pos_tagger_bert.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "2b1b350c-5a3c-434c-bcb2-0672625d154d"
    }
   },
   "source": [
    "# NATURAL LANGUAGE PROCESSING WITH TRANSFORMERS\n",
    "# Α Pos Tagger trained on UD treebank with fine-tuning a BERT model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Keras BERT implementation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "nbpresent": {
     "id": "f61bc4e3-579a-44c9-a7c6-4dc1f9fbf7cb"
    }
   },
   "outputs": [],
   "source": [
    "# First install some extra packages\n",
    "# ! pip install pyconll\n",
    "# ! pip install pydot\n",
    "# ! pip install graphiz\n",
    "# ! pip install bert-tensorflow"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "nbpresent": {
     "id": "f4a0e2d1-af80-44ba-a384-fab4d5d6c703"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n",
      "WARNING: Logging before flag parsing goes to stderr.\n",
      "W0527 23:36:22.545466 140718967228160 __init__.py:56] Some hub symbols are not available because TensorFlow version is less than 1.14\n"
     ]
    }
   ],
   "source": [
    "from numpy.random import seed\n",
    "seed(1)\n",
    "from tensorflow import set_random_seed\n",
    "set_random_seed(2)\n",
    "import pyconll, keras, pickle, os, random, nltk, datetime, warnings, gc, urllib.request, zipfile, collections\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "from sklearn.utils import shuffle\n",
    "from sklearn.metrics import recall_score, precision_score, classification_report, accuracy_score, confusion_matrix, f1_score\n",
    "from sklearn.metrics.classification import UndefinedMetricWarning\n",
    "\n",
    "from keras.models import Sequential, Model\n",
    "from keras.layers import Embedding, Dense, Input, concatenate, Layer, Lambda, Dropout, Activation\n",
    "from keras.callbacks import ModelCheckpoint, EarlyStopping, Callback, TensorBoard\n",
    "from keras import backend as K\n",
    "from keras.models import load_model\n",
    "from keras.utils import plot_model\n",
    "from keras.utils.np_utils import to_categorical\n",
    "\n",
    "import tensorflow as tf\n",
    "import tensorflow_hub as hub\n",
    "from bert.tokenization import FullTokenizer\n",
    "\n",
    "from tqdm import tqdm_notebook\n",
    "from IPython.display import Image \n",
    "warnings.filterwarnings(action='ignore', category=UndefinedMetricWarning)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "0f518892-45cc-4978-97e8-02c1fce60407"
    }
   },
   "source": [
    "# Various plot functions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "nbpresent": {
     "id": "ffc21443-9fd6-4cd3-86e1-a1a518cb6aca"
    }
   },
   "outputs": [],
   "source": [
    "def plot_learning_curve(estimator, title, X, y, ylim=None, cv=None,\n",
    "                        n_jobs=None, train_sizes=np.linspace(.1, 1.0, 5)):\n",
    "    plt.title(title)\n",
    "    if ylim is not None:\n",
    "        plt.ylim(*ylim)\n",
    "    plt.xlabel(\"Training examples\")\n",
    "    plt.ylabel(\"Accuracy\")\n",
    "    train_sizes, train_scores, test_scores = learning_curve(\n",
    "        estimator, X, y, cv=cv, n_jobs=n_jobs, train_sizes=train_sizes, scoring='accuracy')\n",
    "    train_scores_mean = np.mean(train_scores, axis=1)\n",
    "    train_scores_std = np.std(train_scores, axis=1)\n",
    "    test_scores_mean = np.mean(test_scores, axis=1)\n",
    "    test_scores_std = np.std(test_scores, axis=1)\n",
    "    plt.grid()\n",
    "\n",
    "    plt.fill_between(train_sizes, train_scores_mean - train_scores_std,\n",
    "                     train_scores_mean + train_scores_std, alpha=0.1,\n",
    "                     color=\"r\")\n",
    "    plt.fill_between(train_sizes, test_scores_mean - test_scores_std,\n",
    "                     test_scores_mean + test_scores_std, alpha=0.1, color=\"g\")\n",
    "    plt.plot(train_sizes, train_scores_mean, 'o-', color=\"r\",\n",
    "             label=\"Training score\")\n",
    "    plt.plot(train_sizes, test_scores_mean, 'o-', color=\"g\",\n",
    "             label=\"Cross-validation score\")\n",
    "\n",
    "    plt.legend(loc=\"best\")\n",
    "    return plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "nbpresent": {
     "id": "aa581c97-05ee-4e34-9f9d-884ce9ba4b66"
    }
   },
   "outputs": [],
   "source": [
    "def plot_confusion_matrix(f1,\n",
    "                          cm,\n",
    "                          target_names,\n",
    "                          title='Confusion matrix',\n",
    "                          cmap=None,\n",
    "                          normalize=True,\n",
    "                          i=1):\n",
    "    \"\"\"\n",
    "    given a sklearn confusion matrix (cm), make a nice plot\n",
    "\n",
    "    Arguments\n",
    "    ---------\n",
    "    cm:           confusion matrix from sklearn.metrics.confusion_matrix\n",
    "\n",
    "    target_names: given classification classes such as [0, 1, 2]\n",
    "                  the class names, for example: ['high', 'medium', 'low']\n",
    "\n",
    "    title:        the text to display at the top of the matrix\n",
    "\n",
    "    cmap:         the gradient of the values displayed from matplotlib.pyplot.cm\n",
    "                  see http://matplotlib.org/examples/color/colormaps_reference.html\n",
    "                  plt.get_cmap('jet') or plt.cm.Blues\n",
    "\n",
    "    normalize:    If False, plot the raw numbers\n",
    "                  If True, plot the proportions\n",
    "\n",
    "    Usage\n",
    "    -----\n",
    "    plot_confusion_matrix(cm           = cm,                  # confusion matrix created by\n",
    "                                                              # sklearn.metrics.confusion_matrix\n",
    "                          normalize    = True,                # show proportions\n",
    "                          target_names = y_labels_vals,       # list of names of the classes\n",
    "                          title        = best_estimator_name) # title of graph\n",
    "\n",
    "    Citiation\n",
    "    ---------\n",
    "    http://scikit-learn.org/stable/auto_examples/model_selection/plot_confusion_matrix.html\n",
    "\n",
    "    \"\"\"\n",
    "    import matplotlib.pyplot as plt\n",
    "    import numpy as np\n",
    "    import itertools\n",
    "\n",
    "    accuracy = np.trace(cm) / float(np.sum(cm))\n",
    "    misclass = 1 - accuracy\n",
    "    plt.figure(figsize=(10, 6))\n",
    "    if cmap is None:\n",
    "        cmap = plt.get_cmap('Blues')\n",
    "\n",
    "    plt.imshow(cm, interpolation='nearest', cmap=cmap)\n",
    "    plt.title(title)\n",
    "    plt.colorbar()\n",
    "\n",
    "    if target_names is not None:\n",
    "        tick_marks = np.arange(len(target_names))\n",
    "        plt.xticks(tick_marks, target_names, rotation=45)\n",
    "        plt.yticks(tick_marks, target_names)\n",
    "\n",
    "    if normalize:\n",
    "        cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]\n",
    "\n",
    "\n",
    "    thresh = cm.max() / 1.5 if normalize else cm.max() / 2\n",
    "    for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):\n",
    "        if normalize:\n",
    "            plt.text(j, i, \"{:0.4f}\".format(cm[i, j]),\n",
    "                     horizontalalignment=\"center\",\n",
    "                     color=\"white\" if cm[i, j] > thresh else \"black\")\n",
    "        else:\n",
    "            plt.text(j, i, \"{:,}\".format(cm[i, j]),\n",
    "                     horizontalalignment=\"center\",\n",
    "                     color=\"white\" if cm[i, j] > thresh else \"black\")\n",
    "\n",
    "\n",
    "    plt.tight_layout()\n",
    "    plt.ylabel('True label')\n",
    "    plt.xlabel('Predicted label\\naccuracy={:0.4f}; misclass={:0.4f}; f1-score={:0.4f}'.format(accuracy, misclass, f1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "nbpresent": {
     "id": "7626f765-a0ca-4230-b9b4-5aea4c185273"
    }
   },
   "outputs": [],
   "source": [
    "def plot_acc():\n",
    "    plt.plot(history.history['acc'])\n",
    "    plt.plot(history.history['val_acc'])\n",
    "    plt.title('model accuracy')\n",
    "    plt.ylabel('accuracy')\n",
    "    plt.xlabel('epoch')\n",
    "    plt.legend(['train', 'test'], loc='upper left')\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "nbpresent": {
     "id": "4afdbba0-cb73-4bdc-97f7-3cbdd2d19079"
    },
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "def plot_loss():\n",
    "    plt.plot(history.history['loss'])\n",
    "    plt.plot(history.history['val_loss'])\n",
    "    plt.title('model loss')\n",
    "    plt.ylabel('loss')\n",
    "    plt.xlabel('epoch')\n",
