notes_Mallet_topics.ipynb

{
 "metadata": {
  "name": "notes_Mallet_topics"
 }, 
 "nbformat": 2, 
 "worksheets": [
  {
   "cells": [
    {
     "cell_type": "markdown", 
     "source": [
      "# Topic models from Mallet over text of turns "
     ]
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Running Mallet on the text of turns and considering them in a similar task setting as the unigram baseline for predicting app-user reactions."
     ]
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "## Text and reactions"
     ]
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "First, some text of turns to play with."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [
      "import pandas as pd", 
      "import reactions", 
      "import nltk", 
      "import random", 
      "import matplotlib.pyplot as plt", 
      "import os", 
      "from pandas.tools.plotting import scatter_matrix"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 77
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "The raw transcript and reaction info not yet grouped into turns."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "%time r = reactions.link_reactions_to_transcript('data/reactions_oct3_4project.csv','corpora/oct3_coded_transcript_sync.csv')"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "CPU times: user 8.59 s, sys: 0.49 s, total: 9.07 s", 
        "Wall time: 9.08 s"
       ]
      }
     ], 
     "prompt_number": 9
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "print ' '.join(r.columns)", 
      "r2 = r.copy()", 
      "del r2[\"Sync'd start\"]", 
      "del r2[\"Sync'd end\"]", 
      "del r2[\"Time\"]", 
      "del r2[\"Speaker\"]"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "Frame QuestionTopic Reaction_what Reaction_who Speaker Sync'd end Sync'd start Time Tone Topic Transcript UserID start statement turn Speaker_name"
       ]
      }
     ], 
     "prompt_number": 50
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "%time p = reactions.split_reactions_file('data/reactions_oct3_4project.csv')['quest_political']", 
      "' '.join(p.columns)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "CPU times: user 4.79 s, sys: 0.35 s, total: 5.14 s", 
        "Wall time: 5.12 s"
       ]
      }, 
      {
       "output_type": "pyout", 
       "prompt_number": 48, 
       "text": [
        "'UserID party_1 political_views_2 candidate_choice_3 confidence_in_choice_4 likely_to_vote_5 immigration_priority_6 health_care_priority_7 foreign_policy_priority_8 abortion_priority_9 economy_priority_10 immigration_party_11 health_care_party_12 foreign_policy_party_13 abortion_party_14 economy_party_15 interested_23 news_sources_24 economy_candidate_27 foreign_policy_candidate_28 candidate_preferred_29'"
       ]
      }
     ], 
     "prompt_number": 48
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [
      "p2 = p[['UserID','party_1','political_views_2','candidate_choice_3','confidence_in_choice_4','likely_to_vote_5','candidate_preferred_29']]", 
      "p2['party'] = p2.party_1.apply(lambda a: {'closest to democratic party':'democrat', ", 
      "                                          'lean democrat':'democrat',", 
      "                                          'lean republican':'republican',", 
      "                                          'closest to republican party':'republican'}.get(a,'other'))"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 13
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "%time r3 = r2.merge(p2[['UserID','party']])", 
      "' '.join(r3.columns)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "CPU times: user 0.63 s, sys: 0.04 s, total: 0.67 s", 
        "Wall time: 0.67 s"
       ]
      }, 
      {
       "output_type": "pyout", 
       "prompt_number": 47, 
       "text": [
        "'Frame QuestionTopic Reaction_what Reaction_who Tone Topic Transcript UserID start statement turn Speaker_name party'"
       ]
      }
     ], 
     "prompt_number": 47
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Group by turns"
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [
      "st = r3.groupby(['statement']).first()[['Speaker_name','Transcript','turn']]", 
      "t = pd.DataFrame({'speaker':st.groupby('turn').first().Speaker_name, ", 
      "                  'reactions':r3.groupby('turn').count().Speaker_name, ", 
      "                  'statements':st.groupby('turn').count().turn, ", 
      "                  'text':st.groupby('turn').apply(lambda x: ''.join(x.Transcript)),", 
      "                  'agree':r3[r3.Reaction_what=='Agree'].groupby('turn').count().turn,", 
      "                  'agree_dem':r3[(r3.party=='democrat') & (r3.Reaction_what=='Agree')].groupby('turn').count().turn,", 
      "                  'agree_rep':r3[(r3.party=='republican') & (r3.Reaction_what=='Agree')].groupby('turn').count().turn,", 
      "                  'disagree':r3[r3.Reaction_what=='Disagree'].groupby('turn').count().turn,", 
      "                  'disagree_dem':r3[(r3.party=='democrat') & (r3.Reaction_what=='Disagree')].groupby('turn').count().turn,", 
      "                  'disagree_rep':r3[(r3.party=='republican') & (r3.Reaction_what=='Disagree')].groupby('turn').count().turn,", 
      "                  })", 
      "t['words'] = t.text.apply(lambda txt: [t.lower() for t in nltk.tokenize.word_tokenize(txt) if t.isalpha()])", 
      "t['word_count'] = t.words.apply(lambda words: len(words))", 
      "t['r_per_st'] = 1.0 * t.reactions / t.statements", 
      "t['r_per_w'] = 1.0 * t.reactions / t.word_count", 
      "t['a_to_d_dems'] = t.agree_dem / t.disagree_dem", 
      "t['a_to_d_reps'] = t.agree_rep / t.disagree_rep"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 15
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Unigram features."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [
      "ranked_unigrams = nltk.FreqDist([w for word_list in t.words for w in word_list]).keys()", 
      "MAX_FEATURES = 700 # avoid overfitting", 
