7 Vector Space Models & Text Classification

7.1 Seminar

7.1.1 Exercise

This exercise is designed to get you working with quanteda. The focus will be on exploring the package and getting some texts into the corpus object format. quanteda package has several functions for creating a corpus of texts which we will use in this exercise.

  1. Getting Started.

    You will first need to install some packages:

    install.packages(c("tm", "readtext", "quanteda"))

    You will also need to install the package quantedaData from github using the install_github function from the devtools package:

  2. Exploring quanteda functions.


    Look at the Quick Start vignette, and browse the manual for quanteda. You can use example() function for any function in the package, to run the examples and see how the function works. Of course you should also browse the documentation, especially ?corpus to see the structure and operations of how to construct a corpus. The website http://quanteda.io has extensive documentation.

  3. Making a corpus and corpus structure

    1. From a vector of texts already in memory.

      The simplest way to create a corpus is to use a vector of texts already present in R’s global environment. Some text and corpus objects are built into the package, for example data_char_ukimmig2010 is the UTF-8 encoded set of 9 UK party manifesto sections from 2010, that deal with immigration policy. addresses. Try using corpus() on this set of texts to create a corpus.

      Once you have constructed this corpus, use the summary() method to see a brief description of the corpus. The names of the corpus data_char_ukimmig2010 should have become the document names.

      immig_corpus <- corpus(data_char_ukimmig2010)
      Corpus consisting of 9 documents:
               Text Types Tokens Sentences
                BNP  1125   3280        88
          Coalition   142    260         4
       Conservative   251    499        15
             Greens   322    679        21
             Labour   298    683        29
             LibDem   251    483        14
                 PC    77    114         5
                SNP    88    134         4
               UKIP   346    723        27
      Source:  /home/travis/build/UCLSPP/PUBLG088/_book/data/* on x86_64 by travis
      Created: Wed Dec 20 23:40:53 2017
    2. From a directory of text files.

      The readtext() function from the readtext package can read (almost) any set of files into an object that you can then call the corpus() function on, to create a corpus. (See ?readtext for an example.)

      Here you are encouraged to select any directory of plain text files of your own.
      How did it work? Try using docvars() to assign a set of document-level variables. If you do not have a set of text files to work with, then you can use the UK 2010 manifesto texts on immigration, in the Day 8 folder, like this:

      manfiles <- readtext("https://github.com/kbenoit/ME114/raw/master/day8/UKimmigTexts.zip")
      mycorpus <- corpus(manfiles)
    3. From .csv or .json files — see the documentation for the package readtext (help(package = "readtext")).

      Here you can try one of your own examples, or just file this in your mental catalogue for future reference.

  4. Explore some phrases in the text.

    You can do this using the kwic (for “key-words-in-context”) to explore a specific word or phrase.

    kwic(data_corpus_inaugural, "terror", 3)
        [1797-Adams, 1325]             violence, by | terror |
     [1933-Roosevelt, 112] unreasoning, unjustified | terror |
     [1941-Roosevelt, 287]          by a fatalistic | terror |
       [1961-Kennedy, 866]     uncertain balance of | terror |
        [1981-Reagan, 813]       Americans from the | terror |
      [1997-Clinton, 1055]        the fanaticism of | terror |
      [1997-Clinton, 1655]   strong defense against | terror |
        [2009-Obama, 1632]         aims by inducing | terror |
     , intrigue,               
     which paralyzes needed    
     , we proved               
     that stays the            
     of runaway living         
     . And they                
     and destruction.          
     and slaughtering innocents

    Try substituting your own search terms, or working with your own corpus.

    head(kwic(data_corpus_inaugural, "america", 3))
     [1793-Washington, 63]      people of united | America |
          [1797-Adams, 16]     middle course for | America |
         [1797-Adams, 427]         the people of | America |
        [1797-Adams, 1419]         the people of | America |
        [1797-Adams, 2004] aboriginal nations of | America |
        [1797-Adams, 2152]         the people of | America |
     . Previous to             
     remained between unlimited
     were not abandoned        
     have exhibited to         
     , and a                   
     and the internal          
    head(kwic(data_corpus_inaugural, "democracy", 3))
        [1825-Adams, 1546] a confederated representative | democracy |
      [1841-Harrison, 525]                    to that of | democracy |
     [1841-Harrison, 1585]       a simple representative | democracy |
     [1841-Harrison, 7463]                   the name of | democracy |
     [1841-Harrison, 7894]                of devotion to | democracy |
      [1921-Harding, 1087]      temple of representative | democracy |
     were a government
     . If such        
     or republic,     
     they speak,      
     . The foregoing  
     , to be          
  5. Create a document-feature matrix, using dfm. First, read the documentation using ?dfm to see the available options.