    "    plt.legend(['train', 'test'], loc='upper left')\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "1585e7f9-881f-46fe-a1e8-ca970e292167"
    }
   },
   "source": [
    "# Load Dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "nbpresent": {
     "id": "75d32121-1074-4cd6-8355-599afeff56db"
    }
   },
   "outputs": [],
   "source": [
    "UD_ENGLISH_TRAIN = 'en_partut-ud-train.conllu'\n",
    "UD_ENGLISH_DEV = 'en_partut-ud-dev.conllu'\n",
    "UD_ENGLISH_TEST = 'en_partut-ud-test.conllu'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "nbpresent": {
     "id": "2eb0b136-1052-4d48-8a66-48b5ebc96ff8"
    }
   },
   "outputs": [],
   "source": [
    "def download_files():\n",
    "    print('Downloading English treebank...')\n",
    "    urllib.request.urlretrieve('http://archive.aueb.gr:8085/files/en_partut-ud-dev.conllu', 'en_partut-ud-dev.conllu')\n",
    "    urllib.request.urlretrieve('http://archive.aueb.gr:8085/files/en_partut-ud-test.conllu', 'en_partut-ud-test.conllu')\n",
    "    urllib.request.urlretrieve('http://archive.aueb.gr:8085/files/en_partut-ud-train.conllu', 'en_partut-ud-train.conllu')\n",
    "    print('Treebank downloaded.')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "nbpresent": {
     "id": "4e617ca6-4cce-42a0-9396-28b9d704df73"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading English treebank...\n",
      "Treebank downloaded.\n"
     ]
    }
   ],
   "source": [
    "download_files()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "43c9097b-37be-4f3d-b3d3-4a54a94c1174"
    }
   },
   "source": [
    "# Preprocessing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "nbpresent": {
     "id": "016836bb-abb2-4235-ab21-df90a6e1e44e"
    }
   },
   "outputs": [],
   "source": [
    "def read_conllu(path):\n",
    "    data = pyconll.load_from_file(path)\n",
    "    tagged_sentences=[]\n",
    "    t=0\n",
    "    for sentence in data:\n",
    "        tagged_sentence=[]\n",
    "        for token in sentence:\n",
    "            if token.upos and token.form:\n",
    "                t+=1\n",
    "                tagged_sentence.append((token.form.lower(), token.upos))\n",
    "        tagged_sentences.append(tagged_sentence)\n",
    "    return tagged_sentences"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "11bff856-f685-49ef-a4e6-b5f2a768a9ad"
    }
   },
   "source": [
    "Load train, development and test set in the appropriate tagged format, tuple (word, pos-tag)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "nbpresent": {
     "id": "fc49598b-75d3-45e9-9fa3-a7b604234760"
    }
   },
   "outputs": [],
   "source": [
    "train_sentences = read_conllu(UD_ENGLISH_TRAIN)\n",
    "val_sentences = read_conllu(UD_ENGLISH_DEV)\n",
    "test_sentences = read_conllu(UD_ENGLISH_TEST)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "3f72d1da-c1d1-4e25-a782-e12a66f1cedf"
    }
   },
   "source": [
    "Print some statistics"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "nbpresent": {
     "id": "57ee9c67-7ec6-4054-91bc-a9aa896ba18a"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tagged sentences in train set:  1781\n",
      "Tagged words in train set: 43518\n",
      "========================================\n",
      "Tagged sentences in dev set:  156\n",
      "Tagged words in dev set: 2722\n",
      "========================================\n",
      "Tagged sentences in test set:  153\n",
      "Tagged words in test set: 3408\n",
      "****************************************\n",
      "Total sentences in dataset: 2090\n"
     ]
    }
   ],
   "source": [
    "print(\"Tagged sentences in train set: \", len(train_sentences))\n",
    "print(\"Tagged words in train set:\", len([item for sublist in train_sentences for item in sublist]))\n",
    "print(40*'=')\n",
    "print(\"Tagged sentences in dev set: \", len(val_sentences))\n",
    "print(\"Tagged words in dev set:\", len([item for sublist in val_sentences for item in sublist]))\n",
    "print(40*'=')\n",
    "print(\"Tagged sentences in test set: \", len(test_sentences))\n",
    "print(\"Tagged words in test set:\", len([item for sublist in test_sentences for item in sublist]))\n",
    "print(40*'*')\n",
    "print(\"Total sentences in dataset:\", len(train_sentences)+len(val_sentences)+len(test_sentences))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "nbpresent": {
     "id": "eb97895a-d649-4318-b3ac-d7d6b6088a12"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[('distribution', 'NOUN'), ('of', 'ADP'), ('this', 'DET'), ('license', 'NOUN'), ('does', 'AUX'), ('not', 'PART'), ('create', 'VERB'), ('an', 'DET'), ('attorney', 'NOUN'), ('-', 'PUNCT'), ('client', 'NOUN'), ('relationship', 'NOUN'), ('.', 'PUNCT')]\n"
     ]
    }
   ],
   "source": [
    "print(train_sentences[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "nbpresent": {
     "id": "573e2ec4-2df3-49f2-913d-3d0af2d6dc36"
    }
   },
   "outputs": [],
   "source": [
    "# Some usefull functions\n",
    "def tag_sequence(sentences):\n",
    "    return [[t for w, t in sentence] for sentence in sentences]\n",
    "\n",
    "def text_sequence(sentences):\n",
    "    return [[w for w, t in sentence] for sentence in sentences]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "533dd9e2-9e0e-43c1-8d13-b096a81a8984"
    }
   },
   "source": [
    "# Build dictionary with tag vocabulary"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "TOTAL TAGS:  17\n"
     ]
    }
   ],
   "source": [
    "tags = set([item for sublist in train_sentences+test_sentences+val_sentences for _, item in sublist])\n",
    "print('TOTAL TAGS: ', len(tags))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Add '-PAD-' tag with index=0 as used in BERT code"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "nbpresent": {
     "id": "98944f66-c2ae-4b29-ba01-99287044a815"
    }
   },
   "outputs": [],
   "source": [
    "tag2int = {}\n",
    "int2tag = {}\n",
    "\n",
    "for i, tag in enumerate(sorted(tags)):\n",
    "    tag2int[tag] = i+1\n",
    "    int2tag[i+1] = tag"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "nbpresent": {
     "id": "a935a8bd-0522-42a1-8fea-285b671df0b8"
    }
   },
   "outputs": [],
   "source": [
    "# Special character for the tags\n",
    "tag2int['-PAD-'] = 0\n",
    "int2tag[0] = '-PAD-'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "nbpresent": {
     "id": "3e55930b-4268-4b18-a7c7-de6675cdcc01"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total tags: 18\n"
     ]
    }
   ],
   "source": [
    "n_tags = len(tag2int)\n",
    "print('Total tags:', n_tags)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['ADJ',\n",
       " 'ADP',\n",
       " 'ADV',\n",
       " 'AUX',\n",
       " 'CCONJ',\n",
       " 'DET',\n",
       " 'INTJ',\n",
       " 'NOUN',\n",
       " 'NUM',\n",
       " 'PART',\n",
       " 'PRON',\n",
       " 'PROPN',\n",
       " 'PUNCT',\n",
       " 'SCONJ',\n",
       " 'SYM',\n",
       " 'VERB',\n",
       " 'X',\n",
       " '-PAD-']"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "list(tag2int)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "6d6c1924-4661-4d3c-a446-ce81c05f9b71"
    }
   },
   "source": [
    "======================================================================================================="
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "0b2c54e8-30d1-4f19-b03e-533a9e27b069"
    }
   },
   "source": [
    "# Parameters"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "nbpresent": {
     "id": "5a212b96-46c2-4147-9dba-9eea08757f36"
    }
   },
   "outputs": [],
   "source": [
    "MAX_SEQUENCE_LENGTH = 70\n",
    "EPOCHS = 30"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "e1ee0fe5-a699-487d-9d36-b0342bc17d79"
    }
   },
   "source": [
    "# Special preprocessing for NNs"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "e2fff7c1-9822-4192-95b8-6721acef446f"
    }
   },
   "source": [
    "### This is a very importand step for the whole process! So be careful, to rerun this step in case you change one of the following parameters:\n",
    "- MAX_SEQUENCE_LENGTH"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "nbpresent": {
     "id": "12c47160-79d6-42d8-822d-e2d958e9ccb2"
    }
   },
   "outputs": [],
   "source": [
    "train_sentences = read_conllu(UD_ENGLISH_TRAIN)\n",
    "val_sentences = read_conllu(UD_ENGLISH_DEV)\n",
    "test_sentences = read_conllu(UD_ENGLISH_TEST)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "f67342ab-29db-4565-8912-1a7efb49f808"
    }
   },
   "source": [
    "Lets examine the train set sentences distribution:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "nbpresent": {
     "id": "844de2c1-e861-45db-8bc0-4e1f47724a95"
    }
