       "  'the',\n",
       "  'point',\n",
       "  'whence',\n",
       "  'I',\n",
       "  'set',\n",
       "  'out',\n",
       "  ',',\n",
       "  'without',\n",
       "  'being',\n",
       "  'aware',\n",
       "  'of',\n",
       "  'the',\n",
       "  'fact',\n",
       "  ';',\n",
       "  'so',\n",
       "  'perfectly',\n",
       "  'uniform',\n",
       "  'seemed',\n",
       "  'the',\n",
       "  'wall',\n",
       "  '.'],\n",
       " ['It',\n",
       "  'never',\n",
       "  'once',\n",
       "  'occurred',\n",
       "  'to',\n",
       "  'me',\n",
       "  'that',\n",
       "  'the',\n",
       "  'fumbling',\n",
       "  'might',\n",
       "  'be',\n",
       "  'a',\n",
       "  'mere',\n",
       "  'mistake',\n",
       "  '.']]"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "word_data = [nltk.word_tokenize(x) for x in train.text]\n",
    "word_data[:2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[24218, 20295, 13975, 6312, 13975]\n",
      "['This', 'process', ',', 'however', ',']\n"
     ]
    }
   ],
   "source": [
    "def get_num_data(data):\n",
    "    num2word = set()\n",
    "    for x in data:\n",
    "        for y in x:\n",
    "            num2word.add(y)\n",
    "    num2word = list(num2word)\n",
    "    word2num = {w:i for i,w in enumerate(num2word)}\n",
    "    new_data = []\n",
    "    for x in data:\n",
    "        new_data.append([])\n",
    "        for y in x:\n",
    "            new_data[-1].append(word2num[y])\n",
    "    return new_data, num2word, word2num\n",
    "\n",
    "num_data, num2word, word2num = get_num_data(word_data)\n",
    "print(num_data[0][:5])\n",
    "print([num2word[x] for x in num_data[0][:5]])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "27752"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(num2word)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [],
   "source": [
    "skip = 3\n",
    "def prepare_training_data(data):\n",
    "    x, y = [], []\n",
    "    for d in data:\n",
    "        for i in range(skip, len(d)-skip):\n",
    "            x.append(d[i-skip:i] + d[i+1: i+skip+1])\n",
    "            y.append(d[i])\n",
    "    return np.array(x), np.array(y)\n",
    "\n",
    "x_embed, y_embed = prepare_training_data(num_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "['This', 'process', ',', ',', 'afforded', 'me']\n",
      "however\n"
     ]
    }
   ],
   "source": [
    "print([num2word[x] for x in x_embed[0]])\n",
    "print(num2word[y_embed[0]])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_1\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "embedding (Embedding)        (None, 6, 64)             1776128   \n",
      "_________________________________________________________________\n",
      "global_average_pooling1d (Gl (None, 64)                0         \n",
      "_________________________________________________________________\n",
      "dropout_2 (Dropout)          (None, 64)                0         \n",
      "_________________________________________________________________\n",
      "dense_3 (Dense)              (None, 27752)             1803880   \n",
      "=================================================================\n",
      "Total params: 3,580,008\n",
      "Trainable params: 3,580,008\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "# CBOW model\n",
    "\n",
    "vocab_size = len(num2word)\n",
    "embedding_dim = 64\n",
    "\n",
    "embedding_model = tf.keras.Sequential([\n",
    "    tf.keras.layers.Embedding(vocab_size, embedding_dim, input_shape=(2*skip,)),\n",
    "    tf.keras.layers.GlobalAveragePooling1D(),\n",
    "    tf.keras.layers.Dropout(0.2),\n",
    "    tf.keras.layers.Dense(vocab_size)\n",
    "])\n",
    "\n",
    "embedding_model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "class PrintSim(tf.keras.callbacks.Callback):\n",
    "    def on_epoch_end(self, epoch, logs=None):\n",
    "        words = [\"the\", \"animal\", \"me\", \"like\", \"love\", \"ship\", \"walk\"]\n",