      "t['unigrams'] = t.words.apply(lambda words: {w:True for w in words if w in ranked_unigrams[:MAX_FEATURES] and not w in stopwords.words('english')})", 
      "t['unigram_count'] = t.unigrams.apply(lambda unigrams: len(unigrams))"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 37
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Remove short turns."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "MIN_WORDS = 30", 
      "t2 = t[t.word_count >= MIN_WORDS]", 
      "print len(t),'->',len(t2)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "190 -> 71"
       ]
      }
     ], 
     "prompt_number": 16
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Label turns."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [
      "t2['label'] = t2.a_to_d_dems >= t2.a_to_d_dems.quantile(.5)"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 17
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "' '.join(t2.columns)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "pyout", 
       "prompt_number": 46, 
       "text": [
        "'agree agree_dem agree_rep disagree disagree_dem disagree_rep reactions speaker statements text words word_count r_per_st r_per_w a_to_d_dems a_to_d_reps unigrams unigram_count label'"
       ]
      }
     ], 
     "prompt_number": 46
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "## LDA features"
     ]
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Write text to input file for Mallet."
     ]
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "A folder for holding Mallet files for this topics vs unigram comparison."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "!rm -r tmp.tVSu", 
      "!mkdir tmp.tVSu"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 213
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "f = open('tmp.tVSu/in.txt','w')", 
      "for i,r in t2.iterrows():", 
      "    f.write(\"%d NA %s\\n\" % (i, ' '.join(r['words'])))", 
      "f.close()"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 214
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "!head -n 1 tmp.tVSu/in.txt"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "1 NA good evening from the magness arena at the university of denver in denver jim lehrer of the pbs newshour and i welcome you to the first of the presidential debates between president barack obama the democratic nominee and former massachusetts governor mitt romney the republican debate and the next three two presidential one presidential are sponsored by the commission on presidential minutes will be about domestic issues and will follow a format designed by the will be six roughly segments with answers for the first question then open discussion for the remainder of each of people offered suggestions on segment subjects or questions via the internet and other means but i made the final selectionsand for the record they were not submitted for approval to the commission or the segments as i announced in advance will be three on the economy and one each on health care the role of government and governing with an emphasis throughout on differences specifics and candidates will also have closing audience here in the hall has promised to remain cheers applause boos hisses among other noisy distracting things so we may all concentrate on what the candidates have to is a noise exception right now though as we welcome president obama and governor gentlemen welcome to you start the economy segment let begin with are the major differences between the two of you about how you would go about creating new jobs you have two minutes each of you have two minutes to coin toss has determined president you go first"
       ]
      }
     ], 
     "prompt_number": 245
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Use Mallet to prep the input file before training topics."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "!mallet/bin/mallet import-file --input tmp.tVSu/in.txt --output tmp.tVSu/in.mallet --keep-sequence --remove-stopwords", 
      "!ls tmp.tVSu/"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "in.mallet  in.txt     keys.txt   log.txt    state.gz   topics.txt"
       ]
      }
     ], 
     "prompt_number": 227
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Train topics."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "!mallet/bin/mallet train-topics --input tmp.tVSu/in.mallet --num-topics 30 --output-state tmp.tVSu/state.gz --optimize-interval 10 --output-topic-keys tmp.tVSu/keys.txt --output-doc-topics tmp.tVSu/topics.txt &2> tmp.tVSu/log.txt"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 228
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Wait until output files appear."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "!ls tmp.tVSu/"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "in.mallet  in.txt     keys.txt   log.txt    state.gz   topics.txt"
       ]
      }
     ], 
     "prompt_number": 222
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "!head -n 2 tmp.tVSu/topics.txt"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "#doc name topic proportion ...", 
        "0\t1\t13\t0.45926618183203816\t8\t0.35710755777295733\t24\t0.0788980922235202\t23\t0.04730063985040274\t28\t0.029975056448629345\t4\t0.009686990895312646\t21\t0.004341785068119908\t3\t0.0014216965243021948\t14\t0.0013318089480645195\t15\t0.0012738606858953686\t20\t0.00107052194643624\t12\t9.539719251060424E-4\t11\t7.910957781564223E-4\t25\t5.310775209234101E-4\t2\t5.305902722393401E-4\t6\t5.30106004709156E-4\t26\t5.286668084779E-4\t1\t4.244008874088659E-4\t17\t4.0297150830554166E-4\t19\t4.0261376798926167E-4\t22\t3.700997635907186E-4\t0\t3.6937669411317057E-4\t5\t3.6620073986661167E-4\t29\t3.621127098850232E-4\t27\t3.5949360578337887E-4\t7\t3.229038652824974E-4\t10\t3.217396348623458E-4\t18\t2.6964341416511753E-4\t16\t2.6619109832790326E-4\t9\t2.2255180512863995E-4\t"
       ]
      }
     ], 
     "prompt_number": 247
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Read Mallet output."