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "plt.style.use(\"ggplot\")\n",
    "plt.hist([len(s) for s in train_sentences], bins=50)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "nbpresent": {
     "id": "d9a2a74a-dff8-4dce-a9d2-b251e8a33976"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Max sentence length: 209\n"
     ]
    }
   ],
   "source": [
    "print('Max sentence length:',len(max(train_sentences+val_sentences, key=len)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "0bff7138-d0ee-43b8-82ae-8d577b6e18f1"
    }
   },
   "source": [
    "To improve speed we will use a MAX_SEQUENCE_LENGTH shorter than the max lengthed sentence. To avoid truncating sequences during padding we split our sentences to MAX_SEQUENCE_LENGTH and so the number of samples increases accordingly. For example, if MAX_SEQUENCE_LENGTH=70, a sentence with length 150 splits in 3 sentences: 150=70+70+10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "nbpresent": {
     "id": "6b60d427-a8d8-4bb7-ad74-6b890098b6f7"
    }
   },
   "outputs": [],
   "source": [
    "def  split(sentences, max):\n",
    "    new=[]\n",
    "    for data in sentences:\n",
    "        new.append(([data[x:x+max] for x in range(0, len(data), max)]))\n",
    "    new = [val for sublist in new for val in sublist]\n",
    "    return new"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "nbpresent": {
     "id": "067299e9-7857-445b-99dd-1df2ddd4d740"
    }
   },
   "outputs": [],
   "source": [
    "train_sentences = split(train_sentences, MAX_SEQUENCE_LENGTH)\n",
    "val_sentences = split(val_sentences, MAX_SEQUENCE_LENGTH)\n",
    "test_sentences = split(test_sentences, MAX_SEQUENCE_LENGTH)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "nbpresent": {
     "id": "b180dbde-a00e-4f4b-aad7-33b0b31ad693"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "70"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(max(train_sentences+val_sentences, key=len))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_sentences = train_sentences + val_sentences"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "nbpresent": {
     "id": "bf4abb0c-ba12-4be4-918d-06172bc011f1"
    }
   },
   "source": [
    "# BERT implementation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "nbpresent": {
     "id": "b2e4df36-f6ed-4961-b2d2-540fcbf651fd"
    }
   },
   "outputs": [],
   "source": [
    "# Initialize session\n",
    "sess = tf.Session()\n",
    "# Params for bert model and tokenization\n",
    "bert_path = \"https://tfhub.dev/google/bert_uncased_L-12_H-768_A-12/1\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "nbpresent": {
     "id": "81781553-5840-4539-8e31-ebf941647e22"
    }
   },
   "outputs": [],
   "source": [
    "train_text = text_sequence(train_sentences)\n",
    "test_text = text_sequence(test_sentences)\n",
    "#val_text = text_sequence(val_sentences)\n",
    "\n",
    "train_label = tag_sequence(train_sentences)\n",
    "test_label= tag_sequence(test_sentences)\n",
    "#val_label= tag_sequence(val_sentences)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "1951"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(train_text)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['distribution',\n",
       "  'of',\n",
       "  'this',\n",
       "  'license',\n",
       "  'does',\n",
       "  'not',\n",
       "  'create',\n",
       "  'an',\n",
       "  'attorney',\n",
       "  '-',\n",
       "  'client',\n",
       "  'relationship',\n",
       "  '.'],\n",
       " ['NOUN',\n",
       "  'ADP',\n",
       "  'DET',\n",
       "  'NOUN',\n",
       "  'AUX',\n",
       "  'PART',\n",
       "  'VERB',\n",
       "  'DET',\n",
       "  'NOUN',\n",
       "  'PUNCT',\n",
       "  'NOUN',\n",
       "  'NOUN',\n",
       "  'PUNCT'])"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_text[0], train_label[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Functions used for BERT representation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 88,
   "metadata": {
    "nbpresent": {
     "id": "1165aa8f-9bd2-48d7-8ad2-3109c960334f"
    }
   },
   "outputs": [],
   "source": [
    "class PaddingInputExample(object):\n",
    "    \"\"\"Fake example so the num input examples is a multiple of the batch size.\n",
    "  When running eval/predict on the TPU, we need to pad the number of examples\n",
    "  to be a multiple of the batch size, because the TPU requires a fixed batch\n",
    "  size. The alternative is to drop the last batch, which is bad because it means\n",
    "  the entire output data won't be generated.\n",
    "  We use this class instead of `None` because treating `None` as padding\n",
    "  battches could cause silent errors.\n",
    "  \"\"\"\n",
    "\n",
    "class InputExample(object):\n",
    "    \"\"\"A single training/test example for simple sequence classification.\"\"\"\n",
    "\n",
    "    def __init__(self, guid, text_a, text_b=None, label=None):\n",
    "        \"\"\"Constructs a InputExample.\n",
    "    Args:\n",
    "      guid: Unique id for the example.\n",
    "      text_a: string. The untokenized text of the first sequence. For single\n",
    "        sequence tasks, only this sequence must be specified.\n",
    "      text_b: (Optional) string. The untokenized text of the second sequence.\n",
    "        Only must be specified for sequence pair tasks.\n",
    "      label: (Optional) string. The label of the example. This should be\n",
    "        specified for train and dev examples, but not for test examples.\n",
    "    \"\"\"\n",
    "        self.guid = guid\n",
    "        self.text_a = text_a\n",
    "        self.text_b = text_b\n",
    "        self.label = label\n",
    "\n",
    "def create_tokenizer_from_hub_module():\n",
    "    \"\"\"Get the vocab file and casing info from the Hub module.\"\"\"\n",
    "    bert_module =  hub.Module(bert_path)\n",
    "    tokenization_info = bert_module(signature=\"tokenization_info\", as_dict=True)\n",
    "    vocab_file, do_lower_case = sess.run(\n",
    "        [\n",
    "            tokenization_info[\"vocab_file\"],\n",
    "            tokenization_info[\"do_lower_case\"],\n",
    "        ]\n",
    "    )\n",
    "\n",
    "    return FullTokenizer(vocab_file=vocab_file, do_lower_case=do_lower_case)\n",
    "\n",
    "def convert_single_example(tokenizer, example, max_seq_length=256):\n",
    "    \"\"\"Converts a single `InputExample` into a single `InputFeatures`.\"\"\"\n",
    "\n",
    "    if isinstance(example, PaddingInputExample):\n",
    "        input_ids = [0] * max_seq_length\n",
    "        input_mask = [0] * max_seq_length\n",
    "        segment_ids = [0] * max_seq_length\n",
    "        label_ids = [0] * max_seq_length\n",
    "        return input_ids, input_mask, segment_ids, label_ids\n",
    "    \n",
    "    tokens_a = example.text_a\n",
    "    if len(tokens_a) > max_seq_length-2:\n",
    "        tokens_a = tokens_a[0 : (max_seq_length-2)]\n",
    "\n",
    "# Token map will be an int -> int mapping between the `orig_tokens` index and\n",
    "# the `bert_tokens` index.\n",
    "\n",
    "# bert_tokens == [\"[CLS]\", \"john\", \"johan\", \"##son\", \"'\", \"s\", \"house\", \"[SEP]\"]\n",
    "# orig_to_tok_map == [1, 2, 4, 6]   \n",
    "    orig_to_tok_map = []              \n",
    "    tokens = []\n",
    "    segment_ids = []\n",
    "    \n",
    "    tokens.append(\"[CLS]\")\n",
    "    segment_ids.append(0)\n",
    "    orig_to_tok_map.append(len(tokens)-1)\n",
    "    #print(len(tokens_a))\n",
    "    for token in tokens_a:       \n",
    "        tokens.extend(tokenizer.tokenize(token))\n",
    "        orig_to_tok_map.append(len(tokens)-1)\n",
    "        segment_ids.append(0)\n",
    "    tokens.append(\"[SEP]\")\n",
    "    segment_ids.append(0)\n",
    "    orig_to_tok_map.append(len(tokens)-1)\n",
    "    input_ids = tokenizer.convert_tokens_to_ids([tokens[i] for i in orig_to_tok_map])\n",
    "    #print(len(orig_to_tok_map), len(tokens), len(input_ids), len(segment_ids)) #for debugging\n",
    "\n",
    "    # The mask has 1 for real tokens and 0 for padding tokens. Only real\n",
    "    # tokens are attended to.\n",
    "    input_mask = [1] * len(input_ids)\n",
    "    \n",
    "    label_ids = []\n",
    "    labels = example.label\n",
    "    label_ids.append(0)\n",
    "    label_ids.extend([tag2int[label] for label in labels])\n",
    "    label_ids.append(0)\n",
    "    #print(len(label_ids)) #for debugging\n",
    "    # Zero-pad up to the sequence length.\n",
    "    while len(input_ids) < max_seq_length:\n",
    "        input_ids.append(0)\n",
    "        input_mask.append(0)\n",
    "        segment_ids.append(0)\n",
    "        label_ids.append(0)\n",
