UCSD-Newspapers.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "First let's change our current working directory to the folder where the data is located using `os.chdir`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import os\n",
    "os.chdir(\"./ucsdHistory_raw\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's confirm that the changes are in effect by checking the current working directory:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "os.getcwd()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's open one of the newspaper's files and have a look at the contents. We'll start with the *Student Newspapers*, just because it's listed first alphabetically. We'll list all the files in the directory, open the first one and inspect the content by displaying the first 2000 characters in the file."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "journal_name = 'Student Newspapers'\n",
    "path = os.path.join(os.getcwd(), journal_name)\n",
    "files = os.listdir(path)\n",
    "files[:10]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`os.listdir()` returns the names of the files but not their path, which we need to be able to open them. For now we'll just use `os.path.join()` again."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "file = os.path.join(path, files[0])\n",
    "with open(file, 'r') as fp:\n",
    "    content = fp.read()\n",
    "    \n",
    "content[:2000]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Woops. Encoding error. Let's fix this by specifying the encoding"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "file = os.path.join(path, files[0])\n",
    "with open(file, 'r', encoding = 'utf8') as fp:\n",
    "    content = fp.read()\n",
    "    \n",
    "content[:2000]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "First, let's get rid of the hyphenation using regular expressions (for which we'll need the **re** package). We'll create a specific function for it:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import re\n",
    "def preprocess(text):\n",
    "    return re.sub('\\-\\n+', '', text)\n",
    "\n",
    "content = preprocess(content)\n",
    "content[:2000]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we want to be able to load all the files in the directory and apply the preprocessing to them. To do that, we'll first define a function that will return the full path of all the files in the directory and will makes sure that we are only dealing with text files (so that we don't try to open a folder or another filetype that would cause an error):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def load_files(directory):\n",
    "    path = os.path.join(os.getcwd(), directory)\n",
    "    files = os.listdir(path)\n",
    "    files = [os.path.join(path, _) for _ in files if '.txt' in _]\n",
    "    return files\n",
    "\n",
    "load_files(path)[:10]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Once we have all the files, we'll create a new function `load_data` to each file one by one, preprocess it and store the result in memory in an array"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def load_data(files):\n",
    "    X = []\n",
    "    for file in files:\n",
    "        with open(file, 'r', encoding = 'utf8') as fp:\n",
    "            content = fp.read()\n",
    "            content = preprocess(content)\n",
    "        X.append(content)\n",
    "    return X"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's actually run the function and get the data:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "files = load_files(journal_name)\n",
    "documents = load_data(files)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "len(documents)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we'll vectorize the documents. We'll only keep terms that appear in less than 80% of the documents and at least twice in the corpus. We'll limit the number of features to 5000."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from sklearn.feature_extraction.text import TfidfVectorizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "vectorizer = TfidfVectorizer(max_df=0.8, min_df=2, max_features=5000, stop_words='english')\n",
    "X = vectorizer.fit_transform(documents)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Next, we'll apply non-negative matrix factorization to do topic modeling. We'll try with 25 topics first:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from sklearn.decomposition import NMF\n",
    "nmf = NMF(n_components=25, init = 'nndsvd').fit(X)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's have a look at nmf.components_ to understand the output and try to interprete the results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "np.shape(nmf.components_)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "nmf.components_"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "nmf.components_[0].argsort()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "vocabulary = vectorizer.get_feature_names()\n",
    "print(vocabulary[2499])\n",
    "print(vocabulary[4960])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "nmf.components_[0][2499]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "nmf.components_[0][4960]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "nmf.components_[0].argsort()[::-1]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's now print the top 10 words for the first topic:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "[vocabulary[i] for i in nmf.components_[0].argsort()[::-1][:10]]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's create a function that will allow us to do that for all the topics:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def display_topics(model, vocabulary, no_words):\n",
    "    for index, topic in enumerate(model.components_):\n",
    "        print(\"Topic\", index, \":\")\n",
