notebooks/Openstates_Classifier_Demo.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Introduction\n",
    "\n",
    "This notebook provides an example of using a text classification model with Openstates Bills data.\n",
    "\n",
    "It does the following:\n",
    "\n",
    "- Reads in data for Bills in Rhode Island as extracted with Open states [GraphQL](https://docs.openstates.org/en/latest/api/v2/index.html) API, extended to include scraped text. The data file was created using notebook `./scrapers/Openstates_GraphQL_API_extract.ipynb`\n",
    "- Using the 'subject' field - which contains a list of categories for each bill - identifies bills which fall into health-related categories and sets a meta category called 'Health related', being 0 or 1 for each Bill\n",
    "- Using the bill title text, trains a [Fasttext](https://github.com/facebookresearch/fastText) text classifier to predict if a bill is in the health-related category\n",
    "- The model is then used to predict this category on Bills from another state\n",
    "\n",
    "This is a simple example trained on a small amount of data using only bill titles, but with more data and using full bill text the same approach can be used for classifying more complex catagories organizations might be interested in tracking. A model trained this way can be used to assign consistent categories across all states.\n",
    "\n",
    "# Setup\n",
    "\n",
    "The notebook will automatically download required datafiles from Amazon s3. For reference, these files are:\n",
    "\n",
    "1. [rhode_island.pkl](https://datakind-public.s3.amazonaws.com/NLP_data/rhode_island.pkl.zip)\n",
    "2. [Bills_V2.zip](https://datakind-public.s3.amazonaws.com/NLP_data/Bills_V2.zip)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Helper functions\n",
    "\n",
    "The following functions are some useful text pre-processing, model training, prediction and visualization."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Requirement already satisfied: en_core_web_sm==2.1.0 from https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.1.0/en_core_web_sm-2.1.0.tar.gz#egg=en_core_web_sm==2.1.0 in /opt/conda/lib/python3.6/site-packages (2.1.0)\n",
      "\u001b[38;5;2m✔ Download and installation successful\u001b[0m\n",
      "You can now load the model via spacy.load('en_core_web_sm')\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "[nltk_data] Downloading package punkt to /home/jovyan/nltk_data...\n",
      "[nltk_data]   Package punkt is already up-to-date!\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Done\n"
     ]
    }
   ],
   "source": [
    "import pandas as pd\n",
    "import os\n",
    "import sys\n",
    "import pathlib\n",
    "import json\n",
    "import re\n",
    "import pickle\n",
    "from sklearn.model_selection import train_test_split\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "import seaborn as sns\n",
    "from sklearn.metrics import confusion_matrix\n",
    "import numpy as np\n",
    "import itertools\n",
    "from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n",
    "import traceback\n",
    "import subprocess\n",
    "import logging\n",
    "import requests\n",
    "from zipfile import ZipFile\n",
    "from io import BytesIO\n",
    "!{sys.executable} -m spacy download en_core_web_sm\n",
    "import en_core_web_sm\n",
    "\n",
    "from sumy.nlp.tokenizers import Tokenizer\n",
    "from sumy.summarizers.lsa import LsaSummarizer\n",
    "from sumy.parsers.plaintext import PlaintextParser\n",
    "from sumy.summarizers.lex_rank import LexRankSummarizer\n",
    "import nltk\n",
    "nltk.download('punkt')\n",
    "\n",
    "nlp = en_core_web_sm.load()\n",
    "nlp.max_length = 2000000\n",
    "\n",
    "#\n",
    "# Combine text and labels into a format Fasttext needs\n",
    "#\n",
    "def df_to_ft(df, text_field, label_field, file_name, label_format=\"__label__{0}\"):\n",
    "    \"\"\"\n",
    "        Takes a dataframe and writes to a Fastext compliant format.\n",
    "        \n",
    "        Inputs:\n",
    "        \n",
    "        df             - The dataframe\n",
    "        text_field     - The column name of the field with Text that Fasttext will use\n",
    "        label_field    - The column name of the category Fasttext will predict (0/1)\n",
    "        file_name      - Name of file to write data\n",
    "        label_format   - Label format in the Fasttext input file\n",
    "        \n",
    "        Outputs:\n",
    "        \n",
    "        A file in the right format for Fasttext.\n",
    "        \n",
    "    \"\"\"\n",
    "    labels = df[label_field].apply(lambda val: label_format.format(int(val)) + \" \")\n",
    "    text = df[text_field]\n",
    "    combined = labels.str.cat(text)\n",
    "    print(f\"Writing {len(combined)} lines\")\n",
    "    with open(file_name, 'w') as output:\n",
    "        for i, val in combined.iteritems():\n",
    "            output.write(val + \"\\n\")\n",
    "#\n",
    "# Train a Fasttext model\n",
    "#\n",
    "def train_fastText(train, text_field, ddir, vecs, lr, pretrainedVectors, epoch, wordNgrams, \n",
    "                   model_name, threads, quantize, response_field):\n",
    "    \n",
    "    \"\"\"\n",
    "        Wrapper function to train fasttext model\n",
    "        \n",
    "        Inputs:\n",
    "        \n",
    "        train             - Training data frame\n",
    "        ddir              - Directory to output any Fasttext files\n",
    "        vecs              - The number of vectors to use in the Fasttext model\n",
    "        pretrainedVectors - (Optional) Pre-trained vector file. WIll try and download if not available.\n",
    "        epoch             - Number of epochs to train the model\n",
    "        wordNgrams        - NGrams to use (1,2,3 etc)\n",
    "        model_name        - Name of output model file\n",
    "        threads           - Number of threads to use\n",
    "        quantize          - If a (smaller) quantized model file should be produced\n",
    "        response_field    - The field being predicted\n",
    "        \n",
    "        Outputs:\n",
    "        \n",
    "        A Fasttext model file\n",
    "        \n",
    "    \"\"\"\n",
    "    \n",
    "    print(\"Creating fasttext training file\")\n",
    "    input_file = ddir + \"/fasttext.txt\"\n",
    "    df_to_ft(train, text_field, response_field, input_file)\n",
    "    \n",
    "    exec_str = '/home/jovyan/fastText/fasttext supervised -input ' + input_file + ' -dim ' + str(vecs) + \\\n",
    "                  ' -lr ' + str(lr) + ' -epoch ' + str(epoch) + ' -wordNgrams ' + str(wordNgrams) + \\\n",
    "                  ' -output ' + ddir + '/' + model_name \n",
    "    if pretrainedVectors != '':\n",
    "        # Get some pretrained vectors from 'https://fasttext.cc/docs/en/english-vectors.html'\n",
    "        if not os.path.isfile(ddir + pretrainedVectors):\n",
    "            url = 'https://dl.fbaipublicfiles.com/fasttext/vectors-english/' + pretrainedVectors + '.zip'\n",
    "            print(\"Downloading \" + url + \" ...\")\n",
    "            resp = requests.get(url).content\n",
    "            zipfile = ZipFile(BytesIO(resp))\n",
    "            zipfile.extractall(ddir)\n",
    "            print(\"Done\")\n",
    "        else:\n",
    "            print(\"Already have prebuilt vector file, no need to download\")\n",
    "        exec_str = exec_str + ' -pretrainedVectors ' + ddir + pretrainedVectors\n",
    "        \n",
    "    # For reproducibility, remove for better performance\n",
    "    if threads != '':\n",
    "        exec_str = exec_str + \" -thread \" + threads\n",
    "    print(exec_str)\n",
    "    !{exec_str}\n",
    "    if quantize:\n",
    "        exec_str = '/home/jovyan/fastText/fasttext quantize -input ' + input_file + ' -dim ' + str(vecs) + \\\n",
    "                  ' -lr ' + str(lr) + ' -epoch ' + str(epoch) + ' -wordNgrams ' + str(wordNgrams) + \\\n",