    "\n",
    "    assert len(input_ids) == max_seq_length\n",
    "    assert len(input_mask) == max_seq_length\n",
    "    assert len(segment_ids) == max_seq_length\n",
    "    assert len(label_ids) == max_seq_length\n",
    "\n",
    "    return input_ids, input_mask, segment_ids, label_ids\n",
    "\n",
    "def convert_examples_to_features(tokenizer, examples, max_seq_length=256):\n",
    "    \"\"\"Convert a set of `InputExample`s to a list of `InputFeatures`.\"\"\"\n",
    "\n",
    "    input_ids, input_masks, segment_ids, labels = [], [], [], []\n",
    "    for example in tqdm_notebook(examples, desc=\"Converting examples to features\"):\n",
    "        input_id, input_mask, segment_id, label = convert_single_example(\n",
    "            tokenizer, example, max_seq_length\n",
    "        )\n",
    "        input_ids.append(input_id)\n",
    "        input_masks.append(input_mask)\n",
    "        segment_ids.append(segment_id)\n",
    "        labels.append(label)\n",
    "    return (\n",
    "        np.array(input_ids),\n",
    "        np.array(input_masks),\n",
    "        np.array(segment_ids),\n",
    "        np.array(labels),\n",
    "    )\n",
    "\n",
    "def convert_text_to_examples(texts, labels):\n",
    "    \"\"\"Create InputExamples\"\"\"\n",
    "    InputExamples = []\n",
    "    for text, label in zip(texts, labels):\n",
    "        InputExamples.append(\n",
    "            InputExample(guid=None, text_a=text, text_b=None, label=label)\n",
    "        )\n",
    "    return InputExamples"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "nbpresent": {
     "id": "c5d22fdb-2f46-44b2-9962-cea0cc6927a8"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:From /home/dbnet/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py:3632: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Colocations handled automatically by placer.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "W0527 23:36:38.126359 140718967228160 deprecation.py:323] From /home/dbnet/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py:3632: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Colocations handled automatically by placer.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:tensorflow:Saver not created because there are no variables in the graph to restore\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "I0527 23:36:39.406475 140718967228160 saver.py:1483] Saver not created because there are no variables in the graph to restore\n"
     ]
    }
   ],
   "source": [
    "# Instantiate tokenizer\n",
    "tokenizer = create_tokenizer_from_hub_module()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Alignment code (token -> tag)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Following suggestions from WordPiece tokenization part in https://github.com/google-research/bert, for POS tagging task, input sentences are converted to the following format:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [],
   "source": [
    "# orig_tokens = [\"John\", \"Johanson\", \"'s\",  \"house\"]\n",
    "# labels      = [\"NNP\",  \"NNP\",      \"POS\", \"NN\"]\n",
    "# bert_tokens = ['[CLS]\", \"john\", \"johan\", \"##son\", \"'\",   \"s\",  \"house\", \"[SEP]\"]\n",
    "# bert_labels = ['[CLS]\", \"NNP\",  \"NNP\",   \"##\",    \"POS\", \"##\", \"NN\",    \"[SEP]\"] "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "So, due to **Wordpiece** tokenizer we must take special care for the correct alignment of token to tag. This is done in function *convert_single_example(tokenizer, example, max_seq_length=256)* with the following piece of code:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "def bert_labels(labels):\n",
    "    train_label_bert = []\n",
    "    train_label_bert.append('-PAD-')\n",
    "    for i in labels:\n",
    "        train_label_bert.append(i)\n",
    "    train_label_bert.append('-PAD-')\n",
    "    print('BERT labels:', train_label_bert)    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "tokens_a = train_text[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "orig_to_tok_map = []              \n",
    "tokens = []\n",
    "segment_ids = []\n",
    "tokens.append(\"[CLS]\")\n",
    "segment_ids.append(0)\n",
    "orig_to_tok_map.append(len(tokens)-1)\n",
    "for token in tokens_a:\n",
    "    #orig_to_tok_map.append(len(tokens)) # keep first piece of tokenized term\n",
    "    tokens.extend(tokenizer.tokenize(token))\n",
    "    orig_to_tok_map.append(len(tokens)-1) # # keep last piece of tokenized term -->> gives better results!\n",
    "    segment_ids.append(0)\n",
    "tokens.append(\"[SEP]\")\n",
    "segment_ids.append(0)\n",
    "orig_to_tok_map.append(len(tokens)-1)\n",
    "input_ids = tokenizer.convert_tokens_to_ids([tokens[i] for i in orig_to_tok_map])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "nbpresent": {
     "id": "b31bd2f4-6a3b-4836-888a-5362ebdbc492"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original tokens: ['creative', 'commons', 'makes', 'no', 'warranties', 'regarding', 'the', 'information', 'provided', ',', 'and', 'disclaims', 'liability', 'for', 'damages', 'resulting', 'from', 'its', 'use', '.']\n"
     ]
    }
   ],
   "source": [
    "print('Original tokens:',tokens_a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "nbpresent": {
     "id": "b31bd2f4-6a3b-4836-888a-5362ebdbc492"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "BERT tokens: ['[CLS]', 'creative', 'commons', 'makes', 'no', 'warrant', '##ies', 'regarding', 'the', 'information', 'provided', ',', 'and', 'disc', '##lai', '##ms', 'liability', 'for', 'damages', 'resulting', 'from', 'its', 'use', '.', '[SEP]']\n"
     ]
    }
   ],
   "source": [
    "print('BERT tokens:',tokens)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "BERT labels: ['-PAD-', 'PROPN', 'PROPN', 'VERB', 'DET', 'NOUN', 'VERB', 'DET', 'NOUN', 'VERB', 'PUNCT', 'CCONJ', 'VERB', 'NOUN', 'ADP', 'NOUN', 'VERB', 'ADP', 'DET', 'NOUN', 'PUNCT', '-PAD-']\n"
     ]
    }
   ],
   "source": [
    "bert_labels(train_label[2])"
   ]
  },
  {
   "cell_type": "raw",
   "metadata": {},
   "source": [
    "So the allignment 'warranties' -> 'NOUN' becomes after alligning '##ies' -> 'NOUN'\n",
    "We tested the other (proposed in run_classifier.py) possible allignment 'warrant' -> 'NOUN', but with worse results!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[0, 1, 2, 3, 4, 6, 7, 8, 9, 10, 11, 12, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24]"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "orig_to_tok_map"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[101, 5541, 7674, 3084, 2053, 3111, 4953, 1996, 2592, 3024, 1010, 1998, 5244, 14000, 2005, 12394, 4525, 2013, 2049, 2224, 1012, 102]\n"
     ]
    }
   ],
   "source": [
    "print(input_ids)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Test the conversion of an example to BERT representation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "nbpresent": {
     "id": "a8524cd7-a390-455a-b94d-995aabcbfb02"
    }
   },
   "outputs": [],
   "source": [
    "\"\"\"Create InputExamples\"\"\"\n",
    "InputExamples = []\n",
    "for text, label in zip(train_text[0:1], train_label[0:1]):\n",
    "    InputExamples.append(\n",
    "        InputExample(guid=None, text_a=text, text_b=None, label=label)\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {
    "nbpresent": {
     "id": "da8ff012-152f-4533-975c-b23c3a63611c"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "05b407623dc24822bb052b4b7eb87354",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Converting examples to features', max=1, style=ProgressStyle(…"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "input_ids, input_masks, segment_ids, labels = [], [], [], []\n",
    "for example in tqdm_notebook(InputExamples, desc=\"Converting examples to features\"):\n",
    "    input_id, input_mask, segment_id, label = convert_single_example(\n",
    "        tokenizer, example, MAX_SEQUENCE_LENGTH+2\n",
    "    )\n",
    "    input_ids.append(input_id)\n",
    "    input_masks.append(input_mask)\n",
    "    segment_ids.append(segment_id)\n",
    "    labels.append(label)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Convert examples to BERT representation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 101,
   "metadata": {
    "nbpresent": {
     "id": "eb4915fc-babf-4bf9-8789-ab05ef24be45"
    }
   },
   "outputs": [],
   "source": [
    "# Convert data to InputExample format\n",
    "train_examples = convert_text_to_examples(train_text, train_label)\n",
    "test_examples = convert_text_to_examples(test_text, test_label)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 102,
   "metadata": {
    "nbpresent": {