    "        print(\" \".join([vocabulary[i] for i in topic.argsort()[::-1][:no_words]]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "display_topics(nmf, vocabulary, 15)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Note that we can also use nmf to get the distribution of each of our topic within a specific document (what percentage of document a concerns topic 1, 2 or 3? ) by normalizing the results of the NMF over the corpus:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "doc_topics = nmf.transform(X)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "np.shape(doc_topics)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "doc_topics[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "sum(doc_topics[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "doc_topics = doc_topics / np.sum(doc_topics, axis=1, keepdims=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "doc_topics[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's have a look at what metadata we can work with by opening the csv file corresponding to the corpus we're looking at. We will use the **pandas** package to load the csv file."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "# we've stored the name of the journal in a variable earlier, we can reuse this here so that we can more easily switch to another corpus in the future\n",
    "df = pd.read_csv(journal_name + '.csv')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's look at the columns"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "df.columns"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "And the first entries..."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "df.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We notice that the filenames listed do not exactly correspond to the names of the raw text files we have used before, nor are they listed in the same order:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "files[:10]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "For that reason we will need to process our filenames and match them to the appropriate row in the dataframe as well as to the appropriate topic distribution in the NMF matrix we created earlier from the corpus.\n",
    "\n",
    "We notice that the Object Unique ID of each row can be found in the filenames of the text files we have, so let's use that to match the raw text to the rows of the csv. Let's do a test with the first file:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "file_id = files[0].split('/')[-1].split('-')[0]\n",
    "file_id"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "df['Object Unique ID'].str.contains(file_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "df[df['Object Unique ID'].str.contains(file_id)]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's try to extract the date for this particular row. There are several date columns (BeginDate, EndDate, ...) but they all seem to be equivalent, so let's go with BeginDate..."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "test_date = df['Begin date'][df['Object Unique ID'].str.contains(file_id)]\n",
    "test_date"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We are getting the right value, but the format of the output tells us that the data type returned is not exactly standard. Let's confirm this:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "type(test_date)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This is a variable type specific to pandas, so we might run into trouble when trying to perform standard operations over it or sending it to some package's function. Let's make sure we only get the value and not the **pandas** Series cell:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "test_date = df['Begin date'][df['Object Unique ID'].str.contains(file_id)].values\n",
    "print(test_date)\n",
    "print(type(test_date))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Almost there..."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "test_date = df['Begin date'][df['Object Unique ID'].str.contains(file_id)].values[0]\n",
    "print(test_date)\n",
    "print(type(test_date))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Ok. So now we have a standard string. Which is good, but not ideal in case we want to perform some calculations on the date (substraction or extracting the month, year...). We can use the **dateparser** package to help us:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import dateparser\n",
    "parsed = dateparser.parse(test_date)\n",
    "print(type(parsed))\n",
    "print(parsed)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "A little parsing mistake here but we can handle that later. Let's iterate over all our files and store the results:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from collections import defaultdict\n",
    "yearbook = defaultdict(list)\n",
    "\n",
    "for i, file in enumerate(files):\n",
    "    doc_year = dateparser.parse(df['Begin date'][df['Object Unique ID'].str.contains(file.split('/')[-1].split('-')[0])].values[0]).year\n",
    "    \n",
    "    # handles the 2k bug\n",
    "    if doc_year > 2018:\n",
    "        doc_year -= 100\n",
    "    \n",
    "    # fetches the corresponding distribution for the document. our files and topic distribution matrix have the same index\n",
    "    topic_distribution = doc_topics[i]\n",
    "    \n",
    "    # stores the result in an array\n",
    "    yearbook[doc_year].append(topic_distribution)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's compute the mean distribution of topics for each year (maybe absolute values would be better) and store the result in a dataframe"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "finals = {v: np.mean(yearbook[v], axis = 0) for v in yearbook}\n",
    "res = pd.DataFrame.from_dict(finals)\n",
    "res.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's transpose and plot the data:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "\n",
    "res.transpose().plot(figsize=(12,12))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}