    "                  ' -output ' + ddir + '/' + model_name \n",
    "        print(exec_str)\n",
    "        !{exec_str}\n",
    "        \n",
    "#\n",
    "# Make a prediction using Fasttext\n",
    "#\n",
    "def predict_fasttext(df, text_field, model_name):\n",
    "    \n",
    "    \"\"\"\n",
    "        Wrapper function to make Fassttext prediction\n",
    "        \n",
    "        Inputs:\n",
    "        \n",
    "        df                - Date frame\n",
    "        text_field        - Column name of text field\n",
    "        model_name        - Full path to trained Fasttext model file\n",
    "        \n",
    "        Outputs:\n",
    "        \n",
    "        df                - Input data frame, but with new columsn 'predict' for prediction and 'probability'\n",
    "        \n",
    "    \"\"\"\n",
    "    \n",
    "    # Location of fastText binary, adjust as appropriate\n",
    "    ft = str(pathlib.Path.home()) + \"/fastText/fasttext\"\n",
    "    \n",
    "    # Input file\n",
    "    with open( ddir + '/input.txt', 'w') as output:\n",
    "        for item in df[text_field]:\n",
    "            output.write(\"%s\\n\" % item)\n",
    "        \n",
    "    # Run prediction\n",
    "    print(model_name)\n",
    "    response = subprocess.run([ft,\"predict-prob\",\n",
    "                              ddir + \"/\" + model_name,\n",
    "                              ddir + \"/input.txt\"], stdout=subprocess.PIPE)\n",
    "    \n",
    "    response = response.stdout.decode('utf-8').split(\"\\n\")\n",
    "    df['ft_raw'] = response[0:len(response)-1]\n",
    "            \n",
    "    df['prediction'], df['probability'] = df['ft_raw'].str.split(' ', 1).str\n",
    "    df['probability'] = df['probability'].apply(float)\n",
    "    \n",
    "    def map_response(x):\n",
    "        # Translate fasttext labels to something human\n",
    "        if \"__label__1\" in x:\n",
    "            response = 1.0\n",
    "        else:\n",
    "            response = 0.0\n",
    "        return response\n",
    "    \n",
    "    df['prediction'] = df['prediction'].apply(map_response)\n",
    "\n",
    "    return df\n",
    "    \n",
    "def plot_confusion_matrix(cm, classes,\n",
    "                          normalize=False,\n",
    "                          title='Confusion matrix',\n",
    "                          cmap=plt.cm.Blues):\n",
    "    \"\"\"\n",
    "    This function prints and plots the confusion matrix.\n",
    "    Normalization can be applied by setting `normalize=True`.\n",
    "    \n",
    "    See: https://scikit-learn.org/stable/auto_examples/model_selection/plot_confusion_matrix.html for more details\n",
    "    \"\"\"\n",
    "    \n",
    "    if normalize:\n",
    "        cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]\n",
    "        print(\"Normalized confusion matrix\")\n",
    "    else:\n",
    "        print('Confusion matrix, without normalization')\n",
    "\n",
    "    plt.imshow(cm, interpolation='nearest', cmap=cmap)\n",
    "    plt.title(title)\n",
    "    plt.colorbar()\n",
    "    tick_marks = np.arange(len(classes))\n",
    "    plt.xticks(tick_marks, classes, rotation=45)\n",
    "    plt.yticks(tick_marks, classes)\n",
    "\n",
    "    fmt = '.2f' if normalize else 'd'\n",
    "    thresh = cm.max() / 2.\n",
    "    for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):\n",
    "        plt.text(j, i, format(cm[i, j], fmt),\n",
    "                 horizontalalignment=\"center\",\n",
    "                 color=\"white\" if cm[i, j] > thresh else \"black\")\n",
    "\n",
    "    plt.ylabel('True label')\n",
    "    plt.xlabel('Predicted label')\n",
    "    plt.tight_layout()    \n",
    "\n",
    "def output_metrics(actual,predicted, orig_len, response):\n",
    "    \n",
    "    \"\"\"\n",
    "    Wrapper to output useful prediction metrics \n",
    "    \n",
    "    Inputs:\n",
    "        \n",
    "    actual            - Vector of actual (0/1)\n",
    "    predicted         - Vector of predicted (0/1)\n",
    "    orig_len          - Vector length before any filtering was applied \n",
    "    response          - Name of response variable\n",
    "\n",
    "    Outputs:\n",
    "\n",
    "    Prints confusion matrix and acfuracy, precision, recal and f1.\n",
    "        \n",
    "    \"\"\"\n",
    "   \n",
    "    # Compute confusion matrix\n",
    "    cnf_matrix = confusion_matrix(predicted,actual)\n",
    "    np.set_printoptions(precision=2)\n",
    "\n",
    "    # Plot non-normalized confusion matrix\n",
    "    plt.figure()\n",
    "    plot_confusion_matrix(cnf_matrix, classes=['NOT ' + response,response],\n",
    "                          title='Confusion matrix, without normalization')\n",
    "\n",
    "    # Plot normalized confusion matrix\n",
    "    plt.figure()\n",
    "    plot_confusion_matrix(cnf_matrix, classes=['NOT ' + response,response], normalize=True,\n",
    "                          title='Normalized confusion matrix')\n",
    "\n",
    "    plt.show()\n",
    "    \n",
    "    # Accuracy if we had a dumb model predicting everything as sojo\n",
    "    all_positives = [1] * len(actual)\n",
    "    baseline_all_sojo = accuracy_score(all_positives, actual)\n",
    "    \n",
    "    print(\"\\n\\n=========== PREDICTION =============\")\n",
    "    test_no = len(actual)\n",
    "    print(\"Number of records: \" + str(test_no) + \" [\" + str(orig_len-test_no) + \", \" + \\\n",
    "                       str(round(100*(1-test_no/orig_len),2)) + \\\n",
    "                       \"% filtered out]\")\n",
    "    accuracy = accuracy_score(predicted, actual)\n",
    "    print(\"prediction accuracy: %.2f%%\" % (accuracy * 100.0) + \\\n",
    "          \" [vs. \" + str(round(baseline_all_sojo*100,2)) +\"% if assumed all \" + response +\" ]\")\n",
    "    recall = recall_score(predicted, actual)\n",
    "    print(\"prediction recall: %.2f%%\" % (recall * 100.0))\n",
    "    precision = precision_score(predicted, actual)\n",
    "    print(\"prediction precision: %.2f%%\" % (precision * 100.0))\n",
    "    f1 = f1_score(predicted, actual)\n",
    "    print(\"prediction f1: %.2f%%\" % (f1 * 100.0))\n",
    "    print(\"===================================\\n\\n\")\n",
    "    \n",
    "# Text preprocessor\n",
    "def preprocessFastText(df, text_field, removeNumbers = False, removeStopWords=True, charLengthMin=1, \\\n",
    "                      lemmatize=True):\n",
    "    \n",
    "    \"\"\"\n",
    "    Text pre-processing function, initially based on https://www.kaggle.com/mschumacher/using-fasttext-models-for-robust-embeddings,\n",
    "    extended for stopwords and lemmatization\n",
    "    \n",
    "    Inputs:\n",
    "        \n",
    "    df                - Data frame with text\n",
    "    text_field        - Name of column with text\n",
    "    removeNumbers     - Flag to indicate whether to remove numbers\n",
    "    removeStopWords   - Flag to indicate whether to remove stopwords\n",
    "    charLengthMin     - Minimum allowed character length of words\n",
    "    lemmatize         - Flag to indicate if we should lemmatize\n",
    "\n",
    "    Outputs:\n",
    "\n",
    "    df                - Input data frame with new column <text_field>_pp for pre-processed text\n",
    "        \n",
    "    \"\"\"\n",
    "    \n",
    "    def preprocessStr(s):\n",
    "        \"\"\"\n",
    "        Given a text, cleans and normalizes it. Feel free to add your own stuff.\n",
    "        \"\"\"\n",
    "        s = s.lower()\n",
    "        \n",
    "        # Remove stopwords\n",
    "        if removeStopWords == True:\n",
    "            text_no_stopwords = nlp(s)\n",
    "            text_no_stopwords = [t.text for t in text_no_stopwords if (t.is_alpha and not (t.is_stop or t.like_num))]\n",
    "            s = ' '.join(text_no_stopwords)\n",
    "            \n",
    "        # Lemmatize\n",