     "id": "4f427047-23e0-4dd4-b8b7-333c93eb88ce"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "f565fbd6aa244d6ba93b83772f5a2987",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Converting examples to features', max=1951, style=ProgressSty…"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "4c13179fbdb74f6c8563b29c3e2e911a",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Converting examples to features', max=154, style=ProgressStyl…"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# Convert to features\n",
    "(train_input_ids, train_input_masks, train_segment_ids, train_labels_ids\n",
    ") = convert_examples_to_features(tokenizer, train_examples, max_seq_length=MAX_SEQUENCE_LENGTH+2)\n",
    "(test_input_ids, test_input_masks, test_segment_ids, test_labels_ids\n",
    ") = convert_examples_to_features(tokenizer, test_examples, max_seq_length=MAX_SEQUENCE_LENGTH+2)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## See the form of BERT representation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "72\n",
      "72\n",
      "72\n",
      "72\n"
     ]
    }
   ],
   "source": [
    "print(len(train_input_ids[0]))\n",
    "print(len(train_input_masks[0]))\n",
    "print(len(train_segment_ids[0]))\n",
    "print(len(train_labels_ids[0]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {
    "nbpresent": {
     "id": "e3ab4f77-4587-4fd0-8fbc-b06840e852f5"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 101, 4353, 1997, 2023, 6105, 2515, 2025, 3443, 2019, 4905, 1011,\n",
       "       7396, 3276, 1012,  102,    0,    0,    0,    0,    0,    0,    0,\n",
       "          0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "          0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "          0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "          0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "          0,    0,    0,    0,    0,    0])"
      ]
     },
     "execution_count": 45,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_input_ids[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {
    "nbpresent": {
     "id": "996d81bf-c1ce-4d09-8315-8d3c4512ddb8"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0,\n",
       "       0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
       "       0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
       "       0, 0, 0, 0, 0, 0])"
      ]
     },
     "execution_count": 46,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_input_masks[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {
    "nbpresent": {
     "id": "d16548f1-b1a4-4003-b90e-c28164d4012d"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
       "       0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
       "       0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
       "       0, 0, 0, 0, 0, 0])"
      ]
     },
     "execution_count": 47,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_segment_ids[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {
    "nbpresent": {
     "id": "4343b8c0-e6e3-4198-b209-968b28ec4fa4"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 0,  8,  2,  6,  8,  4, 10, 16,  6,  8, 13,  8,  8, 13,  0,  0,  0,\n",
       "        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,\n",
       "        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,\n",
       "        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,\n",
       "        0,  0,  0,  0])"
      ]
     },
     "execution_count": 48,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_labels_ids[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {
    "nbpresent": {
     "id": "080448a8-1f00-46b8-b4c9-e8aff53658ae"
    }
   },
   "outputs": [],
   "source": [
    "# One-hot encode labels\n",
    "train_labels = to_categorical(train_labels_ids, num_classes=n_tags)\n",
    "test_labels = to_categorical(test_labels_ids, num_classes=n_tags)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {
    "nbpresent": {
     "id": "2a4bc11b-a307-4121-8c42-44dde65ae4cf"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([ 101, 4353, 1997, 2023, 6105, 2515, 2025, 3443, 2019, 4905, 1011,\n",
       "        7396, 3276, 1012,  102,    0,    0,    0,    0,    0,    0,    0,\n",
       "           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,\n",
       "           0,    0,    0,    0,    0,    0]),\n",
       " array([[1., 0., 0., ..., 0., 0., 0.],\n",
       "        [0., 0., 0., ..., 0., 0., 0.],\n",
       "        [0., 0., 1., ..., 0., 0., 0.],\n",
       "        ...,\n",
       "        [1., 0., 0., ..., 0., 0., 0.],\n",
       "        [1., 0., 0., ..., 0., 0., 0.],\n",
       "        [1., 0., 0., ..., 0., 0., 0.]], dtype=float32))"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_input_ids[0], train_labels[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Model building"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {
    "nbpresent": {
     "id": "4098aad1-cdf7-4282-ba3b-70897cdcfe1d"
    }
   },
   "outputs": [],
   "source": [
    "class BertLayer(Layer):\n",
    "    def __init__(self, output_representation='sequence_output', trainable=True, **kwargs):\n",
    "        self.bert = None\n",
    "        super(BertLayer, self).__init__(**kwargs)\n",
    "        \n",
    "        self.trainable = trainable\n",
    "        self.output_representation = output_representation\n",
    "\n",
    "    def build(self, input_shape):\n",
    "        # SetUp tensorflow Hub module\n",
    "        self.bert = hub.Module(bert_path,\n",
    "                               trainable=self.trainable, \n",
    "                               name=\"{}_module\".format(self.name))\n",
    "\n",
    "        # Assign module's trainable weights to model\n",
    "        # Remove unused layers and set trainable parameters\n",
    "        # s = [\"/cls/\", \"/pooler/\", 'layer_11', 'layer_10', 'layer_9', 'layer_8', 'layer_7', 'layer_6']\n",
    "        s = [\"/cls/\", \"/pooler/\"]\n",
    "        self.trainable_weights += [var for var in self.bert.variables[:] if not any(x in var.name for x in s)]\n",
    "            \n",
    "        for var in self.bert.variables:\n",
    "            if var not in self._trainable_weights:\n",
    "                self._non_trainable_weights.append(var)\n",
    "                \n",
    "        # See Trainable Variables\n",
    "        #tf.logging.info(\"**** Trainable Variables ****\")\n",
    "        #for var in self.trainable_weights:\n",
    "        #    init_string = \", *INIT_FROM_CKPT*\"\n",
    "        #    tf.logging.info(\"  name = %s, shape = %s%s\", var.name, var.shape, init_string)\n",
    "            \n",
    "        print('Trainable weights:',len(self.trainable_weights))\n",
    "        super(BertLayer, self).build(input_shape)\n",
    "\n",
    "    def call(self, inputs, mask=None):\n",
    "        inputs = [K.cast(x, dtype=\"int32\") for x in inputs]\n",
    "        input_ids, input_mask, segment_ids = inputs\n",
    "        bert_inputs = dict(\n",
    "            input_ids=input_ids, input_mask=input_mask, segment_ids=segment_ids\n",
    "        )\n",
    "        result = self.bert(inputs=bert_inputs, signature=\"tokens\", as_dict=True)[\n",
    "            self.output_representation\n",
    "        ]\n",
    "        return result\n",
    "\n",
    "    def compute_mask(self, inputs, mask=None):\n",
    "        return K.not_equal(inputs[0], 0.0)   \n",
    "\n",
    "    def compute_output_shape(self, input_shape):\n",
    "        if self.output_representation == 'pooled_output':\n",
    "            return (None, 768)\n",
    "        else:\n",
    "            return (None, None, 768)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {
    "nbpresent": {
     "id": "177cc483-831b-4a10-b659-e1c6cf653877"
    }
   },
   "outputs": [],
   "source": [
    "#del model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {
    "nbpresent": {
     "id": "08a4002f-f656-47f2-814d-447e0bd56858"
    }
   },
   "outputs": [],
   "source": [
    "# Build model\n",
    "def build_model(max_seq_length):\n",
    "    seed = 0 \n",
    "    in_id = keras.layers.Input(shape=(max_seq_length,), name=\"input_ids\")\n",
    "    in_mask = keras.layers.Input(shape=(max_seq_length,), name=\"input_masks\")\n",
    "    in_segment = keras.layers.Input(shape=(max_seq_length,), name=\"segment_ids\")\n",
    "    bert_inputs = [in_id, in_mask, in_segment]\n",
    "    \n",
    "    np.random.seed(seed)\n",
    "    bert_output = BertLayer()(bert_inputs)\n",
    "\n",
    "    np.random.seed(seed)\n",
    "    outputs = keras.layers.Dense(n_tags, activation=keras.activations.softmax)(bert_output)\n",
    "\n",
    "    np.random.seed(seed)\n",