    mydfm <- dfm(data_corpus_inaugural, remove = stopwords("english"))
    Document-feature matrix of: 58 documents, 9,221 features (92.6% sparse).
    topfeatures(mydfm, 20)
             ,          .          -     people          ; government 
          7026       4945        762        575        565        564 
            us        can       upon       must      great        may 
           478        471        371        366        340        338 
        states      shall      world    country      every     nation 
           333        314        311        304        298        293 
         peace        one 
           254        252 

    Experiment with different dfm options, such as stem = TRUE. The function dfm_trim() allows you to reduce the size of the dfm following its construction.

    dim(dfm(data_corpus_inaugural, stem = TRUE))
    [1]   58 5541
    dim(dfm_trim(mydfm, min_count = 5, min_docfreq = 0.01))
    [1]   58 2596

    Grouping on a variable is an excellent feature of dfm(), in fact one of my favorites.
    For instance, if you want to aggregate all speeches by presidential name, you can execute

    mydfm <- dfm(data_corpus_inaugural, groups = "President")
    Document-feature matrix of: 35 documents, 9,357 features (88.3% sparse).
     [1] "Washington" "Adams"      "Jefferson"  "Madison"    "Monroe"    
     [6] "Jackson"    "Van Buren"  "Harrison"   "Polk"       "Taylor"    
    [11] "Pierce"     "Buchanan"   "Lincoln"    "Grant"      "Hayes"     
    [16] "Garfield"   "Cleveland"  "McKinley"   "Roosevelt"  "Taft"      
    [21] "Wilson"     "Harding"    "Coolidge"   "Hoover"     "Truman"    
    [26] "Eisenhower" "Kennedy"    "Johnson"    "Nixon"      "Carter"    
    [31] "Reagan"     "Bush"       "Clinton"    "Obama"      "Trump"     

    Note that this groups Theodore and Franklin D. Roosevelt together – to separate them we would have needed to add a firstname variable using docvars() and grouped on that as well.

    Do this to aggregate the Irish budget corpus (data_corpus_irishbudget2010) by political party, when creating a dfm.

    mydfm <- dfm(data_corpus_inaugural, remove = stopwords("english"), remove_punct = TRUE, stem = TRUE)
    topfeatures(mydfm, 20)
       nation    govern     peopl        us       can     state     great 
          675       657       623       478       471       450       373 
         upon     power      must   countri     world       may     shall 
          371       370       366       355       339       338       314 
        everi constitut      peac     right       law      time 
          298       286       283       276       271       267 
    irish_dfm <- dfm(data_corpus_irishbudget2010, groups = "party")
  6. Explore the ability to subset a corpus.

    There is a corpus_subset() method defined for a corpus, which works just like R’s normal subset() command. For instance if you want a wordcloud of just Obama’s two inagural addresses, you would need to subset the corpus first:

    obamadfm <- dfm(corpus_subset(data_corpus_inaugural, President=="Obama"))

    Try producing that plot without the stopwords. See dfm_remove() to remove stopwords from the dfm object directly, or supply the remove argument to dfm().

    obamadfm <- dfm(corpus_subset(data_corpus_inaugural, President=="Obama"), remove = stopwords("SMART"), remove_punct = TRUE)
    Warning: 'stopwords(language = "SMART")' is deprecated.
    Use 'stopwords(source = "smart")' instead.
    See help("Deprecated")

  7. Preparing and pre-processing texts

    1. “Cleaning”" texts

      It is common to “clean” texts before processing, usually by removing punctuation, digits, and converting to lower case. Look at the documentation for char_tolower() and use the command on the data_char_sampletext text (you can load this from quantedaData using data(data_char_sampletext). Can you think of cases where cleaning could introduce homonymy?

      sample_text_lowercase <- char_tolower(data_char_sampletext)
    2. Tokenizing texts

      In order to count word frequencies, we first need to split the text into words through a process known as tokenization. Look at the documentation for quanteda’s tokens() function. Use the tokens command on data_char_sampletext, and examine the results. Are there cases where it is unclear where the boundary between two words lies? You can experiment with the options to tokens.