    "        if lemmatize == True:\n",
    "            lemmas = nlp(s)\n",
    "            lemmas = [t.lemma_ for t in lemmas if (t.is_alpha and not (t.is_stop or t.like_num))]\n",
    "            s = \" \".join(lemmas)\n",
    "        \n",
    "        # Remove \\n and unicode\n",
    "        s = re.sub(r'\\\\n',' ',s)\n",
    "        s = re.sub(r'\\\\u[0-9]*b?',' ',s)\n",
    "        \n",
    "        # Words split by line\n",
    "        s = re.sub(r'\\-','',s)\n",
    "        \n",
    "        # Remove '_' so there is no confusion with labels\n",
    "        s = re.sub(r'_',' ',s)\n",
    "        \n",
    "        # Remove very short words\n",
    "        if charLengthMin > 1:\n",
    "            s = re.sub(r'\\b\\w{1,'+ str(charLengthMin) +'}\\b', '', s)\n",
    "                \n",
    "        # Replace ips\n",
    "        s = re.sub(r'\\d{1,3}\\.\\d{1,3}\\.\\d{1,3}\\.\\d{1,3}', ' _ip_ ', s)\n",
    "        # Isolate punctuation\n",
    "        #s = re.sub(r'([\\'\\\"\\.\\(\\)\\!\\?\\-\\\\\\/\\,])', r' \\1 ', s)\n",
    "        s = re.sub(r'([\\'\\\"\\.\\(\\)\\!\\?\\-,])', r' \\1 ', s)\n",
    "        # Remove some special characters\n",
    "        s = re.sub(r'([\\;\\:\\|•«\\n])', ' ', s)\n",
    "        \n",
    "        # Remove very short words\n",
    "        s = re.sub(r'\\b\\w{1,'+ str(charLengthMin) +'}\\b', ' ', s)\n",
    "        \n",
    "        # Replace numbers and symbols with language\n",
    "        if removeNumbers == True:\n",
    "            s = re.sub(r'[0-9]+', '', s)\n",
    "        else:\n",
    "            s = s.replace('&', ' and ')\n",
    "            s = s.replace('@', ' at ')\n",
    "            s = s.replace('0', ' zero ')\n",
    "            s = s.replace('1', ' one ')\n",
    "            s = s.replace('2', ' two ')\n",
    "            s = s.replace('3', ' three ')\n",
    "            s = s.replace('4', ' four ')\n",
    "            s = s.replace('5', ' five ')\n",
    "            s = s.replace('6', ' six ')\n",
    "            s = s.replace('7', ' seven ')\n",
    "            s = s.replace('8', ' eight ')\n",
    "            s = s.replace('9', ' nine ')\n",
    "            \n",
    "        # Left over from web stuff, removing these\n",
    "        s = re.sub(r' s ',' ',s)\n",
    "        s = re.sub(r' t ',' ',s)\n",
    "        s = re.sub(r' d ',' ',s)\n",
    "\n",
    "        s = re.sub('  ',' ',s)\n",
    "                \n",
    "        return s\n",
    "    \n",
    "    df2 = df.copy()\n",
    "    df2[text_field + '_pp'] = df2[text_field].apply(preprocessStr)\n",
    "    \n",
    "    return df2\n",
    "\n",
    "def summarizeText(text, sentences=5):\n",
    "    \"\"\"\n",
    "    Automatically summarize a text using Lsa summarized\n",
    "    \n",
    "    Inputs:\n",
    "        \n",
    "    text              - Data frame with text\n",
    "    sentences         - Number of sentences to summarize with (default=5)\n",
    "\n",
    "    Outputs:\n",
    "\n",
    "    summarized_text \n",
    "        \n",
    "    \"\"\"\n",
    "    parser = PlaintextParser.from_string(text,Tokenizer(\"english\"))\n",
    "    #for sentence in LexRankSummarizer()(parser.document, sentences):\n",
    "    #    print(sentence)\n",
    "    summarized = \"\"\n",
    "    for sentence in LsaSummarizer()(parser.document,sentences):\n",
    "        summarized = summarized + \". \" + str(sentence)\n",
    "    return summarized\n",
    "\n",
    "print(\"Done\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Analysis\n",
    " \n",
    "Read the training datafile (and download is not already available), set the text field we will use and take a peek at some data.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading https://datakind-public.s3.amazonaws.com/NLP_data/rhode_island.pkl.zip ...\n",
      "Done\n"
     ]
    }
   ],
   "source": [
    "# Directory where we store data files\n",
    "ddir = './data/'\n",
    "\n",
    "file = \"rhode_island.pkl\"\n",
    "if not os.path.isfile(ddir + file):\n",
    "    s3_link = 'https://datakind-public.s3.amazonaws.com/NLP_data/' + file + '.zip'\n",
    "    print('Downloading ' + s3_link + ' ...')\n",
    "    resp = requests.get(s3_link).content\n",
    "    zipfile = ZipFile(BytesIO(resp))\n",
    "    zipfile.extractall(ddir)\n",
    "\n",
    "df = pickle.load(open(ddir + file,\"rb\"))\n",
    "    \n",
    "print(\"Done\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Setting field which will contain the text we will classify ...\n",
      "AN ACT RELATING TO MOTOR AND OTHER VEHICLES - MOTOR VEHICLE OFFENSES (Makes driving under the influence with any child under thirteen (13) years of age in the vehicle a felony punishable by up to five (5) years imprisonment, a five thousand dollar ($5,000) fine, and a loss of license of up to two (2) years.)\n",
      "Done\n"
     ]
    }
   ],
   "source": [
    "# Flag which column name we will use for text and pre-processed text \n",
    "text_field     = 'model_text'\n",
    "text_field_pp  = text_field + '_pp'\n",
    "    \n",
    "# Create a column of text\n",
    "print(\"Setting field which will contain the text we will classify ...\")\n",
    "\n",
    "# Use just title text\n",
    "df[text_field] = df['title']\n",
    "\n",
    "# Use title and summarized text\n",
    "#df['summarized'] = df['billText'].apply(lambda x: summarizeText(x),5)\n",
    "#df[text_field] = df['title'] + '\\n\\n\\n' + df['summarized']\n",
    "\n",
    "# Use title and full text\n",
    "#df[text_field] = df['title'] #+ '\\n\\n\\n' + df['billText']\n",
    "\n",
    "# Print example \n",
    "texts = list(df[text_field])\n",
    "print(texts[1400])\n",
    "print(\"Done\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Average word count of text field we generated: 36.33\n"
     ]
    }
   ],
   "source": [
    "df['word_count'] = df[text_field].apply(lambda x: len(str(x).split(\" \")))\n",
    "average_wordcount = df['word_count'].mean()\n",
    "print(\"Average word count of text field we generated: \" + str(round(average_wordcount,2)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Labels analysis\n",
    "\n",
    "Let's look at the labels we have in the 'subject' field. Note: Bills have more than one category, so this field contains a list."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "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>subject</th>\n",
       "      <th>Count</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>215</th>\n",
       "      <td>Taxation</td>\n",
       "      <td>262</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>207</th>\n",
       "      <td>State Affairs</td>\n",
       "      <td>258</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>35</th>\n",
       "      <td>Cities &amp; Towns</td>\n",
       "      <td>219</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>76</th>\n",
       "      <td>Education</td>\n",
       "      <td>195</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>154</th>\n",
       "      <td>Motor Vehicles</td>\n",
       "      <td>184</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>106</th>\n",
       "      <td>Health &amp; Safety</td>\n",
       "      <td>169</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>178</th>\n",
       "      <td>Property</td>\n",
       "      <td>161</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>49</th>\n",
       "      <td>Criminal Offense</td>\n",
       "      <td>144</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>25</th>\n",
       "      <td>Business &amp; Professions</td>\n",
       "      <td>138</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>45</th>\n",
       "      <td>Courts</td>\n",
       "      <td>121</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>128</th>\n",
       "      <td>Labor</td>\n",
       "      <td>118</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>34</th>\n",
       "      <td>Children and Families</td>\n",