    "        weights = self.model.get_layer(index=0).get_weights()[0]\n",
    "        for w in words:\n",
    "            num = word2num[w]\n",
    "            sim = np.sum((weights - weights[num])**2, axis=-1)\n",
    "            closest = sim.argsort()[:10]\n",
    "            print(f\"closest to '{w}' are: {[num2word[x] for x in closest]}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "def loss(labels, logits):\n",
    "    return tf.keras.losses.sparse_categorical_crossentropy(labels, logits, from_logits=True)\n",
    "\n",
    "embedding_model.compile(optimizer=tf.keras.optimizers.Adam(), loss=loss)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 429957 samples, validate on 47773 samples\n",
      "Epoch 1/5\n",
      "429312/429957 [============================>.] - ETA: 0s - loss: 7.1949closest to 'the' are: ['the', 'a', 'his', 'its', 'an', '.', 'my', 'any', 'our', 'this']\n",
      "closest to 'animal' are: ['animal', 'large', 'covered', 'Automaton', 'feeble', 'interior', 'flower', 'presence', 'rain', 'evidence']\n",
      "closest to 'me' are: ['me', 'be', 'him', 'enabled', 'recall', 'she', 'paid', 'they', 'described', 'Do']\n",
      "closest to 'like' are: ['like', 'Automaton', 'wondered', 'metal', 'remain', 'open', 'reached', 'opening', 'called', 'flower']\n",
      "closest to 'love' are: ['love', 'mouth', 'Lollipop', 'loved', 'entire', 'rather', 'east', 'purpose', 'misery', 'bed']\n",
      "closest to 'ship' are: ['ship', 'green', 'waters', 'breast', 'Greek', 'parts', 'storm', 'temple', 'weak', 'artist']\n",
      "closest to 'walk' are: ['walk', 'art', 'rank', 'breast', 'weak', 'walked', 'solitude', 'story', 'shut', 'born']\n",
      "429957/429957 [==============================] - 34s 79us/sample - loss: 7.1942 - val_loss: 6.7414\n",
      "Epoch 2/5\n",
      "429824/429957 [============================>.] - ETA: 0s - loss: 6.5892closest to 'the' are: ['the', 'a', 'his', 'an', 'its', 'some', 'those', 'our', 'any', 'their']\n",
      "closest to 'animal' are: ['animal', 'hem', 'grove', 'buckle', 'winds', 'Drum', 'rain', 'nearest', 'accent', 'stranger']\n",
      "closest to 'me' are: ['me', 'him', 'what', 'seemed', 'nothing', 'so', 'ever', 'enough', 'myself', 'no']\n",
      "closest to 'like' are: ['like', 'till', 'giver', 'guarding', 'versed', 'Islands', 'archway', 'producing', 'thickets', 'Caspian']\n",
      "closest to 'love' are: ['love', 'speech', 'fear', 'misery', 'means', 'habits', 'talked', 'despair', 'Paris', 'use']\n",
      "closest to 'ship' are: ['ship', 'sands', 'storm', 'shrank', 'members', 'slope', 'ravages', 'minutest', 'Greek', 'tempest']\n",
      "closest to 'walk' are: ['walk', 'preferred', 'endeavoured', 'solitude', 'probable', 'procured', 'project', 'struggle', 'add', 'demand']\n",
      "429957/429957 [==============================] - 34s 80us/sample - loss: 6.5891 - val_loss: 6.5630\n",
      "Epoch 3/5\n",
      "429312/429957 [============================>.] - ETA: 0s - loss: 6.4206closest to 'the' are: ['the', 'its', 'a', 'some', 'those', 'an', 'our', 'his', 'their', 'any']\n",
      "closest to 'animal' are: ['animal', 'opening', 'beach', 'lake', 'nearest', 'buckle', 'winds', 'bog', 'cemetery', 'stranger']\n",
      "closest to 'me' are: ['me', 'him', 'seemed', 'enough', 'us', 'them', 'myself', 'nothing', 'able', 'needless']\n",
      "closest to 'like' are: ['like', 'till', 'Pole', 'despite', 'pathless', 'throughout', 'untameable', 'downward', 'letting', 'whilst']\n",
      "closest to 'love' are: ['love', 'fear', 'talked', 'pain', 'talking', 'Paris', 'smiles', 'Italy', 'despair', 'sorrow']\n",
      "closest to 'ship' are: ['ship', 'fabric', 'soil', 'mosques', 'stream', 'twilight', 'august', 'tenth', 'subtle', 'woman']\n",
      "closest to 'walk' are: ['walk', 'declare', 'irksome', 'indulge', 'submit', 'remind', 'dwell', 'sympathize', 'morrow', 'ride']\n",
      "429957/429957 [==============================] - 34s 78us/sample - loss: 6.4205 - val_loss: 6.4803\n",