    "    model = keras.models.Model(inputs=bert_inputs, outputs=outputs)\n",
    "    np.random.seed(seed)\n",
    "    model.compile(optimizer=keras.optimizers.Adam(lr=0.00004), loss=keras.losses.categorical_crossentropy, metrics=['accuracy'])   \n",
    "    model.summary(100)\n",
    "    return model\n",
    "\n",
    "def initialize_vars(sess):\n",
    "    sess.run(tf.local_variables_initializer())\n",
    "    sess.run(tf.global_variables_initializer())\n",
    "    sess.run(tf.tables_initializer())\n",
    "    K.set_session(sess)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {
    "nbpresent": {
     "id": "c1391d48-0f3d-4b90-b1b4-4d8cae268050"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Trainable weights: 197\n",
      "INFO:tensorflow:Saver not created because there are no variables in the graph to restore\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "I0527 23:37:14.097841 140718967228160 saver.py:1483] Saver not created because there are no variables in the graph to restore\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "____________________________________________________________________________________________________\n",
      "Layer (type)                     Output Shape          Param #     Connected to                     \n",
      "====================================================================================================\n",
      "input_ids (InputLayer)           (None, 72)            0                                            \n",
      "____________________________________________________________________________________________________\n",
      "input_masks (InputLayer)         (None, 72)            0                                            \n",
      "____________________________________________________________________________________________________\n",
      "segment_ids (InputLayer)         (None, 72)            0                                            \n",
      "____________________________________________________________________________________________________\n",
      "bert_layer_1 (BertLayer)         (None, None, 768)     110104890   input_ids[0][0]                  \n",
      "                                                                   input_masks[0][0]                \n",
      "                                                                   segment_ids[0][0]                \n",
      "____________________________________________________________________________________________________\n",
      "dense_1 (Dense)                  (None, None, 18)      13842       bert_layer_1[0][0]               \n",
      "====================================================================================================\n",
      "Total params: 110,118,732\n",
      "Trainable params: 108,905,490\n",
      "Non-trainable params: 1,213,242\n",
      "____________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = build_model(MAX_SEQUENCE_LENGTH+2) # We sum 2 for [CLS], [SEP] tokens"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 55,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "plot_model(model, to_file='model.png', show_shapes=True)\n",
    "Image('model.png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {
    "nbpresent": {
     "id": "e5b1510d-5965-4f77-8c1c-a019bb081441"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1951, 72)"
      ]
     },
     "execution_count": 56,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_input_ids.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {
    "nbpresent": {
     "id": "e2ced586-54d2-4f87-87b4-11447a074836"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1951, 72)"
      ]
     },
     "execution_count": 57,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_input_masks.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {
    "nbpresent": {
     "id": "d4e8c809-bc83-428b-824d-4640127d0a6f"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1951, 72)"
      ]
     },
     "execution_count": 58,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_segment_ids.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1951, 72, 18)"
      ]
     },
     "execution_count": 59,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_labels.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Model training (BERT fine-tuning)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "metadata": {},
   "outputs": [],
   "source": [
    "initialize_vars(sess)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 83,
   "metadata": {
    "nbpresent": {
     "id": "fc7dd289-a946-479a-838e-e25e6384f16a"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 1951 samples, validate on 154 samples\n",
      "Epoch 1/30\n",
      "1951/1951 [==============================] - 34s 17ms/step - loss: 0.4391 - acc: 0.8772 - val_loss: 0.1254 - val_acc: 0.9651\n",
      "\n",
      "Epoch 00001: val_acc improved from -inf to 0.96510, saving model to bert_tagger.h5\n",
      "Epoch 2/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0719 - acc: 0.9800 - val_loss: 0.1186 - val_acc: 0.9682\n",
      "\n",
      "Epoch 00002: val_acc improved from 0.96510 to 0.96819, saving model to bert_tagger.h5\n",
      "Epoch 3/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0297 - acc: 0.9915 - val_loss: 0.1298 - val_acc: 0.9704\n",
      "\n",
      "Epoch 00003: val_acc improved from 0.96819 to 0.97043, saving model to bert_tagger.h5\n",
      "Epoch 4/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0120 - acc: 0.9968 - val_loss: 0.1495 - val_acc: 0.9701\n",
      "\n",
      "Epoch 00004: val_acc did not improve from 0.97043\n",
      "Epoch 5/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0057 - acc: 0.9988 - val_loss: 0.1654 - val_acc: 0.9700\n",
      "\n",
      "Epoch 00005: val_acc did not improve from 0.97043\n",
      "Epoch 6/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0032 - acc: 0.9992 - val_loss: 0.1809 - val_acc: 0.9709\n",
      "\n",
      "Epoch 00006: val_acc improved from 0.97043 to 0.97089, saving model to bert_tagger.h5\n",
      "Epoch 7/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0034 - acc: 0.9994 - val_loss: 0.2043 - val_acc: 0.9630\n",
      "\n",
      "Epoch 00007: val_acc did not improve from 0.97089\n",
      "Epoch 8/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0038 - acc: 0.9993 - val_loss: 0.1772 - val_acc: 0.9721\n",
      "\n",
      "Epoch 00008: val_acc improved from 0.97089 to 0.97208, saving model to bert_tagger.h5\n",
      "Epoch 9/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0014 - acc: 0.9998 - val_loss: 0.1824 - val_acc: 0.9744\n",
      "\n",
      "Epoch 00009: val_acc improved from 0.97208 to 0.97440, saving model to bert_tagger.h5\n",
      "Epoch 10/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 0.0011 - acc: 0.9998 - val_loss: 0.1866 - val_acc: 0.9718\n",
      "\n",
      "Epoch 00010: val_acc did not improve from 0.97440\n",
      "Epoch 11/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 7.8762e-04 - acc: 0.9999 - val_loss: 0.1970 - val_acc: 0.9717\n",
      "\n",
      "Epoch 00011: val_acc did not improve from 0.97440\n",
      "Epoch 12/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 5.2191e-04 - acc: 0.9999 - val_loss: 0.2039 - val_acc: 0.9713\n",
      "\n",
      "Epoch 00012: val_acc did not improve from 0.97440\n",
      "Epoch 13/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 2.8219e-04 - acc: 0.9999 - val_loss: 0.2116 - val_acc: 0.9726\n",
      "\n",
      "Epoch 00013: val_acc did not improve from 0.97440\n",
      "Epoch 14/30\n",
      "1951/1951 [==============================] - 19s 10ms/step - loss: 4.8439e-04 - acc: 0.9999 - val_loss: 0.2118 - val_acc: 0.9713\n",
      "\n",
      "Epoch 00014: val_acc did not improve from 0.97440\n",
      "Training completed in 298.059595 seconds\n"
     ]
    }
   ],
   "source": [
    "t_ini = datetime.datetime.now()\n",
    "\n",
    "cp = ModelCheckpoint(filepath=\"bert_tagger.h5\",\n",
    "                     monitor='val_acc',\n",
    "                     save_best_only=True,\n",
    "                     save_weights_only=True,\n",
    "                     verbose=1)\n",
    "\n",
    "early_stopping = EarlyStopping(monitor = 'val_acc', patience = 5)\n",
    "\n",
    "history = model.fit([train_input_ids, train_input_masks, train_segment_ids], \n",
    "                    train_labels,\n",
    "                    validation_data=([test_input_ids, test_input_masks, test_segment_ids], test_labels),\n",
    "                    #validation_split=0.3,\n",
    "                    epochs=EPOCHS,\n",
    "                    batch_size=16,\n",
    "                    shuffle=True,\n",
    "                    verbose=1,\n",