       "      <td>117</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>134</th>\n",
       "      <td>Licenses</td>\n",
       "      <td>114</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>149</th>\n",
       "      <td>Minors</td>\n",
       "      <td>111</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>85</th>\n",
       "      <td>Fees</td>\n",
       "      <td>104</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>50</th>\n",
       "      <td>Criminal Procedure</td>\n",
       "      <td>98</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>184</th>\n",
       "      <td>Public Utilities</td>\n",
       "      <td>94</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>141</th>\n",
       "      <td>Marriage</td>\n",
       "      <td>93</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>122</th>\n",
       "      <td>Insurance</td>\n",
       "      <td>89</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>77</th>\n",
       "      <td>Elections</td>\n",
       "      <td>69</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>12</th>\n",
       "      <td>Attorney General</td>\n",
       "      <td>64</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>142</th>\n",
       "      <td>Medical</td>\n",
       "      <td>63</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>69</th>\n",
       "      <td>Domestic Relations</td>\n",
       "      <td>63</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>108</th>\n",
       "      <td>Health Care Plans/Services</td>\n",
       "      <td>62</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>191</th>\n",
       "      <td>Retirement</td>\n",
       "      <td>59</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>119</th>\n",
       "      <td>Human Services</td>\n",
       "      <td>58</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>33</th>\n",
       "      <td>Charter</td>\n",
       "      <td>57</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>105</th>\n",
       "      <td>Handicapped</td>\n",
       "      <td>57</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>175</th>\n",
       "      <td>Police &amp; Fire</td>\n",
       "      <td>55</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>109</th>\n",
       "      <td>Health Insurance</td>\n",
       "      <td>51</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>54</th>\n",
       "      <td>Day Care</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>48</th>\n",
       "      <td>Credit Cards</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>125</th>\n",
       "      <td>Johnston</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>121</th>\n",
       "      <td>Incorporate</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>110</th>\n",
       "      <td>Heating Fuels</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>62</th>\n",
       "      <td>Dept. of Corrections</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>206</th>\n",
       "      <td>Sports</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>64</th>\n",
       "      <td>Dept. of Employment and Training</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>103</th>\n",
       "      <td>Grandparents Rights</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>228</th>\n",
       "      <td>Vital Statistics</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>92</th>\n",
       "      <td>Fire Marshall</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>90</th>\n",
       "      <td>Fire Detection</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>237</th>\n",
       "      <td>West Greenwich</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>132</th>\n",
       "      <td>Libraries</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>131</th>\n",
       "      <td>Legal Opinion</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>Amendment</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>202</th>\n",
       "      <td>Smithfield</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>133</th>\n",
       "      <td>Library Services</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>136</th>\n",
       "      <td>Life Insurance</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>138</th>\n",
       "      <td>Liquor</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>139</th>\n",
       "      <td>Lobbying</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>38</th>\n",
       "      <td>Code of Ethics</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>156</th>\n",
       "      <td>National Guard</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>160</th>\n",
       "      <td>North Providence</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>37</th>\n",
       "      <td>Claims</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>167</th>\n",
       "      <td>Parental Rights</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>186</th>\n",
       "      <td>Railroads</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>192</th>\n",
       "      <td>Richmond</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>194</th>\n",
       "      <td>Rules</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>83</th>\n",
       "      <td>Exeter</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>244 rows × 2 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                              subject  Count\n",
       "215                          Taxation    262\n",
       "207                     State Affairs    258\n",
       "35                     Cities & Towns    219\n",
       "76                          Education    195\n",
       "154                    Motor Vehicles    184\n",
       "106                   Health & Safety    169\n",
       "178                          Property    161\n",
       "49                   Criminal Offense    144\n",
       "25             Business & Professions    138\n",
       "45                             Courts    121\n",
       "128                             Labor    118\n",
       "34              Children and Families    117\n",
       "134                          Licenses    114\n",
       "149                            Minors    111\n",
       "85                               Fees    104\n",
       "50                 Criminal Procedure     98\n",
       "184                  Public Utilities     94\n",
       "141                          Marriage     93\n",
       "122                         Insurance     89\n",
       "77                          Elections     69\n",
       "12                   Attorney General     64\n",
       "142                           Medical     63\n",
       "69                 Domestic Relations     63\n",
       "108        Health Care Plans/Services     62\n",
       "191                        Retirement     59\n",
       "119                    Human Services     58\n",
       "33                            Charter     57\n",
       "105                       Handicapped     57\n",
       "175                     Police & Fire     55\n",
       "109                  Health Insurance     51\n",
       "..                                ...    ...\n",
       "54                           Day Care      2\n",
       "48                       Credit Cards      2\n",
       "125                          Johnston      1\n",
       "121                       Incorporate      1\n",
       "110                     Heating Fuels      1\n",
       "62               Dept. of Corrections      1\n",
       "206                            Sports      1\n",
       "64   Dept. of Employment and Training      1\n",
       "103               Grandparents Rights      1\n",
       "228                  Vital Statistics      1\n",
       "92                      Fire Marshall      1\n",
       "90                     Fire Detection      1\n",
       "237                    West Greenwich      1\n",
       "132                         Libraries      1\n",
       "131                     Legal Opinion      1\n",