      Try tokenizing the sentences from data_char_sampletext into sentences, using tokens(x, what = "sentence").

      tokens(sample_text_lowercase, remove_hyphens = TRUE)
      tokens(sample_text_lowercase, remove_numbers = TRUE, remove_punct = TRUE, remove_symbols = TRUE)
      tokens(sample_text_lowercase, what = "sentence")
      tokens from 1 document.
      text1 :
      [1] "instead we have a fine gael-labour party government, coming into power promising real change but slavishly following the previous government's policy. that policy has been dictated by the imf, the eu and the ecb not to bail out the irish people but to salvage german, french, british banks and those of other countries from their disastrous and frenzied embrace of irish bankers and speculators in the irish property market bubble. it is criminal that an irish government would ever slavishly agree to make a vassal state of the republic of ireland and its people, to squeeze tribute from our people to save the capitalist banks of europe and in doing so destroy the lives of hundreds of thousands of our people now plunged into unemployment, financial hardship and social dislocation. in this budget and the past four years the austerity policy means €25 billion has been reefed out of the irish economy in pursuit of a policy of cringing acceptance of the diktats of the financial markets. can this government not see that, not only is this immoral and unjust in the extreme, it is decimating the domestic economy?"                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         
      [2] "as we are tired of pointing out, if we savage the ability of the majority of our people to purchase goods and utilise services, then tens of thousands of workers depending on this demand for their jobs will be thrown on the scrapheap of unemployment and, tragically, that is what is happening. all the key indicators in the domestic economy show the abject failure of austerity. private investment has collapsed, vat receipts are more than €400 million behind and unemployment has risen by thousands since this government entered office. much is made of the growth in exports. every job in the export sector is vital and we defend it but the type of investment and the capital intensive nature of the investment that goes into exports means that it is not where the hundreds of thousands of jobs we need will be created in the next period of years. on the other hand, taking €670 million from the pockets of ordinary people through the vat increases and taking other money in the cuts in child benefit and elsewhere will further add to the downward spiral of austerity. the interest relief for householders who are trapped in the nightmare of negative equity and extortionate monthly mortgage payments will be welcomed to a degree but is dismally inadequate. that generation of workers who are trapped in this nightmare are victims of the extortion perpetrated on them in the housing market by the immoral speculators legislated for by fianna fáil and the pds. the huge proportion of their incomes that continues to go to the banks massively dislocates the economy. otherwise those funds would be going to the purchase of goods and services in the domestic economy, stimulating demand and sustaining tens of thousands of jobs for tens of thousands of people who are now on the dole, unfortunately."
    3. Stemming.

      Stemming removes the suffixes using the Porter stemmer, found in the SnowballC library. The quanteda function* to invoke the stemmer end with *_wordstem. Apply stemming to the data_char_sampletext (using char_wordstem()) and examine the results. Why does it not appear to work, and what do you need to do to make it work? How would you apply this to the sentence-segmented vector?

      # Word-segmented vector
      sample_word_stems <- char_wordstem(tokens(data_char_sampletext)[[1]])
       [1] "Instead"     "we"          "have"        "a"           "Fine"       
       [6] "Gael-Labour" "Parti"       "Govern"      ","           "come"       
      # Sentence-segmented vector
      sentence_boundaries <- c(0, which(sample_word_stems == "."))
      sample_sentence_word_stems <- list()
      for(i in 1:(length(sentence_boundaries)-1)) {
      sample_sentence_word_stems[[i]] <- paste0(sample_word_stems[(sentence_boundaries[i] + 1):(sentence_boundaries[i+1])], collapse = " ")
      [1] "Instead we have a Fine Gael-Labour Parti Govern , come into power promis real chang but slavish follow the previous Govern polici ."
      [1] "That polici has been dictat by the IMF , the EU and the ECB not to bail out the Irish peopl but to salvag German , French , British bank and those of other countri from their disastr and frenzi embrac of Irish banker and specul in the Irish properti market bubbl ."
      [1] "It is crimin that an Irish Govern would ever slavish agre to make a vassal State of the Republ of Ireland and it peopl , to squeez tribut from our peopl to save the capitalist bank of Europ and in do so destroy the live of hundr of thousand of our peopl now plung into unemploy , financi hardship and social disloc ."
    4. Applying “pre-processing” to the creation of a dfm.