      "Epoch 4/5\n",
      "429824/429957 [============================>.] - ETA: 0s - loss: 6.3004closest to 'the' are: ['the', 'its', 'some', 'those', 'our', 'every', 'their', 'an', 'any', 'these']\n",
      "closest to 'animal' are: ['animal', \"N'bangus\", 'breeze', 'winds', 'casement', 'ramble', 'preparations', 'impediments', 'beach', 'nimbleness']\n",
      "closest to 'me' are: ['me', 'him', 'seemed', 'us', 'enough', 'them', 'go', 'needless', 'seem', 'himself']\n",
      "closest to 'like' are: ['like', 'scull', 'till', 'pathless', 'matchmen', 'hypocritical', 'Skai', 'whilst', 'letting', 'nauseous']\n",
      "closest to 'love' are: ['love', 'sickness', 'Italy', 'fear', 'despair', 'yours', 'pain', 'wretchedness', 'misery', 'dreamt']\n",
      "closest to 'ship' are: ['ship', 'stream', 'lowest', 'closes', 'guarding', 'twilight', 'fabric', 'subtle', 'august', 'flour']\n",
      "closest to 'walk' are: ['walk', 'submit', 'declare', 'retire', 'irksome', 'apologize', 'indulge', 'dwell', 'lead', 'sail']\n",
      "429957/429957 [==============================] - 34s 80us/sample - loss: 6.3003 - val_loss: 6.4102\n",
      "Epoch 5/5\n",
      "429568/429957 [============================>.] - ETA: 0s - loss: 6.1796closest to 'the' are: ['the', 'its', 'every', 'some', 'those', 'our', 'these', 'Eastern', 'flights', 'masses']\n",
      "closest to 'animal' are: ['animal', 'nimbleness', 'casement', \"N'bangus\", 'impediments', 'ramble', 'lamb', 'breeze', 'shutter', 'devil']\n",
      "closest to 'me' are: ['me', 'him', 'us', 'herself', 'them', 'enough', 'seem', 'cease', 'himself', 'needless']\n",
      "closest to 'like' are: ['like', 'weasel', 'producing', 'workman', 'Solomon', 'Through', 'scull', 'rectangle', 'observant', 'matchmen']\n",
      "closest to 'love' are: ['love', 'sickness', 'despair', 'aberration', 'distress', 'revenge', 'misery', 'paining', 'Elizabeth', 'fear']\n",
      "closest to 'ship' are: ['ship', 'guarding', 'tenantless', 'Top', 'gorge', 'Jew', 'shaft', 'twilight', 'disdained', 'diagnosis']\n",
      "closest to 'walk' are: ['walk', 'submit', 'retire', 'lead', 'sail', 'declare', 'ride', 'reference', 'revert', 'dwell']\n",
      "429957/429957 [==============================] - 34s 79us/sample - loss: 6.1792 - val_loss: 6.3432\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x1f4d5eb4a08>"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "embedding_model.fit(x_embed, y_embed, batch_size=256, epochs=5, validation_split=0.1, callbacks=[PrintSim()])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(27752, 64)"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "embedding_weights = embedding_model.get_layer(index=0).get_weights()[0]\n",
    "embedding_weights.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[',',\n",
       " 'the',\n",
       " 'of',\n",
       " '.',\n",
       " 'and',\n",
       " 'to',\n",
       " 'I',\n",
       " 'a',\n",
       " 'in',\n",
       " 'was',\n",
       " 'that',\n",
       " ';',\n",
       " 'my',\n",
       " 'had',\n",
       " 'with',\n",
       " 'it',\n",
       " 'his',\n",
       " 'as',\n",
       " 'he',\n",
       " 'not']"
      ]
     },
     "execution_count": 34,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "def get_words_sorted_by_usage():\n",
    "    cnt = np.zeros(vocab_size)\n",
    "    for s in num_data:\n",
    "        for x in s:\n",
    "            cnt[x]+=1\n",
    "    return cnt.argsort()[::-1]\n",
    "\n",
    "most_used = get_words_sorted_by_usage()\n",
    "[num2word[x] for x in most_used[:20]]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(100, 3)"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data = embedding_weights[most_used[:100]]\n",
    "\n",
    "from sklearn.decomposition import PCA\n",
    "\n",
    "pca = PCA(n_components=3)\n",
    "\n",
    "pc = pca.fit_transform(data)\n",
    "pc.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1080x1080 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "fig, ax = plt.subplots(figsize=(15, 15))\n",