       "8                           Amendment      1\n",
       "202                        Smithfield      1\n",
       "133                  Library Services      1\n",
       "136                    Life Insurance      1\n",
       "138                            Liquor      1\n",
       "139                          Lobbying      1\n",
       "38                     Code of Ethics      1\n",
       "156                    National Guard      1\n",
       "160                  North Providence      1\n",
       "37                             Claims      1\n",
       "167                   Parental Rights      1\n",
       "186                         Railroads      1\n",
       "192                          Richmond      1\n",
       "194                             Rules      1\n",
       "83                             Exeter      1\n",
       "\n",
       "[244 rows x 2 columns]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "subjects = []\n",
    "for index, row in df.iterrows():\n",
    "    for s in row[\"subject\"]:\n",
    "        subjects.append({'identifier':row[\"identifier\"], 'subject':s})\n",
    "\n",
    "subjects = pd.DataFrame(subjects)\n",
    "summary = pd.DataFrame({'Count': subjects.groupby('subject')['identifier'].count()}).reset_index()\n",
    "summary = summary.sort_values(by='Count',ascending=False)\n",
    "\n",
    "display(summary)\n",
    "\n",
    "labels = list(summary[\"subject\"])\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Classification - using Fasttext to predict bill categories\n",
    "\n",
    "Next, we will train a [Fasttext](https://github.com/facebookresearch/fastText) classifier for predicting whether bills in Rhode Island are health related or not.\n",
    "\n",
    "Note: For more information on the Fasttext model parameters, please see [Fasttext](https://github.com/facebookresearch/fastText).\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Set our 'Health related' category"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Done\n"
     ]
    }
   ],
   "source": [
    "# Utility function to set our 'Health related' cateogry based on the list of categories on each bill\n",
    "def isHealthRelated(subjects):\n",
    "    health_subjects = ['Health & Safety','Health Care Plans/Services','Mental Health', 'Health Insurance', \\\n",
    "                       'Drugs','Prescription drugs','Dept. of Health','Medical']\n",
    "    if len(set(subjects).intersection(set(health_subjects))) > 0:\n",
    "        response=1\n",
    "    else:\n",
    "        response=0\n",
    "    return response\n",
    "\n",
    "response_field        = 'Health related' \n",
    "\n",
    "df[response_field] = df['subject'].apply(lambda x: isHealthRelated(x))\n",
    "\n",
    "print(\"Done\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Data split \n",
    "\n",
    "Split our data into test and training datasets."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train:\n"
     ]
    },
    {
     "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>Health related</th>\n",
       "      <th>Count</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0</td>\n",
       "      <td>1321</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>1</td>\n",
       "      <td>249</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   Health related  Count\n",
       "0               0   1321\n",
       "1               1    249"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test:\n"
     ]
    },
    {
     "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>Health related</th>\n",
       "      <th>Count</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0</td>\n",
       "      <td>335</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>1</td>\n",
       "      <td>58</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   Health related  Count\n",
       "0               0    335\n",
       "1               1     58"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "train, test = train_test_split(df, test_size=0.2, random_state=123)\n",
    "train = train.copy().reset_index()\n",
    "test  = test.copy().reset_index()\n",
    "\n",
    "print(\"Train:\")\n",
    "\n",
    "summary = pd.DataFrame({'Count': train.groupby(response_field)['id'].count()}).reset_index()\n",
    "display(summary)\n",
    "\n",
    "print(\"Test:\")\n",
    "summary = pd.DataFrame({'Count': test.groupby(response_field)['id'].count()}).reset_index()\n",
    "display(summary)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Preprocess text\n",
    "\n",
    "Now we pre-process text as required for Fasttext."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Pre-processing train ...\n",
      "Saving ./data/train.pkl\n",
      "Pre-processing test ...\n",
      "Saving ./data/test.pkl\n",
      "Done\n"
     ]
    }
   ],
   "source": [
    "# Pre-process text and save\n",
    "for f in [\"train\",\"test\"]:\n",
    "    data = eval(f)\n",
    "    print(\"Pre-processing \" + f + \" ...\")\n",
    "    data = preprocessFastText(data, text_field, lemmatize=False, removeStopWords=True)\n",
    "    file = \"./data/\" + f + \".pkl\"\n",
    "    print(\"Saving \" + file)\n",
    "    pickle.dump(data, open(file, \"wb\" ))\n",
    "    \n",
    "print(\"Done\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training\n",
    "\n",
    "Let's train our model!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Creating fasttext training file\n",
      "Writing 1570 lines\n",
      "/home/jovyan/fastText/fasttext supervised -input ./data//fasttext.txt -dim 100 -lr 0.2 -epoch 550 -wordNgrams 3 -output ./data//fastext_model-Healthrelated-0.2-100-550-3---\n",
      "Read 0M words\n",
      "Number of words:  4265\n",
      "Number of labels: 2\n",
      "Progress: 100.0% words/sec/thread:  244260 lr:  0.000000 loss:  0.027698 ETA:   0h 0m\n",
      "/home/jovyan/fastText/fasttext quantize -input ./data//fasttext.txt -dim 100 -lr 0.2 -epoch 550 -wordNgrams 3 -output ./data//fastext_model-Healthrelated-0.2-100-550-3---\n",
      "Done\n"
     ]
    }
   ],
   "source": [
    "train = pickle.load(open(\"./data/train.pkl\", \"rb\" ))\n",
    "\n",
    "# Some config fields\n",
    "prefix            = ''\n",
    "\n",
    "# Fasttext model parameters\n",
    "vecs              = 100   #  Number of vectors. If using pretrained, must match file\n",
    "lr                = 0.2\n",
    "pretrainedVectors = ''    # 'wiki-news-100d-1M-subword.vec' # See https://fasttext.cc/docs/en/english-vectors.html for options \n",
    "epoch             = 550\n",
    "wordNgrams        = 3\n",
    "threads           = ''    # '1' for reproducability but slow, set to '' to default\n",
    "quantize          = True  # Used to create a smaller model file, useful for prediction performance\n",
    "\n",
    "# Generate a model name from our parameters\n",
    "fasttext_model_name = prefix + 'fastext_model-' + response_field.replace(' ','').replace('&','') + '-' + str(lr) + '-' + str(vecs) + '-' + str(epoch) + '-' + str(wordNgrams) + '---' +  pretrainedVectors\n",
    "\n",
    "# Train the model, write a model file\n",
    "train_fastText(train, text_field_pp, ddir, vecs, lr, pretrainedVectors, epoch, wordNgrams, \n",
    "               fasttext_model_name, threads, quantize, response_field)\n",
    "\n",
    "print(\"Done\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Prediciton\n",
    "\n",
    "Now we use our trained model to predict on the test data."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Predicting ...\n",
      "fastext_model-Healthrelated-0.2-100-550-3---.ftz\n",
      "Confusion matrix, without normalization\n",