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "turns = []", 
      "feats = []", 
      "header = True", 
      "for line in open('tmp.tVSu/topics.txt','r'):", 
      "    if header:", 
      "        header = False", 
      "        continue", 
      "    fs = line.split('\\t')", 
      "    turns.append(int(fs[1]))", 
      "    feats.append({int(f):float(v) for f,v in zip(fs[2::2], fs[3::2])})", 
      "#mo = pd.DataFrame({'turn':turns, 'topics':feats})", 
      "mo = pd.DataFrame({'topics':feats}, index=pd.Int64Index(turns,name='turn'))", 
      "#mo = mo.set_index('turn')", 
      "mo.head(2)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "html": [
        "<div style=\"max-height:1000px;max-width:1500px;overflow:auto;\">", 
        "<table border=\"1\" class=\"dataframe\">", 
        "  <thead>", 
        "    <tr style=\"text-align: right;\">", 
        "      <th></th>", 
        "      <th>topics</th>", 
        "    </tr>", 
        "    <tr>", 
        "      <th>turn</th>", 
        "      <th></th>", 
        "    </tr>", 
        "  </thead>", 
        "  <tbody>", 
        "    <tr>", 
        "      <th>1</th>", 
        "      <td> {0: 0.00036937669411317057, 1: 0.0004244008874...</td>", 
        "    </tr>", 
        "    <tr>", 
        "      <th>2</th>", 
        "      <td> {0: 0.0003875201854763252, 1: 0.01026908909979...</td>", 
        "    </tr>", 
        "  </tbody>", 
        "</table>", 
        "</div>"
       ], 
       "output_type": "pyout", 
       "prompt_number": 271, 
       "text": [
        "                                                 topics", 
        "turn                                                   ", 
        "1     {0: 0.00036937669411317057, 1: 0.0004244008874...", 
        "2     {0: 0.0003875201854763252, 1: 0.01026908909979..."
       ]
      }
     ], 
     "prompt_number": 271
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "### Merge topic features with other data"
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "t3 = t2.join(mo)", 
      "t3[['text','label','unigrams','topics']].head(2)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "html": [
        "<div style=\"max-height:1000px;max-width:1500px;overflow:auto;\">", 
        "<table border=\"1\" class=\"dataframe\">", 
        "  <thead>", 
        "    <tr style=\"text-align: right;\">", 
        "      <th></th>", 
        "      <th>text</th>", 
        "      <th>label</th>", 
        "      <th>unigrams</th>", 
        "      <th>topics</th>", 
        "    </tr>", 
        "    <tr>", 
        "      <th>turn</th>", 
        "      <th></th>", 
        "      <th></th>", 
        "      <th></th>", 
        "      <th></th>", 
        "    </tr>", 
        "  </thead>", 
        "  <tbody>", 
        "    <tr>", 
        "      <th>1</th>", 
        "      <td> Good evening from the Magness Arena at the Uni...</td>", 
        "      <td> True</td>", 
        "      <td> {'all': True, 'domestic': True, 'questions': T...</td>", 
        "      <td> {0: 0.00036937669411317057, 1: 0.0004244008874...</td>", 
        "    </tr>", 
        "    <tr>", 
        "      <th>2</th>", 
        "      <td> Well, thank you very much, Jim, for this oppor...</td>", 
        "      <td> True</td>", 
        "      <td> {'sector': True, 'all': True, 'code': True, 'j...</td>", 
        "      <td> {0: 0.0003875201854763252, 1: 0.01026908909979...</td>", 
        "    </tr>", 
        "  </tbody>", 
        "</table>", 
        "</div>"
       ], 
       "output_type": "pyout", 
       "prompt_number": 282, 
       "text": [
        "                                                   text label  \\", 
        "turn                                                            ", 
        "1     Good evening from the Magness Arena at the Uni...  True   ", 
        "2     Well, thank you very much, Jim, for this oppor...  True   ", 
        "", 
        "                                               unigrams  \\", 
        "turn                                                      ", 
        "1     {'all': True, 'domestic': True, 'questions': T...   ", 
        "2     {'sector': True, 'all': True, 'code': True, 'j...   ", 
        "", 
        "                                                 topics  ", 
        "turn                                                     ", 