    "\n",
    "ax.scatter(pc[:, 0], pc[:, 1], c=pc[:, 2])\n",
    "\n",
    "for i in range(pc.shape[0]):\n",
    "    ax.annotate(num2word[most_used[i]], (pc[i, 0], pc[i, 1]))\n",
    "    \n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_closest(v, n=5):\n",
    "    return np.sum((embedding_weights - v)**2, axis=-1).argsort()[:n]\n",
    "\n",
    "def arithmetic(words, weights, n=5):\n",
    "    v = np.zeros(embedding_dim)\n",
    "    for i in range(len(words)):\n",
    "        v += embedding_weights[word2num[words[i]]]*weights[i]\n",
    "    return [num2word[x] for x in get_closest(v, n)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['palace', 'ferocity', 'moulding', 'floors', 'locking']"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Not enough text\n",
    "arithmetic([\"king\", \"woman\", \"man\"], [1, 1, -1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# Да вземем някъв предварително трениран embedding, за по-забавно на български\n",
    "\n",
    "import fasttext\n",
    "\n",
    "ft = fasttext.load_model('cc.bg.300.bin')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "bg_num2word = ft.words\n",
    "bg_word2num = {w:i for i,w in enumerate(bg_num2word)}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "bg_embed = np.zeros((len(bg_num2word), 300))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "for i, w in enumerate(bg_num2word):\n",
    "    bg_embed[i] = ft[w]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "del ft"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_closest_bg(v, n=5):\n",
    "    return np.sum((bg_embed - v)**2, axis=-1).argsort()[:n]\n",
    "\n",
    "def arithmetic_bg(words, weights, n=5):\n",
    "    v = np.zeros(300)\n",
    "    for i in range(len(words)):\n",
    "        v += bg_embed[bg_word2num[words[i]]]*weights[i]\n",
    "    return [bg_num2word[x] for x in get_closest_bg(v, n)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['цар', 'царя', 'крал', 'царят', 'царете']"
      ]
     },
     "execution_count": 45,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "arithmetic_bg([\"цар\"], [1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['цар', 'царица', 'царя', 'царете', 'царят']"
      ]
     },
     "execution_count": 46,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "arithmetic_bg([\"цар\", \"мъж\", \"жена\"], [1, -1, 1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['Лондон', 'Париж', 'Франция', 'Англия', 'Лондон-']"
      ]
     },
     "execution_count": 47,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "arithmetic_bg([\"Лондон\", \"Великобритания\", \"Франция\"], [1, -1, 1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['боза', 'боза.', 'бозата', 'сливова', 'ракийка']"
      ]
     },
     "execution_count": 48,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "arithmetic_bg([\"боза\", \"лош\", \"добър\"], [1, -1, 1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['земя.', 'воине', 'воин.', 'земя-', 'воинът']"
      ]
     },
     "execution_count": 49,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "arithmetic_bg([\"воин\", \"земя\"], [0.5, 0.5])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Да видим и един 1D CNN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "19579"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(num_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "876"
      ]
     },
     "execution_count": 51,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "max([len(x) for x in num_data])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[24218., 20295., 13975., ..., 27752., 27752., 27752.],\n",
       "       [ 3056., 18067.,  2516., ..., 27752., 27752., 27752.],\n",
       "       [23024., 27285.,  7817., ..., 27752., 27752., 27752.],\n",
       "       ...,\n",
       "       [ 5590.,  4100., 14093., ..., 27752., 27752., 27752.],\n",
       "       [15001., 22375.,  7533., ..., 27752., 27752., 27752.],\n",
       "       [ 3199., 14481., 10664., ..., 27752., 27752., 27752.]])"