      quanteda’s dfm() function makes it wasy to pass the cleaning arguments to clean, which are executed as part of the tokenization implemented by dfm(). Compare the steps required in a similar text preparation package, tm:

      crude <- tm_map(crude, content_transformer(tolower))
      crude <- tm_map(crude, removePunctuation)
      crude <- tm_map(crude, removeNumbers)
      crude <- tm_map(crude, stemDocument)
      tdm <- TermDocumentMatrix(crude)
      # same in quanteda
      crudeCorpus <- corpus(crude)
      crudeDfm <- dfm(crudeCorpus)

      Inspect the dimensions of the resulting objects, including the names of the words extracted as features. It is also worth comparing the structure of the document-feature matrixes returned by each package. tm uses the slam simple triplet matrix format for representing a sparse matrix.

      It is also – in fact almost always – useful to inspect the structure of this object:

      List of 6
       $ i       : int [1:1877] 42 77 100 136 148 158 163 164 174 180 ...
       $ j       : int [1:1877] 1 1 1 1 1 1 1 1 1 1 ...
       $ v       : num [1:1877] 1 2 1 1 2 2 1 1 2 2 ...
       $ nrow    : int 848
       $ ncol    : int 20
       $ dimnames:List of 2
        ..$ Terms: chr [1:848] "abdulaziz" "abil" "abl" "about" ...
        ..$ Docs : chr [1:20] "127" "144" "191" "194" ...
       - attr(*, "class")= chr [1:2] "TermDocumentMatrix" "simple_triplet_matrix"
       - attr(*, "weighting")= chr [1:2] "term frequency" "tf"

      THis indicates that we can extract the names of the words from the tm TermDocumentMatrix object by getting the rownames from inspecting the tdm:

      head(tdm$dimnames$Terms, 20)
       [1] "abdulaziz" "abil"      "abl"       "about"     "abov"     
       [6] "abroad"    "accept"    "accord"    "across"    "activ"    
      [11] "add"       "address"   "adher"     "advantag"  "advis"    
      [16] "after"     "again"     "against"   "agenc"     "agre"     

      Compare this to the results of the same operations from quanteda. To get the “words” from a quanteda object, you can use the featnames() function:

      features_quanteda <- featnames(crudeDfm)
      head(features_quanteda, 20)
       [1] "diamond"  "shamrock" "corp"     "said"     "that"     "effect"  
       [7] "today"    "it"       "had"      "cut"      "contract" "price"   
      [13] "for"      "crude"    "oil"      "by"       "dlrs"     "a"       
      [19] "barrel"   "the"     
      Formal class 'dfm' [package "quanteda"] with 15 slots
        ..@ settings    :List of 1
        .. ..$ : NULL
        ..@ weightTf    :List of 3
        .. ..$ scheme: chr "count"
        .. ..$ base  : NULL
        .. ..$ K     : NULL
        ..@ weightDf    :List of 5
        .. ..$ scheme   : chr "unary"
        .. ..$ base     : NULL
        .. ..$ c        : NULL
        .. ..$ smoothing: NULL
        .. ..$ threshold: NULL
        ..@ smooth      : num 0
        ..@ ngrams      : int 1
        ..@ skip        : int 0
        ..@ concatenator: chr "_"
        ..@ version     : int [1:3] 0 99 9031
        ..@ docvars     :'data.frame':    20 obs. of  15 variables:
        .. ..$ datetimestamp: chr [1:20] "1987-02-26 17:00:56" "1987-02-26 17:34:11" "1987-02-26 18:18:00" "1987-02-26 18:21:01" ...
        .. ..$ description  : chr [1:20] "" "" "" "" ...
        .. ..$ id           : chr [1:20] "127" "144" "191" "194" ...
        .. ..$ language     : chr [1:20] "en" "en" "en" "en" ...
        .. ..$ origin       : chr [1:20] "Reuters-21578 XML" "Reuters-21578 XML" "Reuters-21578 XML" "Reuters-21578 XML" ...
        .. ..$ topics       : chr [1:20] "YES" "YES" "YES" "YES" ...
        .. ..$ lewissplit   : chr [1:20] "TRAIN" "TRAIN" "TRAIN" "TRAIN" ...
        .. ..$ cgisplit     : chr [1:20] "TRAINING-SET" "TRAINING-SET" "TRAINING-SET" "TRAINING-SET" ...
        .. ..$ oldid        : chr [1:20] "5670" "5687" "5734" "5737" ...
        .. ..$ places       : chr [1:20] "usa" "usa" "canada" "usa" ...
        .. ..$ author       : chr [1:20] NA "BY TED D'AFFLISIO, Reuters" NA NA ...
        .. ..$ orgs         : chr [1:20] NA "opec" NA NA ...
        .. ..$ people       : chr [1:20] NA NA NA NA ...
        .. ..$ exchanges    : chr [1:20] NA NA NA NA ...
        ..@ i           : int [1:2086] 0 0 0 10 17 0 1 2 3 4 ...
        ..@ p           : int [1:876] 0 1 2 5 25 36 41 48 65 69 ...
        ..@ Dim         : int [1:2] 20 875
        ..@ Dimnames    :List of 2
        .. ..$ docs    : chr [1:20] "127" "144" "191" "194" ...
        .. ..$ features: chr [1:875] "diamond" "shamrock" "corp" "said" ...
        ..@ x           : num [1:2086] 2 1 1 3 1 3 11 1 1 3 ...
        ..@ factors     : list()