    "                    callbacks=[cp, early_stopping]\n",
    "                   ) \n",
    "\n",
    "t_fin = datetime.datetime.now()\n",
    "print('Training completed in {} seconds'.format((t_fin - t_ini).total_seconds()))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Classification report"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 103,
   "metadata": {
    "nbpresent": {
     "id": "ce12f232-8d4a-4093-9977-88402c6a80f4"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Trainable weights: 197\n",
      "INFO:tensorflow:Saver not created because there are no variables in the graph to restore\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "I0527 23:46:49.031078 140718967228160 saver.py:1483] Saver not created because there are no variables in the graph to restore\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "____________________________________________________________________________________________________\n",
      "Layer (type)                     Output Shape          Param #     Connected to                     \n",
      "====================================================================================================\n",
      "input_ids (InputLayer)           (None, 72)            0                                            \n",
      "____________________________________________________________________________________________________\n",
      "input_masks (InputLayer)         (None, 72)            0                                            \n",
      "____________________________________________________________________________________________________\n",
      "segment_ids (InputLayer)         (None, 72)            0                                            \n",
      "____________________________________________________________________________________________________\n",
      "bert_layer_3 (BertLayer)         (None, None, 768)     110104890   input_ids[0][0]                  \n",
      "                                                                   input_masks[0][0]                \n",
      "                                                                   segment_ids[0][0]                \n",
      "____________________________________________________________________________________________________\n",
      "dense_3 (Dense)                  (None, None, 18)      13842       bert_layer_3[0][0]               \n",
      "====================================================================================================\n",
      "Total params: 110,118,732\n",
      "Trainable params: 108,905,490\n",
      "Non-trainable params: 1,213,242\n",
      "____________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = build_model(MAX_SEQUENCE_LENGTH+2) \n",
    "model.load_weights('bert_tagger.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 104,
   "metadata": {
    "nbpresent": {
     "id": "615aa234-1351-4d3e-84ff-379d1fc4f5cf"
    }
   },
   "outputs": [],
   "source": [
    "y_pred = model.predict([test_input_ids, test_input_masks, test_segment_ids]).argmax(-1)\n",
    "y_true = test_labels.argmax(-1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 105,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 0,  8, 13, 12,  9, 13,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,\n",
       "        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,\n",
       "        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,\n",
       "        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,\n",
       "        0,  0,  0,  0])"
      ]
     },
     "execution_count": 105,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "y_true[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 106,
   "metadata": {
    "nbpresent": {
     "id": "4d68934e-3464-4fa5-bdfe-33f1bee74efc"
    }
   },
   "outputs": [],
   "source": [
    "def y2label(zipped, mask=0):\n",
    "    out_true = []\n",
    "    out_pred = []\n",
    "    for zip_i in zipped:\n",
    "        a, b = tuple(zip_i)\n",
    "        if a != mask :\n",
    "            out_true.append(int2tag[a])\n",
    "            out_pred.append(int2tag[b])\n",
    "    return out_true, out_pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 107,
   "metadata": {
    "nbpresent": {
     "id": "6412a274-6f2e-4c27-9fb2-0fd0916ab42a"
    }
   },
   "outputs": [],
   "source": [
    "y_zipped = zip(y_true.flat, y_pred.flat)\n",
    "y_true, y_pred = y2label(y_zipped)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 108,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(3408, 3408)"
      ]
     },
     "execution_count": 108,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(y_true), len(y_pred)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 109,
   "metadata": {
    "nbpresent": {
     "id": "0e51f9f4-b16f-4e14-a204-8167e2897848"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "------------ Result of Bert fine-tuned model ----------\n",
      "\n",
      "              precision    recall  f1-score   support\n",
      "\n",
      "         ADJ     0.9217    0.8929    0.9070       224\n",
      "         ADP     0.9698    0.9857    0.9776       488\n",
      "         ADV     0.9597    0.9084    0.9333       131\n",
      "         AUX     0.9957    0.9957    0.9957       234\n",
      "       CCONJ     0.9896    0.9896    0.9896        96\n",
      "         DET     0.9977    1.0000    0.9989       439\n",
      "        INTJ     1.0000    1.0000    1.0000         2\n",
      "        NOUN     0.9699    0.9841    0.9769       753\n",
      "         NUM     0.9385    1.0000    0.9683        61\n",
      "        PART     0.9565    1.0000    0.9778        66\n",
      "        PRON     1.0000    0.9434    0.9709       106\n",
      "       PROPN     0.9625    0.8556    0.9059        90\n",
      "       PUNCT     1.0000    1.0000    1.0000       339\n",
      "       SCONJ     0.9333    0.8235    0.8750        51\n",
      "        VERB     0.9521    0.9755    0.9636       326\n",
      "           X     0.6667    1.0000    0.8000         2\n",
      "\n",
      "   micro avg     0.9727    0.9727    0.9727      3408\n",
      "   macro avg     0.9509    0.9596    0.9525      3408\n",
      "weighted avg     0.9727    0.9727    0.9724      3408\n",
      "\n"
     ]
    }
   ],
   "source": [
    "name='Bert fine-tuned model'\n",
    "print('\\n------------ Result of {} ----------\\n'.format(name))\n",
    "print(classification_report(y_true, y_pred, digits=4))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 110,
   "metadata": {
    "nbpresent": {
     "id": "0e51f9f4-b16f-4e14-a204-8167e2897848"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.9727\n",
      "f1-macro score: 0.9525\n"
     ]
    }
   ],
   "source": [
    "print(\"Accuracy: {0:.4f}\".format(accuracy_score(y_true, y_pred)))\n",
    "print('f1-macro score: {0:.4f}'.format(f1_score(y_true, y_pred, average='macro')))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Confusion matrix"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 111,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x432 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "tags = sorted(set(y_pred+y_true))\n",
    "cnf_matrix = confusion_matrix(y_true, y_pred)\n",
    "plot_confusion_matrix(f1_score(y_true, y_pred, average='macro'), cnf_matrix, target_names=tags, title=name, normalize=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Make a prediction for a test sample:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 112,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_pred = model.predict([test_input_ids, test_input_masks, test_segment_ids], batch_size=16).argmax(-1)\n",
    "y_true = test_labels_ids\n",
    "\n",
    "def make_prediction(i=16):\n",
    "    note = ''\n",
    "    sent = []\n",
    "    print(\"{:10} {:5} : {:5}\".format(\"Word\", \"True\", \"Predicted\"))\n",
    "    print(35*'-')\n",
    "    for w, true, pred in zip(test_input_ids[i], y_true[i], y_pred[i]):\n",
    "        if tokenizer.convert_ids_to_tokens([w])[0]!='[PAD]' and \\\n",
    "            tokenizer.convert_ids_to_tokens([w])[0]!='[CLS]' and \\\n",
    "            tokenizer.convert_ids_to_tokens([w])[0]!='[SEP]':\n",
    "            if int2tag[true] != int2tag[pred]: note='<<--- Error!'\n",
    "            print(\"{:10} {:5} : {:5} {:5}\".format(tokenizer.convert_ids_to_tokens([w])[0], int2tag[true], int2tag[pred], note))\n",
    "            note=''"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 113,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Word       True  : Predicted\n",
      "-----------------------------------\n",
      "any        DET   : DET        \n",
      "use        NOUN  : NOUN       \n",