      ]
     },
     "execution_count": 52,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "num_x = np.ones((len(num_data), max([len(x) for x in num_data])))*vocab_size\n",
    "for i in range(len(num_data)):\n",
    "    num_x[i, :len(num_data[i])] = num_data[i]\n",
    "    \n",
    "num_x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(19579, 3)"
      ]
     },
     "execution_count": 53,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "to_num = {'EAP':0, 'MWS':1, 'HPL':2}\n",
    "num_y = np.eye(3)[np.array([to_num[x] for x in train.author])]\n",
    "num_y.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_2\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "embedding_1 (Embedding)      (None, None, 64)          1776192   \n",
      "_________________________________________________________________\n",
      "dropout_3 (Dropout)          (None, None, 64)          0         \n",
      "_________________________________________________________________\n",
      "conv1d (Conv1D)              (None, None, 32)          10272     \n",
      "_________________________________________________________________\n",
      "conv1d_1 (Conv1D)            (None, None, 64)          6208      \n",
      "_________________________________________________________________\n",
      "global_max_pooling1d (Global (None, 64)                0         \n",
      "_________________________________________________________________\n",
      "dense_4 (Dense)              (None, 3)                 195       \n",
      "=================================================================\n",
      "Total params: 1,792,867\n",
      "Trainable params: 1,792,867\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = tf.keras.Sequential([\n",
    "    tf.keras.layers.Embedding(vocab_size+1, embedding_dim, input_shape=(None,)),\n",
    "    tf.keras.layers.Dropout(0.2),\n",
    "    tf.keras.layers.Conv1D(32, 5, activation='relu'),\n",
    "    tf.keras.layers.Conv1D(64, 3, dilation_rate=2, activation='relu'),\n",
    "    tf.keras.layers.GlobalMaxPooling1D(),\n",
    "    tf.keras.layers.Dense(3, activation='softmax')\n",
    "])\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 17621 samples, validate on 1958 samples\n",
      "Epoch 1/10\n",
      "17621/17621 [==============================] - 11s 650us/sample - loss: 0.8403 - categorical_crossentropy: 0.8403 - accuracy: 0.5949 - val_loss: 0.5508 - val_categorical_crossentropy: 0.5508 - val_accuracy: 0.7829\n",
      "Epoch 2/10\n",
      "17621/17621 [==============================] - 8s 479us/sample - loss: 0.3332 - categorical_crossentropy: 0.3332 - accuracy: 0.8743 - val_loss: 0.4661 - val_categorical_crossentropy: 0.4661 - val_accuracy: 0.8192\n",
      "Epoch 3/10\n",
      "17621/17621 [==============================] - 8s 479us/sample - loss: 0.1446 - categorical_crossentropy: 0.1446 - accuracy: 0.9505 - val_loss: 0.5319 - val_categorical_crossentropy: 0.5319 - val_accuracy: 0.8202\n",
      "Epoch 4/10\n",
      "17621/17621 [==============================] - 9s 497us/sample - loss: 0.0717 - categorical_crossentropy: 0.0717 - accuracy: 0.9759 - val_loss: 0.6401 - val_categorical_crossentropy: 0.6401 - val_accuracy: 0.8177\n",
      "Epoch 5/10\n",
      "17621/17621 [==============================] - 9s 497us/sample - loss: 0.0386 - categorical_crossentropy: 0.0386 - accuracy: 0.9880 - val_loss: 0.7350 - val_categorical_crossentropy: 0.7350 - val_accuracy: 0.8177\n",
      "Epoch 6/10\n",
      "17621/17621 [==============================] - 9s 502us/sample - loss: 0.0234 - categorical_crossentropy: 0.0234 - accuracy: 0.9928 - val_loss: 0.8277 - val_categorical_crossentropy: 0.8277 - val_accuracy: 0.8115\n",
      "Epoch 7/10\n",
      "17621/17621 [==============================] - 9s 482us/sample - loss: 0.0141 - categorical_crossentropy: 0.0141 - accuracy: 0.9961 - val_loss: 0.9327 - val_categorical_crossentropy: 0.9327 - val_accuracy: 0.8126\n",
      "Epoch 8/10\n",
      "17621/17621 [==============================] - 8s 463us/sample - loss: 0.0109 - categorical_crossentropy: 0.0109 - accuracy: 0.9968 - val_loss: 1.0026 - val_categorical_crossentropy: 1.0026 - val_accuracy: 0.8095\n",
      "Epoch 9/10\n",
      "17621/17621 [==============================] - 8s 454us/sample - loss: 0.0079 - categorical_crossentropy: 0.0079 - accuracy: 0.9978 - val_loss: 1.1331 - val_categorical_crossentropy: 1.1331 - val_accuracy: 0.7998\n",
      "Epoch 10/10\n",
      "17621/17621 [==============================] - 8s 457us/sample - loss: 0.0062 - categorical_crossentropy: 0.0062 - accuracy: 0.9978 - val_loss: 1.1875 - val_categorical_crossentropy: 1.1875 - val_accuracy: 0.8023\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x1f4d7f8ee08>"
      ]
     },
     "execution_count": 55,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.compile(tf.keras.optimizers.Adam(1e-3), loss='categorical_crossentropy', metrics = ['categorical_crossentropy', 'accuracy'])\n",
    "\n",
    "model.fit(num_x, num_y, batch_size=64, epochs=10, validation_split=0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_3\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "embedding_2 (Embedding)      (None, None, 64)          1776192   \n",
      "_________________________________________________________________\n",
      "dropout_4 (Dropout)          (None, None, 64)          0         \n",
      "_________________________________________________________________\n",
      "conv1d_2 (Conv1D)            (None, None, 32)          10272     \n",
      "_________________________________________________________________\n",
      "conv1d_3 (Conv1D)            (None, None, 64)          6208      \n",
      "_________________________________________________________________\n",
      "global_max_pooling1d_1 (Glob (None, 64)                0         \n",
      "_________________________________________________________________\n",
      "dense_5 (Dense)              (None, 32)                2080      \n",
      "_________________________________________________________________\n",
      "dense_6 (Dense)              (None, 3)                 99        \n",
      "=================================================================\n",
      "Total params: 1,794,851\n",
      "Trainable params: 18,659\n",
      "Non-trainable params: 1,776,192\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "embed_init = np.vstack([embedding_weights, np.zeros(embedding_dim)])\n",
    "\n",
    "model_pre = tf.keras.Sequential([\n",
    "    tf.keras.layers.Embedding(\n",
    "        vocab_size+1, \n",
    "        embedding_dim, \n",
    "        input_shape=(None,),\n",
    "        embeddings_initializer = tf.constant_initializer(embed_init),\n",
    "        trainable = False\n",
    "    ),\n",
    "    tf.keras.layers.Dropout(0.2),\n",