      What proportion of the crudeDfm are zeros? Compare the sizes of tdm and crudeDfm using the object.size() function.

       FALSE   TRUE 
      0.1192 0.8808 
      print(object.size(crudeDfm), units= "Mb")
      0.1 Mb
      print(object.size(tdm), units= "Mb")
      0.1 Mb
  8. Keywords-in-context

    1. quanteda provides a keyword-in-context function that is easily usable and configurable to explore texts in a descriptive way. Type ?kwic to view the documentation.

    2. For the Irish budget debate speeches corpus for the year 2010, called data_corpus_irishbudget2010, experiment with the kwic function, following the syntax specified on the help page for kwic. kwic can be used either on a character vector or a corpus object. What class of object is returned? Try assigning the return value from kwic to a new object and then examine the object by clicking on it in the environment pane in RStudio (or using the inspection method of your choice).

      northern_kwic <- kwic(data_corpus_irishbudget2010, "Northern")
      class(northern_kwic) # Class  = data.frame or kwic object
      [1] "kwic"       "data.frame"
    3. Use the kwic function to discover the context of the word “clean”. Is this associated with environmental policy?

      clean_kwic <- kwic(data_corpus_irishbudget2010, "clean") # No, more like corruption!
         [2010_BUDGET_03_Joan_Burton_LAB, 5266]          pay the cost of the |
            [2010_BUDGET_06_Enda_Kenny_FG, 169]      recovery, a strategy to |
           [2010_BUDGET_06_Enda_Kenny_FG, 1240]       , while the people who |
       [2010_BUDGET_08_Eamon_Gilmore_LAB, 3551] strengths in sectors such as |
       clean | up and restore the gang     
       clean | up the banks and a          
       clean | out the Taoiseach's offices-
       clean | technology, food and the    
    4. By default, kwic explores all words related to the word, since it interprets the pattern as a “regular expression”. What if we wanted to see only the literal, entire word “disaster”? Hint: Look at the arguments using ?kwic.

      disaster_kwic <- kwic(data_corpus_irishbudget2010, "disaster", valuetype = "fixed")
  9. Descriptive statistics

    1. We can extract basic descriptive statistics from a corpus from its document feature matrix. Make a dfm from the 2010 Irish budget speeches corpus.

      irish_dfm <- dfm(data_corpus_irishbudget2010, remove_punct = TRUE)
    2. Examine the most frequent word features using textstat_frequency(). What are the five most frequent word in the corpus? (Note: There is a also a topfeatures() command that works in a similar way.)

      head(textstat_frequency(irish_dfm), 5)
        feature frequency rank docfreq
      1     the      3598    1      14
      2      to      1633    2      14
      3      of      1537    3      14
      4     and      1359    4      14
      5      in      1231    5      14
       the   to   of  and   in 
      3598 1633 1537 1359 1231 
    3. quanteda provides a function to count syllables in a word — nsyllable(). Try the function at the prompt. The code below will apply this function to all the words in the corpus, to give you a count of the total syllables in the corpus.

      # count syllables from texts in the 2010 speech corpus 
      textsyls <- nsyllable(texts(data_corpus_irishbudget2010))
      # sum the syllable counts 
      [1] 80284

      How would you get the total syllables per text?

       [1] 13370  6313  9126 10569 10260  6172  3155  6081  1869  1842  1517
      [12]  2293  1793  5924