      "Normalized confusion matrix\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAVAAAAEYCAYAAAAK467YAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDIuMi4zLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvIxREBQAAIABJREFUeJzt3Xnc1WP+x/HX+77vJEpFiyzJJJQtSnYaS7IMmrGvJZLdMGYszQ+DYRi7wWjsM/Y1+xJJRSJJiELWSCkqFdXn98d1nXw77rPc517Oue/78+xxHvc53/U633P6nGv7XpfMDOecc1VXVuwEOOdcfeUB1DnnCuQB1DnnCuQB1DnnCuQB1DnnCuQB1DnnCuQBtMRJaibpcUnfS3qgGsc5TNJzNZm2YpG0g6QPSuV8kjpJMkkVdZWm+kLSNEm7xufnSPpPLZzjJkl/renj5nVu7wdaMyQdCpwObAjMBSYAF5vZqGoe9wjgZGBbM1tc7YSWOEkGdDGzqcVOSyaSpgHHmNkL8XUn4BOgSU1/RpJuB74wsyE1edy6kn6tauB4/ePxtq+J41WX50BrgKTTgauBvwPtgY7ADcC+NXD4dYAPG0PwzIfn8mqPX9sCmJk/qvEAWgLzgAOybNOUEGC/io+rgaZxXW/gC+AMYAYwHRgQ110A/AT8HM8xEDgf+G/i2J0AAyri6/7Ax4Rc8CfAYYnloxL7bQuMA76Pf7dNrBsBXAiMjsd5DmiT4b2l0v/nRPr3A/YEPgS+A85JbN8LeBWYE7e9HlghrhsZ38v8+H4PShz/L8DXwF2pZXGfzvEcW8TXawAzgd55fHZ3AGfE52vGc58QX68Xj6u0890FLAUWxDT+OfEZHAV8Fs9/bp6f/3KfS1xm8fyD4mf/UzzX4xnehwGDgSnAbOBf/FK6LAOGAJ/Gz+dOoGXad2dgTPfIxLIBwOfxeIOBLYGJ8XO7PnHuzsCLwKz4vv8HtEqsnwbsGp+fT/zuxs99XuKxGDg/rjsL+Ijw3XsP6BeXdwUWAkviPnPi8tuBixLnPBaYGj+/YcAa+Vyrgv7/FzsA1fcH0Dd++BVZtvkb8BrQDmgLjAEujOt6x/3/BjQhBJ4fgdbpX7oMr1Nf+ApgZeAHYIO4rgOwUXzen/gfFVg1fnmOiPsdEl+vFtePiF/g9YFm8fWlGd5bKv3/F9N/LPAtcDfQAtgoful/E7fvAWwdz9sJeB84Le0Lvl4lx/8HIRA1IxHQEv9h3gdWAp4F/pnnZ3c0MSgBh8b3fF9i3WOJNCTPN40YFNI+g6ExfZsBi4CueXz+yz6Xyq4BacEhw/sw4AmgFaH08y3QN/E+pgK/AZoDDwN3paX7TsJ3p1li2U3AikCf+Pk9GtO/JiEQ7xSPsR6wW/xs2hKC8NWVXSvSvruJbbrHNG8eXx9A+CEsI/yIzgc6ZLley64RsDMhkG8R03QdMDKfa1XIw4vw1bcaMNOyF7EPA/5mZjPM7FtCzvKIxPqf4/qfzewpwq/rBgWmZymwsaRmZjbdzN6tZJu9gClmdpeZLTaze4DJwO8S29xmZh+a2QLgfsKXPJOfCfW9PwP3Am2Aa8xsbjz/u8CmAGb2ppm9Fs87Dfg3sFMe7+k8M1sU07McMxtKyFGMJfxonJvjeCkvAztIKgN2BC4Dtovrdorrq+ICM1tgZm8DbxMCKeT+/GvCpWY2x8w+A17il8/rMOBKM/vYzOYBZwMHpxXXzzez+WnX9kIzW2hmzxEC2D0x/V8CrwCbA5jZVDN7Pn423wJXkvvzXEZSW0JwPtnM3orHfMDMvjKzpWZ2H+Gz7ZXnIQ8DbjWz8Wa2KL7fbWI9dUqma1VlHkCrbxbQJkf90RqEIlTKp3HZsmOkBeAfCbmFKjGz+YRf7MHAdElPStowj/Sk0rRm4vXXVUjPLDNbEp+n/hN+k1i/ILW/pPUlPSHpa0k/EOqN22Q5NsC3ZrYwxzZDgY2B6+J/nJzM7CPCj1V3YAdCzuQrSRtQWADNdM1yff41oSrnriDU1ad8Xsnx0j+/TJ9nO0n3Svoyfp7/JffnSdy3CfAgcLeZ3ZtYfqSkCZLmSJpD+FzzOiZp7zf+aMyi8O92Vh5Aq+9VQhFnvyzbfEVoDErpGJcVYj6hqJqyenKlmT1rZrsRcmKTCYElV3pSafqywDRVxY2EdHUxs1WAcwj1jNlk7SoiqTmhXvEW4HxJq1YhPS8D+xPqYb+Mr48EWhN6UlQ5PZXI9vkv93lKWu7zLOBc+Zx7McsHxOqc45K4/6bx8zyc3J9nynWEes5lPQwkrUP4zp5EqFJqBUxKHDNXWpd7v5JWJpQSa+W77QG0mszse0L9378k7SdpJUlNJO0h6bK42T3AEEltJbWJ2/+3wFNOAHaU1FFSS0IRBQBJ7SXtE780iwi5qyWVHOMpYH1Jh0qqkHQQ0I2QA6ttLQj1tPNi7vj4tPXfEOrrquIa4E0zOwZ4klB/B4Ck8yWNyLLvy4T/rCPj6xGEbmOjErnqdFVNY7bP/21gI0ndJa1IqCeszrkqO/cfJa0bf2j+TqjnraleHS2IDTqS1gTOzGcnSccRcvmHmtnSxKqVCUHy27jdAEIONOUbYC1JK2Q49N3AgHg9mxLe79hYXVTjPIDWADO7ktAHdAjhg/+c8J/y0bjJRcAbhFbMd4DxcVkh53oeuC8e602WD3plhNb8rwgtkDsBJ1RyjFnA3nHbWYSW5L3NbGYhaaqiPxEabOYSchr3pa0/H7gjFt8OzHUwSfsSGvIGx0WnA1tIOiy+XpvQmyCTlwlBIBVARxFyhCMz7hFyXUNiGv+UK41k+fzN7ENCI9MLhLq+9H7DtwDd4rkepepuJfQcGEnolbGQ8ANRUy4gNNh8T/jxejjP/Q4h/DB8JWlefJxjZu8BVxBKdt8Am7D85/cioU79a0m/+r6a2XDgr8BDhF4enYGDC3lj+fCO9K5BkzQB2CX+aDhXozyAOudcgbwI75xzBfIA6pxzBfIA6pxzBfLBAxoZVTQzrdCi2MmoV7p37VjsJNRLb41/c6aZta3OMcpXWcds8a9uPlvGFnz7rJn1rc45qsMDaCOjFVrQdIOcvYNcwsgx1xY7CfVSixXL0+92qzJbvCDr93XhhH/le4dSrfAA6pwrXRKUlRc7FRl5AHXOlTYPoM45VwiBSret2wOoc650Cc+BOudcYbwO1DnnCqd8R8erex5AnXOly1vhnXOuUIKy0g1Tpdu85ZxzAGXK/MhC0oqSXpf0tqR3JV0Ql68raaykKZLuSw3OLKlpfD01ru+UM2k18Pacc652pFrhMz2yWwTsbGabEea96itpa8IMr1eZWRfCbLQD4/YDgdlmth5wVdwuKw+gzrkSpoIDqAXz4ssm8WGEqY8fjMvv4Jf5zPaNr4nrd5Gyt2B5AHXOlTaVZX6EGXHfSDwGLberVB5nJZgBPA98BMxJzAn1Bb/M2LkmcYbSuP57woR0GZVu7axzzuVuhZ9pZj0zrYwTA3aX1Ap4BOha2Waps2VZVynPgTrnSljhRfgkM5tDmHF1a6CVpFTmcS1+mWL6C8IkhMT1LQmTM2bkAdQ5V9qkzI+su6ltzHkiqRmwK/A+8BKwf9zsKOCx+HxYfE1c/6LlmDTOi/DOudKlavUD7UCYIruckFm838yekPQecK+ki4C3CFNHE//eJWkqIeeZczpkD6DOudJW4J1IZjYR2LyS5R8DvSpZvhA4oCrn8ADqnCttfi+8c84VwO+Fd865wggoKyvdtm4PoM650iUq751ZIjyAOudKmDwH6pxzhfIA6pxzhfAivHPOFUZehHfOucJ5AHXOuUJ4Ed455wrjRXjnnKsG5Zj7qJg8gDrnSpcgx6waReUB1NWJpitU8MItp7HCChVUlJfzyAtvcdFNT3HbxUexRbeO/Lx4CW9M+pSTLr6HxYuXArBDjy5cfuYfaFJRzqw58+hzzDVFfhfFdfyggTzz9JO0bduO18dPBODvF17A7bf9hzZt2gJw3t8uYve+exYzmTXOi/Cu0Vv002L6DrqW+Qt+oqKijBdvPZ3nRr/HvU+PY8C5YR6vOy7pz4B+2zL0gVG0bN6Ma845kH1PvIHPv55N29bNi/wOiu+wI47iuONPZNDA/sstP/Hk0zj1j2cUJU21Taiki/ClG9pdgzN/wU8ANKkop6KiHDPj2VHvLVv/xqRPWbNdawAO2qMnjw1/m8+/ng3At7Pn/fqAjcz2O+xI69arFjsZdSsW4TM9su4qrS3pJUnvx3nhT43Lz5f0paQJ8bFnYp+z47zwH0jaPVfyPIC6OlNWJl679yw+G34pL742mXGTPl22rqKijEP26sXzY0JA7bJOO1qtshLPDj2V0f/7M4fu/avxb110843/Yuue3Tl+0EBmz55d7OTUuLKysoyPHBYDZ5hZV8JcSCdK6hbXXWVm3ePjKYC47mBgI6AvcEMczT5z2qrzxrKRZJKuSLz+k6TzE68HSZocH69L2j4ufyT+KkyV9H3iV2LbtOPfLmn/tGUFZ1MkTZPURlIrSScklveW9EShx63kPP0lXZ9jm97p7zfPY0+T1Kbw1NWupUuNrQ++lPV2H0LPjdehW+cOy9Zdc/ZBjB4/ldFvfQRARXkZW3Rdm34n38g+J/6Ls4/ty3od2xUr6SXrmEGDmfj+FMa8Pp7VV+/AOX/5U7GTVONUpoyPbMxsupmNj8/nEuZDWjPLLvsC95rZIjP7BJhKJSPXJ9VmDnQR8PvK/kNL2hs4DtjezDYEBgN3S1rdzPqZWXfgGOCVxK/EmFpMa1Ir4IScW2WRmPGvUL2BKgfQ+uL7eQsY+cYU+mwbMgPnDNqDtq2b8+crHl62zZcz5vDcmPf5ceFPzJozn1Hjp7Lp+tm++41Tu/btKS8vp6ysjP5HH8Obb4wrdpJqVLbieyzCZ50XPnGcToTpPcbGRSdJmijpVkmt47Jl88JHyTnjK1WbAXQxcDPwx0rW/QU408xmAsRfiTuAE2vq5JLOlDQuXqQLEssflfRmrBOp7GJfCnSOud7L47Lmkh6MueX/qZLKF0kjJP1d0svAqXFGwIdiGsZJ2q6SfX4naayktyS9IKl9/KAHA3+Madgh07EkrSbpubj/vynhezbatG5Oy+bNAFixaRN23moDPpj2Df37bcNu23blyLNvJzkB4uMjJrLd5p0pLy+j2YpN2HLjTkz+5OtiJb9kfT19+rLnjw97lG4bbVTE1NSOHEX4mWbWM/G4OX1/Sc2Bh4DTzOwH4EagM9AdmA6kSspVnhe+tlvh/wVMlHRZ2vKNgDfTlr3BL1OK5utySUPSF0rqA3QhZL8FDJO0o5mNBI42s+/iNKfjJD1kZrMSu58FbBxzwUjqTfjl2ogwf/RoYDtgVCXpaWVmO8X97ibUs4yS1BF4Fuiatv0oYGszM0nHAH82szMk3QTMM7N/5jjWecAoM/ubpL2ATL++g5ata1Kc1uzV26zC0L8dQXlZGWVl4qHnx/P0K5OYO+4aPpv+HSPuCK3Ij704gUtufoYPPvmG58e8x7j7z2bpUuP2R8bw3kfTc5ylYRtwxKG88srLzJo5kw06d+ScIecxauTLTJz4NpLouM46XHv9TcVOZs2rRrZAUhNC8PyfmT0MYGbfJNYPBVJVdMvmhY+Sc8ZXqlYDqJn9IOlO4BRgQY7NRY5oX4kzzezBZQf4pQ60T3y8FV83JwTUkcApkvrF5WvH5ckAWpnXzeyLeI4JQCcqD6D3JZ7vCnRLZFZXkdQibfu1gPskdQBWAD7JcP5Mx9oR+D2AmT0pqdIWhPirfDNA2UrtqnqNa8SkKV+xzSH/+NXyFluemnGfq+4czlV3Dq/NZNUrt91196+WHTVgYBFSUodUeD/QWFK8BXjfzK5MLO9gZqlf437ApPh8GKEq8UpgDUJseD3bOeqiH+jVwHjgtsSy94AewIuJZVvE5TVBwCVm9u/lFobc5K7ANmb2o6QRwIp5HG9R4vkSMl+3+YnnZfE8y/1wpJX+rwOuNLNhMW3nZzhutmMVJSA6VxfCvfAFZ0G3A44A3okZH4BzgEMkdSf835lGaI/BzN6VdD8hDi0GTjSzJdlOUOvdmMzsO+B+IPlTeRnwD0mrAcQ30x+4oYZO+yxwdKz7QNKaktoBLYHZMXhuSOjakG4ukJ5TLMRzwEmpF/E9pmsJfBmfJ6sv0tOQ6VgjgcPisj2A1jjXwEiZH9mY2Sgzk5ltmuyyZGZHmNkmcfk+idwoZnaxmXU2sw3M7OlcaaurfqBXAMta481sGHArMEbSZGAocHjyjVSHmT0H3A28Kukd4EFCQHoGqJA0EbgQeK2SfWcBoyVNSjQiFeIUoGdsxHqP0DCU7nzgAUmvADMTyx8H+qUakbIc6wJgR0njCVUWn1Ujvc6VHoX+w5kexaZky6dr+MpWamdNNziw2MmoV7597dpiJ6FearFi+Ztm1rM6x2jWYX37zdGZu02/9/fdq32O6vB74Z1zJa2SXoMlwwOoc65kKRbhS5UHUOdcCSuNus5MPIA650qaF+Gdc64AXoR3zrlq8ADqnHMFKuESvAdQ51zp8iK8c84VrJ62wktaJduOcVw955yrVfW1Ff5dwmglydSnXhvQsRbT5Zxz9bcIb2ZrZ1rnnHN1pZQDaF6jMUk6WNI58flaknrUbrKccy7IMSdSUeUMoAozSP6WMDApwI9AA5w3wDlXaqTMQ9mVQs40nxzotmZ2HLAQlg2QvEKtpso556LyMmV8ZCNpbUkvSXo/TiJ5aly+qqTnJU2Jf1vH5ZJ0rcKU6hMlbZErbfkE0J8llRGnjoijyC/NYz/nnKu2QkekJ0zLcYaZdSXMPnGipG6EiSOHm1kXYHh8DbAHYR6kLoRJGG/MdYJ8Aui/CLPatVWYHngU8OvZwZxzroZJhedAzWx6nDIdM5sLvE+Y531fwjTqxL/7xef7Anda8BrQKk74mFHOjvRmdqekNwmTsQEcYGaTsu3jnHM1JUddZxtJbyRe35xhbvhOhOnJxwLtU9MHmdn0OF8ahOD6eWK3L+KyjFMN5XsnUjnwM6EYX1fzKDnnGjkRZubMYmauKT3i5JIPAafFqdaznS5d1jmP8mmFPxe4hzBP8lqEeZPPzrWfc85VmzIX33MV4cPuakIInv8zs4fj4m9SRfP4d0Zc/gWQ7P++FvBVtuPnk5s8HNjSzIaY2blAL+DIPPZzzrlqEdVqhRdwC/C+mV2ZWDWMX6YRPwp4LLH8yNgavzXwfa6ZgvMpwn+atl0F8HEe+znnXLVVo7/8doT+6+9ImhCXnQNcCtwvaSBhKvAD4rqngD2BqYT+7gNynSDbYCJXEcr/PwLvSno2vu5DaIl3zrlaVZ174c1sFJXXawLsUsn2BpxYlXNky4GmWtrfBZ5MLH+tKidwzrnqKC+BWzYzyTaYyC11mRDnnKtMKdzznknOOlBJnYGLgW7AiqnlZrZ+LabLOeeQ8mttL5Z8WuFvB24j1CXsAdwP3FuLaXLOuWXq+2AiK5nZswBm9pGZDSGMzuScc7VKQJkyP4otn25Mi2J/qo8kDQa+BNrl2Mc552pEWX2uAwX+CDQHTiHUhbYEjq7NRDnnHMRuTPU5gJrZ2Ph0Lr8Mquycc3WiFOo6M8nWkf4RstxIb2a/r5UUOedcJEq7FT5bDvT6OkuFqzObd+3I6LH+0VbFjB8WFTsJjVd+AycXTbaO9MPrMiHOOVeZenknknPOFVtqNKZS5QHUOVfSSjh+5h9AJTU1M68Mcs7VmdScSKUqnxHpe0l6B5gSX28m6bpaT5lzrtGrzoDKdSGfWzmvBfYGZgGY2dv4rZzOuTpSluWRi6RbJc2QNCmx7HxJX0qaEB97JtadHeeF/0DS7rmOn08RvszMPk0bUmpJHvs551y11MBoTLcTumTembb8KjP7Z9q5ugEHAxsR5oB7QdL6ZpYx3uUTxD+X1AswSeWSTgM+rMIbcM65gpWXZX7kYmYjge/yPNW+wL1mtsjMPiFM7dEr2w75BNDjgdOBjsA3wNZxmXPO1aowGpMyPojzwiceg/I89EmSJsYifuu4LNO88Bnlcy/8DEK21jnn6pZy5jRzzgtfiRuBCwm3ql8IXEEYIKnK88LnMyL90MoOYmb5RnrnnCuIqPk7kczsm2XHD/HtifiyVuaFfwEYHh+jCWOBen9Q51ydqOkBlSV1SLzsxy8TaA4DDpbUVNK6QBfg9WzHyqcIf1/aye8Cnq9Sip1zrgDVvZVT0j1Ab0Jd6RfAeUBvSd0JJetpwHEAZvaupPuB94DFwInZWuChsFs51wXWKWA/55yrmmreiWRmh1SyOOOMw2Z2MWHg+LzkUwc6m1/qQMsIXQLOyvcEzjlXqNScSKUqawCNcyFtRpgHCWCpmWVtlXLOuZqjkh7OLmsjUgyWj5jZkvjw4OmcqzNS9TrS17Z8kvC6pC1qPSXOOVeJHB3piyrbnEgVZrYY2B44VtJHwHxCtYSZmQdV51ytqs8DKr8ObAHsV0dpcc655YSO9MVORWbZAqgAzOyjOkqLc84tT2FEplKVLYC2lXR6ppVmdmUtpMc555apjVs5a1K2AFoONKfyG+ydc65OlHAVaNYAOt3M/lZnKXHOuV9RvS3Cl26qnXONQn0uwu9SZ6lwzrnKiJLo75lJxgBqZvkOg++cc7VC5He3T7EUMhqTc87VmXqZA3XOuWIr9TrQUs4dO+ccUuZH7n0rnRd+VUnPS5oS/7aOyyXp2jgv/MR8xgDxAOqcK2GZBxLJs2h/O9A3bdlZwHAz60KYqig1vvEehGk8ugCDCJPPZeUB1BXd9ddeQ4/uG7PFZhtx3TVXFzs5JW3JkiXs0XsrBhzSD4AzTzmOvjtuye479GRw/0OYP29ekVNYs6RQhM/0yCXDvPD7AnfE53fwy3gf+wJ3WvAa0Cpt/qRf8QDqiurdSZO47dahvDLmdV5/822efuoJpk6ZUuxklaxb/309662/wbLX/3fR5TwzchzPvvIGa6y1Nnf8J2emqd6pThE+g/ZmNh0g/m0Xl1d5XngPoK6oJk9+n169tmallVaioqKCHXbciccee6TYySpJ07/8ghefe5qDDx+wbFmLVVYBwMxYtGBBSd+1U4hUI1KWHGgbSW8kHtWZbr3K88J7AHVFtdFGGzNq1EhmzZrFjz/+yDNPP8UXn3+ee8dG6IJzz+Sc8/9OWdny/23/dNKx9Oy6DlOnfkD/Y08oUupqj7L8A2aaWc/E4+Y8DvlNqmge/86Iy2tlXviSJWle2uv+kq4v8Fi9JT2ReL5tYt3tkvavXmqXO9cIST1zbHOapJWqeNxl76G+2LBrV87401/Yu+9u7LNXXzbddDMqKrx3Xbrhzz7Fam3askn3XzcM//P6obz+7ies12VDHn/kgSKkrnbVwoj0w4Cj4vOjgMcSy4+MrfFbA9+nivoZ01ZoChq43sC2uTbKJH4A1b22pwFVCqD1Vf+jB/LquPG88NJIWq+6Kuut16XYSSo5b4wdwwvPPMl23dfn5GOPZMwrIzj1uP7L1peXl/O7fvvz9BOPFi+RtUBkLr7n04gU54V/FdhA0heSBgKXArtJmgLsFl8DPAV8DEwFhgI5s/MNNoBKaivpIUnj4mO7uLyXpDGS3op/N0jbrxMwGPijpAmSdoirdozbf1xZblRSJ0nvS7oBGA+sLamPpFcljZf0gKTmlex3Y6y7eVfSBXHZKcAawEuSXorLKj2WpL6SJksaBfy+Zq5e3ZoxI5SgPvvsMx579GEOPLiyqbwbt7/830WMnfQRoyd8yHVD72TbHXpz9U23Me3jMN65mfHCs0/RucsGOY5Uz2RpQMonA2pmh5hZBzNrYmZrmdktZjbLzHYxsy7x73dxWzOzE82ss5ltYmZv5Dp+fS8rNZM0IfF6VUI2HOAa4CozGyWpI/As0BWYDOxoZosl7Qr8HfhD6gBmNk3STcA8M/snQPzV6kCYH2rDeI4HK0nPBsAAMztBUhtgCLCrmc2X9BfgdCB9iMBzzew7SeXAcEmbmtm1cTDr35rZzEzHknQZ4ZdyZ8Kv5n2VXaRYsT4IYO2OHbNe0GI45MA/8N13s2hS0YSrr/0XrVu3LnaS6gUz4/QTBzJv7lzMjK4bb8LFl19X7GTVqFK/E6m+B9AFZtY99UJSfyBVt7gr0C3RKrmKpBZAS+AOSV0ILWxN8jzXo2a2FHhPUvsM23wa+48BbA10A0bHNKxAKEqkOzAGuApCkO4GTEzbJtOxNgQ+MbMpAJL+SwyUSbFi/WaAHj16ltzU1MNHvFLsJNQr22y/E9tsvxMADz89oriJqQMlHD/rfQDNpgzYxswWJBdKug54ycz6xeL6iDyPtyh5mAzbzE/b5nkzy1gelbQu8CdgSzObLel2YMXKNq3sWJK6k6ObhXP1nUp4aOIGWwcKPAeclHoRgw2EHOiX8Xn/DPvOBVpU8/yvAdtJWi+efyVJ66dtswoh6H4fc7V7ZEhDpmNNBtaV1Dlu55WHrsEpU+ZHsTXkAHoK0DMOCvAeoWEI4DLgEkmjCfM+VeZxoF9aI1KVmNm3hAB9j6SJhCC4Ydo2bwNvAe8CtwKjE6tvBp6W9FKmY5nZQkKR/cnYiPRpIWl1rlSJMCtnpkexycxLgI1Jjx49bfTYnI2LLmHGD4tyb+R+ZZ3VVnzTzLL2d86l26ab213DXs64vue6Lat9jupoyHWgzrkGoAQymhl5AHXOlbBq3XFU6zyAOudKlijt6YE9gDrnSlopNBZl4gHUOVfSSqG7UiYeQJ1zpavEy/AeQJ1zJUv4tMbOOVewEo6fHkCdc6WtlO+F9wDqnCtp3ojknHMFSN0LX/D+0jTCwDxLgMVm1lPSqoSxczsB04ADzWx2IcdvyIOJOOfqu2qOSB/91sy6J+6ZPwsYbmZdgOHxdUE8gDrnSlotzAu/L3BHfH4HsF+hB/IA6pwrYZln5CzLb154A56T9GZiXfvUbJvxb7tCU+d1oM65kpVHP/qZOYaz287MvpLUDnhe0uQaTJ7nQJ1zpa06Ayqb2Vfx7wzgEaAX8I2kDvHYHYAZhabNA6hzrqQVOqWHpJXjRJJIWhnoA0wizKp7VNzsKOCxQtPmRXjnXOmjizCHAAASqUlEQVSqXmNRe+CRmFOtAO42s2ckjQPuj9OVfwYcUOgJPIA650pWdfqBmtnHwGaVLJ8F7FK9lAUeQJ1zJc3vRHLOuQL5vfDOOVcgH43JOecKIPl4oM45V7jSjZ8eQJ1zpc0bkZxzriDyRiTnnCtE6Ada7FRk5gHUOVfSvBHJOecKUb1bOWudB1DnXMnyIrxzzlWDF+Gdc65ApRs+PYA650pcdWblrG0ys2KnwdUhSd8CnxY7HRm0AWYWOxH1TClfs3XMrG11DiDpGcJ7zGSmmfWtzjmqwwOoKxmS3sgxv41L49esuHxKD+ecK5AHUOecK5AHUFdKbi52Auohv2ZF5HWgzjlXIM+BOudcgTyAukZPpdzR0JU0D6Cu0ZK0laQOZmYeRANJG0laqdjpqC88gLrG7DDgBUmrexAFSSsDg4GhHkTz441IrtGRVGZmS+Pze4B2wOFmNl2SrBH+p0i9b0nrAGcBS4EzzGxhkZNW0jwH6hqdRPDsD8wB2gMvS1qjseZEEz8avQlxYQfgCs+JZuc5UNcoSdoGuA3Yycy+kXQNsAvQx8y+aow5UUn7AZcA+wI9gK2BZsApnhOtnOdAXaNQSa5yOvA6MB/AzE6Ny8ZIat8YgmfqmiSuTTPgQTP7EHgEuAtYH7hBUrPipLK0eQB1DV4yNylpVUmrEYLlWsDuiU3/C0wFVqj7VNattBx2Kg5MBg6RtJOZLTSzN4APCfWhrYqRzlLnRXjXaEg6HdgRaA38B5gGXAc8AawMdAcOM7OvipXGuibpWKAX8AkwClgDOA74N1AOHA4caWbfFi2RJcwHVHaNgqSDgb7AHsDthEDZV9IAYFOgC3BCIwueg4GDgCGEus+VgIuAhcBRwM/AWR48M/McqGuQJK1gZj8lXv8OmAVsD+wM7EMIEJ3M7JPipLK4JP0fcD3wB+AAwo9LObDUzBZLKjezJcVMY6nzOlDX4EhqCWwvaXVJB0rqDawN3AH0MLO+MbgOBv4qqVlD77okqb2k38Tn/WKjUAUwDtjPzPrEYDkAODT2lfXgmYMX4V1DtCKwMXAu0BHY1MxGxEDaWlJ3YCfgWOBAM1tQtJTWnfbAjZImAFsBLxN+UDYFxgNIOgo4Bfh9qq+sy85zoK7BSOUizewb4CtgM+BpoENcfiDwAaFhpBdwgJm9V5zU1i0zmwi8AgwErjaz74AZwDXAlnHuoUGEH5QPipfS+sXrQF2DkN7xPQbTboT6vRWBx8xsrKQOwNdAhZn9XJzU1o1KrsmmhNznmcAfzezJuDyVkWpuZj/UfUrrLw+grt5L6+d5PKH4Ph24EzDgZELLcktCkf5IM5tbpOTWibRr8gegLTDazN6RtD9wGXAk0JVw19HxjeHmgZrmRXhX7yUCxSmE1uT7gN8CtxCC5lWEIv3awHkNPXjCctfkZOB0Qt/XeyQdY2YPEnKhfyWMSHWjB8/CeAB19ZakbjHHiaRVCHWdexNyVD8BzxGCZ3szuwk4KNYFNliSOsc7rZDUg9Bl67fAXEJufBdJx5nZQ4Qfm33N7O2iJbie8wDq6qVYb7cJsFXMVf0AXAxsSBgMYz/gUWA14G+x287iYqW3tiloSujXeaakVczsTeB4YFdCV6VNCC3u50o6EphnZt8XL9X1nwdQVy/FbjZPEW7D7CVpkJnNAxYBM8xsEbAFMBw41swWNORiqgWLCH1buwJ/ltTazL4GVifUCQN8TLhl81nvqlR93ojk6pVKWpabAnsR7qIZS6j3fApYQAigezb0rkqVXJOOhJzoO8ClhPrPhwi9DzoSum99WIy0NjQeQF29kdayPABoCvxsZrfEluW+wNNm9pCkteO6r4uY5FqXdk32AL4n5DbnEDrKv0Xo67kysBvwspl9VKTkNjgeQF29ExuODiW0JI8h9PV8nhBADwCeMrM7ipfCuhdHmtqHcB32INySuYgw2tQUYIgPilzzvA7UlbzkfeqxMWhboB9h+LnngCdi/eejwP/isgYt7Zp0BnY2s96EblvfAp+a2WfAiYRxT1sUI50NnedAXUlLK6L2JfTnPIoQENoQhqVbIOlM4BUze614qa0badekE6F70kWE8U27E+o4F8ZqjWHAEh8YpHZ4DtSVtESg2I1wR9G7hPq9w4GjY/A8kFCk/6ZoCa1DiWtyEDAU+CKuOhg4NAbPgYQqjpYePGuPj8bkSl6c7OxAwnw9S4ALJa0LPCrpC6Az4fbMRjOup6QDCHcRnWBmSyTdDMwGHpM0glAvfJgPhly7PIC6+uA7YF1gU0mtzGyOmR0taUtCd6XZZvZlcZNY55oQ7rp6gNBI9BrwNiEXOgu4x8ymFC95jYPXgbqSlqrvU5iG+O+Eid/ubwz3s+cSx+88GzjVzJ4tdnoaI8+BupIRR0Ffmni9bNpdM3tV0nmEATCaSbqlMQyEXNkwfanXZnaHpJ+ASyQ1MbMnipbQRsobkVzJSAXP2C0ndXuiEW/1NrORhDtrdid0om/wEg1GK6ReS6pIrL+H0NfzbEkrN/SpSUqNF+Fd0UnaHOhmZv+TdBJhqo0phKmHR5vZ3Dh4iMUA0qyh5z4lrUN4v59JOpUw9cZSMzs2rq8ws8WJ7VfxwZDrnudAXVFJKic0EB0j6SJgB8KdNK8DvwP2lNQ8beCLBntHTRxVqRVhZKn9JR0GHEK4x38tSa8AWJg1s0lqPw+exeE5UFc0iQaipsAuwAmEkZSOjutTo8uPBR6Odxs1aKl6YEk7EW4YABhjZv+J658EVjKz3xYtkW4Zz4G6opC0KqHLDYQBf5sTpuDYStIRAGZ2I/AR4e6aBl+3J6k1YR4ngM+AxwiDgPRM1AvvBTSR9HRxUumSPAfqiiLWaV4B7Enot7gzYcDj3xEGwrjfzP4bt21lZnOKlda6ojDd8p6E8Tw3MLNekrYi3M/+BuGe/4/jtmub2efFS60Dz4G6Oiapl6T/xDrNlwg5zx/MbGFsFHkJuBUYKOlQgIYePCV1l3SqmU0g3FXVj5Abx8zGEhrTNgcOkPSbuNyDZwnwAOrq2kRgC0lXASOBbYCvJD0Cy4LlNEIjyshiJbKuxG5HrYE9JA0G/gFcQGgwOkxSeey+9TCwDuF2TVcivAjv6kwMBkskrUjIab5tZoMltQBuIuRGHyC0wp/Y0HOeKbGP5zbAKcALZnZjbEDrRhiabzXCJHlPeGt7afEA6mpd2vBrK8bRgpoRgsNkMzs21oleB6wHnG5m7xYxybWukjuMKoDtgdMI8xXdKOkYYCvCpHB7N/RrUh95AHW1qpJpODoC0+JtiE35JYgeF7dZ2czmFy/FtS/tmvyeMOXwAjMbJak3IYg+ZWY3x2u0io+qVJq8DtTVqkSgOAE4BngZuF7ShcBSoA+wtaTr4/YNOnjCctfkFODPwNrAUEl/MLMRhLnsD5Z0kpkt8uBZujyAulolqSyO3dmHMGdPN8Kwa1sBVxK6Lm0JXF60RBZBHIrvd0BvwpQbM4DLJB1iZi8D/0eYosSVMB+NydW4tBGDlgKfSDoc6AEcaGbbS+pBmBBuCnCdmX1avBTXvvQ6T8L77g/sB+xmZjtIOhu4XdJiM3ugGOl0VeMB1NWotPq9/YDVCZ3A3wF+JuQ4AdoSuuY8Yg28Ij7tmmxFuA7vmtmceEfW3XHTbwnF9wnFSamrKg+grkZU0qp8KrA/8AShU/hZhEA6WdJLhK45+zeGDuGJ4HkSYWCQccA+krYgTAi3h6T1Ca3tfRvh6Pr1lgdQV1PWAj4HkNQR6BGLpccRJj17HCgndJBfD/jCGvgcRpJam9ns+Hw7whQcOxMGTfki9nO9UdJXxIYkD571i3djctUS76RpTgiS55nZ1bHrzaXAb4AVgL3iCENHEaYe/rh4Ka4bcTzPIcDdZvZSoiFtdWA7Qr/On2I1xxPJsT1d/eE5UFctsXg6V9LuhFkyfzKzGyRNJ7Su/zEGz8MJ0+yOKGJy61I58AlhTM+fCaNKnQL8ZGabA8RrcjQwmlD/6eoZz4G6gqU6vadGR5e0JzCMMGf7o4RcaGfgR0L3pYMb09008RbVM4D1gXOBFsBw4GpgVUJx/ggzm1S0RLpq8QDqCiJpI0Ij0aNm9raktYDbCK3t/YGT4xQdXYEOwAcNvX5P0i7A+nEcUyS1A14APgDmEGYVbUro+7kyMMx86uF6zYvwrlA/EXJWO0tajTCC0H1mdr2k/wLPSFrdzK4A3i9mQuvQD4S7rL4zs/sI3bRuIEzFPIjQE+EOM7upiGl0NchzoK5gkroQbkXcmtAQcnZiXS/gXkLn+TkNva9niqSewPPAEkIu/J64vDNhBP42hOL8gsZyTRoyD6Aub3Ew39lmNjsxn1FHQmvzB8CTZjY5sX1TM1tUrPQWi6RNCGOZHm9m9ybmOVoHmGtm3xU5ia6G+L3wLi+xgegBQkNIan5ymdlnwD8J01DsH4NHyk91n9LiM7N3CF2WbpQ0ON7Oipl96sGzYfEA6nKKXZSuBgZYmKe8DJYLoh8SWtw3BvrGAYJpzEVUMxtHuLPohjiMn2uAvBHJZRWD57+B7wkd5onFUVkUn0+VdA6hbq9R5jzTmdmbcdCUH4udFlc7vA7UZSRpR+BaINU4dBLwbzMbFtcnB8lIH23IuQbPc6Aum3nAsWY2LnZVugcYFGPl44ncp3nwdI2RB1D3K3Gw31UBEUYOwsxmSXqKMHrQcTFuPuGB0zVmHkDdciTtRRiT8kFgF+AdSZ+Y2cVm9l0MokuBv8SBf58pZnqdKyavA3XLSOoGPAYcZWZjJLUmdJI/gjAF8T/idi0J3XReawzjeTqXiedAXbIBqD0wPgbP8thh/iXCyEL7SGplZnPM7HtJD3rx3TV23g/UQewcT5jYrJnCnO2prkoLgbHAtsAWqR08eDrnAbTRi/doD4n3cH8OrAEcEAOkACxMq/si3p/RueV4AHUtCS3r+xOGWDsZuDaOHt8KQNJhhLtqphcrkc6VIm9EaqRS9Znx+UaEyc6a8csgyP8E5sdHN0Ku9J0iJde5kuQBtBGStCthnMqnCWNVfkHIhQ4mDPh7IzCLMH/PasDnZua5T+fSeBG+cZoJdAQGAFsRhl7bm3Cv+xxCMb6jmX1kZq978HSuch5AGyEzm0BoUV9EGEW9D6FzfA/gIMI8PoMkNSlaIp2rB7wI34jFWzZfAE41s9sllQObEQLqY2bWWKbicK4gHkAbuRhEnwPONbMbip0e5+oTvxOpkYsjLe0KjJO00MxuLXaanKsvPAfqAJC0OfCjmX1Q7LQ4V194AHXOuQJ5K7xzzhXIA6hzzhXIA6hzzhXIA6hzzhXIA6hzzhXIA6grOZKWSJogaZKkByStVI1j9Zb0RHy+j6SzsmzbStIJBZzjfEl/ynd52ja3S9q/CufqJGlSVdPoaocHUFeKFphZdzPbGPiJMErUMgqq/N01s2FmdmmWTVoBVQ6grvHyAOpK3SvAejHn9b6kG4DxwNqS+kh6VdL4mFNtDiCpr6TJkkYBv08dSFJ/SdfH5+0lPSLp7fjYljgWasz9Xh63O1PSOEkTJV2QONa5kj6Q9AKwQa43IenYeJy3JT2UlqveVdIrkj6UtHfcvlzS5YlzH1fdC+lqngdQV7IkVQB7AKmBnDcA7jSzzQkDPQ8BdjWzLYA3gNMlrQgMBX4H7EAY07Qy1wIvm9lmhJGp3gXOAj6Kud8zJfUBugC9gO5AD0k7SuoBHAxsTgjQW+bxdh42sy3j+d4HBibWdQJ2AvYCborvYSDwvZltGY9/rKR18ziPq0N+L7wrRc0kTYjPXwFuIczV9KmZvRaXb00YKX+0JIAVgFeBDYFPzGwKgKT/AoMqOcfOwJEAZrYE+D5O45zUJz7eiq+bEwJqC+ARM/sxnmNYHu9pY0kXEaoJmgPPJtbdb2ZLgSmSPo7voQ+waaJ+tGU894d5nMvVEQ+grhQtMLPuyQUxSM5PLgKeN7ND0rbrThhdvyYIuMTM/p12jtMKOMftwH5m9rak/kDvxLr0Y6Um9DvZzJKBFkmdqnheV4u8CO/qq9eA7SStByBpJUnrA5OBdeNsoxDmeqrMcOD4uG+5pFWAufwyxTOEXOLRibrVNSW1I4zg309SM0ktCNUFubQApsdBqg9LW3eApLKY5t8AH8RzH58a1FrS+pJWzuM8rg55DtTVS2b2bczJ3SOpaVw8xMw+lDQIeFLSTGAUsHElhzgVuFnSQGAJcLyZvSppdOwm9HSsB+0KvBpzwPOAw81svKT7gAnAp4Rqhlz+CoyN27/D8oH6A+BloD0w2MwWSvoPoW50vMLJvwX2y+/quLriozE551yBvAjvnHMF8gDqnHMF8gDqnHMF8gDqnHMF8gDqnHMF8gDqnHMF8gDqnHMF+n+t7a+FBpVGJwAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "=========== PREDICTION =============\n",
      "Number of records: 393 [0, 0.0% filtered out]\n",
      "prediction accuracy: 93.89% [vs. 14.76% if assumed all Health related ]\n",
      "prediction recall: 82.69%\n",
      "prediction precision: 74.14%\n",
      "prediction f1: 78.18%\n",
      "===================================\n",
      "\n",
      "\n"
     ]
    }
   ],
   "source": [
    "test = pickle.load(open(\"./data/test.pkl\", \"rb\" ))\n",
    "\n",
    "print(\"Predicting ...\")\n",
    "predict1 = predict_fasttext(test, text_field_pp, fasttext_model_name + \".ftz\")\n",
    "\n",
    "output_metrics(predict1[response_field],predict1['prediction'], len(test), response_field)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Not too bad. As mentioned, this is a quick demo, with more data and hyper-parameter tuning we could improve performance.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Using a model trained on Bill categories in one state, predict categories on Bills in another state\n",
    "\n",
    "Let's read bills data for South Dakota."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading https://datakind-public.s3.amazonaws.com/NLP_data/Bills_V2.zip\n",
      "2713 bills total\n",
      "BY YEAR\n",
      "year\n",
      "2009    273\n",
      "2010    244\n",
      "2011    224\n",
      "2012    252\n",
      "2013    260\n",
      "2014    248\n",
      "2015    257\n",
      "2016    236\n",
      "2017    219\n",
      "2018    282\n",
      "2019    218\n",
      "Name: bill_id, dtype: int64\n",
      "\n",
      "Done\n"
     ]
    }
   ],
   "source": [
    "state = 'SD'\n",
    "fl = './data/' + state + '_Bills_2.csv'\n",
    "\n",
    "# Get the data files if we don't have them\n",
    "if not os.path.isfile(fl):\n",
    "    s3_link = \"https://datakind-public.s3.amazonaws.com/NLP_data/Bills_V2.zip\"\n",
    "    print(\"Downloading \" + s3_link)\n",
    "    resp = requests.get(s3_link).content\n",
    "    zipfile = ZipFile(BytesIO(resp))\n",
    "    zipfile.extractall(ddir)\n",
    "\n",
    "df2 = pd.read_csv(fl, index_col=0)\n",
    "df2['state'] = fl.replace('./data/','').replace('_Bills_2.csv','')\n",
    "\n",
    "print('{} bills total'.format(df2.shape[0]))\n",
    "print('BY YEAR')\n",
    "print(df2.groupby('year').count()['bill_id'])\n",
    "print()\n",
    "\n",
    "df2['word_count'] = df2['text'].apply(lambda x: len(str(x).split(\" \")))\n",
    "\n",
    "print(\"Done\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Predict\n",
    "\n",
    "Using our model, let's predict which South Dakota bills are health-related."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Pre-processing text ...\n",
      "fastext_model-Healthrelated-0.2-100-550-3---.ftz\n",
      "\n",
      "\n",
      "Example of bills predicted as being health-related:\n",
      "\n",
      "----------------------------------------------\n",
      " An Act to  revise the definition of an accredited prevention or treatment facility for the purposes of the prevention and treatment of alcohol and drug abuse. \n",
      "----------------------------------------------\n",
      " An Act to  revise certain drug registration fees and license fees related to health care facilities. \n",
      "----------------------------------------------\n",
      " An Act to provide for the monitoring of the prescribing and dispensing of controlled substances. \n",
      "----------------------------------------------\n",
      " An Act to establish certain legislative findings pertaining to the decision of a pregnant mother considering termination of her relationship with her child by an abortion, to establish certain procedures to better insure that such decisions are voluntary, uncoerced, and informed, and to revise certain causes of action for professional negligence relating to performance of an abortion. \n",
      "----------------------------------------------\n",
      " An Act to revise the definition of accredited prevention or treatment facilities and to authorize the Division of Drug and Alcohol Abuse to inspect these facilities and access their records. \n",
      "----------------------------------------------\n",
      " An Act to provide for the regulation and licensure of birth centers. \n",
      "----------------------------------------------\n",
      " An Act to establish network adequacy standards, quality assessment and improvement requirements, utilization review and benefit determination requirements, and grievance procedures for managed health care plans, and to repeal certain standards for managed health care plans. \n",
      "----------------------------------------------\n",
      " An Act to revise certain certification levels of emergency medical technicians. \n",
      "----------------------------------------------\n",
      " An Act to establish policies for youth athletes with concussions resulting from participation in youth athletic activities. \n",
      "----------------------------------------------\n",
      " An Act to repeal certain provisions relating to licenses issued by the Board of Pharmacy. \n"
     ]
    }
   ],
   "source": [
    "# Restrict to the state we're interested in\n",
    "test2       = df2[df2['state'] == state]\n",
    "\n",
    "# Concatenate title and text\n",
    "test2[\"model_text\"] = test2['title'] + '\\n\\n\\n' + test2['text']\n",
    "\n",
    "print(\"Pre-processing text ...\")\n",
    "test2 = preprocessFastText(test2, text_field, lemmatize=False, removeStopWords=False)\n",
    "\n",
    "# Make our prediction\n",
    "predict1 = predict_fasttext(test2, text_field_pp, fasttext_model_name + \".ftz\")\n",
    "\n",
    "# Isolate cases where the model thought the bill was 'health and safety' related\n",
    "health_related_bills = predict1.loc[predict1[\"prediction\"]==1]\n",
    "\n",
    "# Output excerpts of the Bill text\n",
    "print(\"\\n\\nExample of bills predicted as being health-related:\\n\")\n",
    "for t in health_related_bills[\"title\"][0:10]:\n",
    "    print(\"----------------------------------------------\")\n",
    "    print(t)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "\n",
    "\n",
    "Those seem like health related Bills, the model was able to classify them reasonably well just using titles.\n",
    "\n",
    "So, if we train a model to recognize health bills using data from one state, we can then use it to predict the same for Bills data in another state. Even if that state doesn't have health bills labelled.\n",
    "\n",
    "Hope this has been helpful, good luck!\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    ".\n"
   ]
  }
 ],
 "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": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}