    "    tf.keras.layers.Conv1D(32, 5, activation='relu'),\n",
    "    tf.keras.layers.Conv1D(64, 3, dilation_rate=2, activation='relu'),\n",
    "    tf.keras.layers.GlobalMaxPooling1D(),\n",
    "    tf.keras.layers.Dense(32, activation='relu'),\n",
    "    tf.keras.layers.Dense(3, activation='softmax')\n",
    "])\n",
    "\n",
    "model_pre.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 17621 samples, validate on 1958 samples\n",
      "Epoch 1/10\n",
      "17621/17621 [==============================] - 5s 271us/sample - loss: 0.9858 - categorical_crossentropy: 0.9858 - accuracy: 0.5079 - val_loss: 0.9191 - val_categorical_crossentropy: 0.9191 - val_accuracy: 0.5598\n",
      "Epoch 2/10\n",
      "17621/17621 [==============================] - 4s 242us/sample - loss: 0.8726 - categorical_crossentropy: 0.8726 - accuracy: 0.6005 - val_loss: 0.8303 - val_categorical_crossentropy: 0.8303 - val_accuracy: 0.6267\n",
      "Epoch 3/10\n",
      "17621/17621 [==============================] - 4s 240us/sample - loss: 0.8212 - categorical_crossentropy: 0.8212 - accuracy: 0.6296 - val_loss: 0.8482 - val_categorical_crossentropy: 0.8482 - val_accuracy: 0.6175\n",
      "Epoch 4/10\n",
      "17621/17621 [==============================] - 4s 240us/sample - loss: 0.7905 - categorical_crossentropy: 0.7905 - accuracy: 0.6479 - val_loss: 0.7801 - val_categorical_crossentropy: 0.7801 - val_accuracy: 0.6599\n",
      "Epoch 5/10\n",
      "17621/17621 [==============================] - 4s 239us/sample - loss: 0.7674 - categorical_crossentropy: 0.7674 - accuracy: 0.6586 - val_loss: 0.7782 - val_categorical_crossentropy: 0.7782 - val_accuracy: 0.6604\n",
      "Epoch 6/10\n",
      "17621/17621 [==============================] - 4s 242us/sample - loss: 0.7432 - categorical_crossentropy: 0.7432 - accuracy: 0.6744 - val_loss: 0.7691 - val_categorical_crossentropy: 0.7691 - val_accuracy: 0.6619\n",
      "Epoch 7/10\n",
      "17621/17621 [==============================] - 4s 235us/sample - loss: 0.7279 - categorical_crossentropy: 0.7279 - accuracy: 0.6813 - val_loss: 0.7542 - val_categorical_crossentropy: 0.7542 - val_accuracy: 0.6645\n",
      "Epoch 8/10\n",
      "17621/17621 [==============================] - 4s 239us/sample - loss: 0.7165 - categorical_crossentropy: 0.7165 - accuracy: 0.6890 - val_loss: 0.7728 - val_categorical_crossentropy: 0.7728 - val_accuracy: 0.6593\n",
      "Epoch 9/10\n",
      "17621/17621 [==============================] - 4s 243us/sample - loss: 0.7051 - categorical_crossentropy: 0.7051 - accuracy: 0.6926 - val_loss: 0.7584 - val_categorical_crossentropy: 0.7584 - val_accuracy: 0.6588\n",
      "Epoch 10/10\n",
      "17621/17621 [==============================] - 4s 231us/sample - loss: 0.6887 - categorical_crossentropy: 0.6887 - accuracy: 0.7001 - val_loss: 0.7649 - val_categorical_crossentropy: 0.7649 - val_accuracy: 0.6665\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x1f4d97234c8>"
      ]
     },
     "execution_count": 58,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model_pre.compile(tf.keras.optimizers.Adam(1e-3), loss='categorical_crossentropy', metrics = ['categorical_crossentropy', 'accuracy'])\n",
    "\n",
    "model_pre.fit(num_x, num_y, batch_size=64, epochs=10, validation_split=0.1)"
   ]
  },
  {
   "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": "3.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}