      "of         ADP   : ADP        \n",
      "the        DET   : DET        \n",
      "work       NOUN  : NOUN       \n",
      "other      ADJ   : ADJ        \n",
      "than       SCONJ : ADP   <<--- Error!\n",
      "as         ADP   : ADP        \n",
      "authorized VERB  : VERB       \n",
      "under      ADP   : ADP        \n",
      "this       DET   : DET        \n",
      "license    NOUN  : NOUN       \n",
      "or         CCONJ : CCONJ      \n",
      "copyright  NOUN  : NOUN       \n",
      "law        NOUN  : NOUN       \n",
      "is         AUX   : AUX        \n",
      "prohibited VERB  : VERB       \n",
      ".          PUNCT : PUNCT      \n"
     ]
    }
   ],
   "source": [
    "make_prediction(i=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Tag an unknown sentence"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 114,
   "metadata": {},
   "outputs": [],
   "source": [
    "sentence_raw = 'Word embeddings provide a dense representation of words and their relative meanings.'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 115,
   "metadata": {},
   "outputs": [],
   "source": [
    "sentence_ini = nltk.word_tokenize(sentence_raw.lower())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Convert the sentence to BERT representation:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 116,
   "metadata": {},
   "outputs": [],
   "source": [
    "sentence_bert = tokenizer.tokenize(sentence_raw)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 117,
   "metadata": {},
   "outputs": [],
   "source": [
    "tokens_a = sentence_ini"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 118,
   "metadata": {},
   "outputs": [],
   "source": [
    "orig_to_tok_map = []              \n",
    "tokens = []\n",
    "segment_ids = []\n",
    "tokens.append(\"[CLS]\")\n",
    "segment_ids.append(0)\n",
    "orig_to_tok_map.append(len(tokens)-1)\n",
    "for token in tokens_a:\n",
    "    #orig_to_tok_map.append(len(tokens)) # keep first piece of tokenized term\n",
    "    tokens.extend(tokenizer.tokenize(token))\n",
    "    orig_to_tok_map.append(len(tokens)-1) # # keep last piece of tokenized term -->> gives better results!\n",
    "    segment_ids.append(0)\n",
    "tokens.append(\"[SEP]\")\n",
    "segment_ids.append(0)\n",
    "orig_to_tok_map.append(len(tokens)-1)\n",
    "input_ids = tokenizer.convert_tokens_to_ids([tokens[i] for i in orig_to_tok_map])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 119,
   "metadata": {
    "nbpresent": {
     "id": "b31bd2f4-6a3b-4836-888a-5362ebdbc492"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original tokens: ['word', 'embeddings', 'provide', 'a', 'dense', 'representation', 'of', 'words', 'and', 'their', 'relative', 'meanings', '.']\n"
     ]
    }
   ],
   "source": [
    "print('Original tokens:',tokens_a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 120,
   "metadata": {
    "nbpresent": {
     "id": "b31bd2f4-6a3b-4836-888a-5362ebdbc492"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "BERT tokens: ['[CLS]', 'word', 'em', '##bed', '##ding', '##s', 'provide', 'a', 'dense', 'representation', 'of', 'words', 'and', 'their', 'relative', 'meanings', '.', '[SEP]']\n"
     ]
    }
   ],
   "source": [
    "print('BERT tokens:',tokens)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 121,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[0, 1, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]"
      ]
     },
     "execution_count": 121,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "orig_to_tok_map"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 122,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['[CLS]',\n",
       " 'word',\n",
       " '##s',\n",
       " 'provide',\n",
       " 'a',\n",
       " 'dense',\n",
       " 'representation',\n",
       " 'of',\n",
       " 'words',\n",
       " 'and',\n",
       " 'their',\n",
       " 'relative',\n",
       " 'meanings',\n",
       " '.',\n",
       " '[SEP]']"
      ]
     },
     "execution_count": 122,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "[tokens[i] for i in orig_to_tok_map]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 123,
   "metadata": {
    "nbpresent": {
     "id": "eb4915fc-babf-4bf9-8789-ab05ef24be45"
    }
   },
   "outputs": [],
   "source": [
    "# Convert data to InputExample format\n",
    "test_example = convert_text_to_examples([sentence_ini], [['-PAD-']*len(sentence_ini)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 124,
   "metadata": {
    "nbpresent": {
     "id": "4f427047-23e0-4dd4-b8b7-333c93eb88ce"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "9a2ca2fe61fe4f1eb193f663b028fa51",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Converting examples to features', max=1, style=ProgressStyle(…"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "# Convert to features\n",
    "(input_ids, input_masks, segment_ids, _\n",
    ") = convert_examples_to_features(tokenizer, test_example, max_seq_length=MAX_SEQUENCE_LENGTH+2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 125,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "Word in BERT layer  | Initial word   : Predicted POS-tag\n",
      "-------------------------------------------------------------\n",
      "word                | word           : NOUN           \n",
      "##s                 | embeddings     : NOUN           \n",
      "provide             | provide        : VERB           \n",
      "a                   | a              : DET            \n",
      "dense               | dense          : ADJ            \n",
      "representation      | representation : NOUN           \n",
      "of                  | of             : ADP            \n",
      "words               | words          : NOUN           \n",
      "and                 | and            : CCONJ          \n",
      "their               | their          : DET            \n",
      "relative            | relative       : ADJ            \n",
      "meanings            | meanings       : NOUN           \n",
      ".                   | .              : PUNCT          \n"
     ]
    }
   ],
   "source": [
    "predictions = model.predict([input_ids, input_masks, segment_ids], batch_size=1).argmax(-1)[0]\n",
    "print(\"\\n{:20}| {:15}: {:15}\".format(\"Word in BERT layer\", 'Initial word', \"Predicted POS-tag\"))\n",
    "print(61*'-')\n",
    "k = 0\n",
    "for i, pred in enumerate(predictions):\n",
    "    try:\n",
    "        if pred!=0:\n",
    "            print(\"{:20}| {:15}: {:15}\".format([tokens[i] for i in orig_to_tok_map][i], sentence_ini[i-1], int2tag[pred]))            \n",
    "            k+=1            \n",
    "    except:\n",
    "        pass"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Frequent types of mistakes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 126,
   "metadata": {},
   "outputs": [],
   "source": [
    "def find_errors(X,y):\n",
    "    error_counter = collections.Counter()\n",
    "    support = 0\n",
    "    for i in range(test_input_ids.shape[0]):\n",
    "        for w, true, pred in zip(test_input_ids[i], y_true[i], y_pred[i]):\n",
    "            if int2tag[true]!='-PAD-':\n",
    "                if true != pred:\n",
    "                    word = tokenizer.convert_ids_to_tokens([w])[0]\n",
    "                    error_counter[word] += 1\n",
    "                support += 1\n",
    "    return error_counter, support"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 127,
   "metadata": {},
   "outputs": [],
   "source": [
    "errors, support = find_errors([test_input_ids],y_true)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total errors/Total words: 93/3408 | Accuracy: 0.9727\n",
      "\n",
      "Most common errors: [('as', 8), ('appointed', 5), ('to', 4), ('part', 3), ('justice', 3), ('general', 3), ('than', 2), ('contracting', 2), ('base', 2), ('around', 2), ('##s', 1), ('##r', 1), ('is', 1), ('that', 1), ('time', 1), ('set', 1), ('interested', 1), ('traveling', 1), ('japanese', 1), ('by', 1)]\n"
     ]
    }
   ],
   "source": [
    "print('Total errors/Total words: {}/{} | Accuracy: {:.4}\\n'.format(sum(errors.values()), support, 1-sum(errors.values())/support))\n",
    "print('Most common errors:', errors.most_common(20))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Learning curves"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 104,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_acc()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 105,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_loss()"
   ]
  }
 ],
 "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": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}