        "1     {0: 0.00036937669411317057, 1: 0.0004244008874...  ", 
        "2     {0: 0.0003875201854763252, 1: 0.01026908909979...  "
       ]
      }
     ], 
     "prompt_number": 282
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "##Experiment with classification"
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [
      "ex = t3"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 313
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "train_rows = random.sample(ex.index, len(ex)*9/10)", 
      "trn = ex.ix[train_rows]", 
      "tst = ex.drop(train_rows)", 
      "print len(trn)", 
      "print len(tst)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "63", 
        "8"
       ]
      }
     ], 
     "prompt_number": 314
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "%time cl = nltk.NaiveBayesClassifier.train(zip(trn.topics, trn.label))"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "CPU times: user 0.02 s, sys: 0.00 s, total: 0.02 s", 
        "Wall time: 0.02 s"
       ]
      }
     ], 
     "prompt_number": 315
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "nltk.classify.accuracy(cl, zip(tst.topics, tst.label))"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "pyout", 
       "prompt_number": 316, 
       "text": [
        "0.625"
       ]
      }
     ], 
     "prompt_number": 316
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "cl.show_most_informative_features(10)"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "Most Informative Features", 
        "                      22 = 0.003089471295239396  False : True   =      1.7 : 1.0", 
        "                       9 = 0.001857789388941636  False : True   =      1.7 : 1.0", 
        "                       4 = 0.0026970996127855377  False : True   =      1.7 : 1.0", 
        "                      13 = 0.0036340230223573552  False : True   =      1.7 : 1.0", 
        "                       7 = 0.0026954954340779884  False : True   =      1.7 : 1.0", 
        "                       5 = 0.003056923525529371  False : True   =      1.7 : 1.0", 
        "                      25 = 0.004433260752509495  False : True   =      1.7 : 1.0", 
        "                      12 = 0.007963444371012015  False : True   =      1.7 : 1.0", 
        "                      15 = 0.010633770702863553  False : True   =      1.7 : 1.0", 
        "                       6 = 0.004425150854173767  False : True   =      1.7 : 1.0"
       ]
      }
     ], 
     "prompt_number": 317
    }, 
    {
     "cell_type": "markdown", 
     "source": [
      "Multiple train/test"
     ]
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [
      "ex2 = t3"
     ], 
     "language": "python", 
     "outputs": [], 
     "prompt_number": 292
    }, 
    {
     "cell_type": "code", 
     "collapsed": false, 
     "input": [
      "accs = []", 
      "for i in range(50):", 
      "    train_rows2 = random.sample(ex2.index, len(ex2)*9/10)", 
      "    trn2 = ex2.ix[train_rows2]", 
      "    tst2 = ex2.drop(train_rows2)", 
      "    ", 
      "    # ~65%", 
      "    cl2 = nltk.NaiveBayesClassifier.train(zip(trn2.topics, trn2.label))", 
      "    ", 
      "    # error..", 
      "    #cl2 = nltk.classify.MaxentClassifier.train(zip(trn2.topics, trn2.label))", 
      "    ", 
      "    # ~50%", 
      "    #cl2 = nltk.classify.DecisionTreeClassifier.train(zip(trn2.topics, trn2.label))", 
      "    ", 
      "    # error..", 
      "    #cl2 = nltk.classify.ConditionalExponentialClassifier.train(zip(trn2.topics, trn2.label))", 
      "    ", 
      "    accs.append(nltk.classify.accuracy(cl2, zip(tst2.topics, tst2.label)))", 
      "print mean(accs)", 
      "a2 = pd.DataFrame({'accuracy':accs})", 
      "figsize(5,2)", 
      "a2.accuracy.hist()", 
      "title('accuracy')", 
      "show()"
     ], 
     "language": "python", 
     "outputs": [
      {
       "output_type": "stream", 
       "stream": "stdout", 
       "text": [
        "0.645"
       ]
      }, 
      {
       "output_type": "display_data", 
       "png": "iVBORw0KGgoAAAANSUhEUgAAATsAAACWCAYAAAChKS9DAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n8H/W0HOA1hkGCwAAAABJRU5ErkJggg==\n"
      }
     ], 
     "prompt_number": 327
    }, 
    {
     "cell_type": "code", 
     "collapsed": true, 
     "input": [], 
     "language": "python", 
     "outputs": []
    }
   ]
  }
 ]
}