Co-occurrences with keyword lists

Playing R with Rammstein texts can be fun when the outcomes are unexpected, and you get plots like

Gut, better. We might even see:

How do we manage making "das Gute" appear?

From general to keyword plots

In the plot based on all the co-occurrences in Rammstein song texts, we can identify a center on the left-hand side:

The meaningful central terms seem to be

"Lust", "Deutschland", "Liebe", "ich", "du", "kalt" and "gut". Choosing these words as keys, we can plot a new picture.

What do we see? There is a clear "ich" - "du" axis. "Liebe" appears together with "ich" and "du", well. But "Lust" seems to be mainly referred only to "ich", as well as "kalt". 

Into the contexts

Finding "kalt" surprising, even at the border of a net with "ich", "du". "Liebe", "Lust", I searched for the occurrences of this adjective. While the noun, "Kälte" does not appear at all, a KWIC-search gives 32 occurences of "kalt" instead. We can transform the immediate context (window= 10) into a matrix, count and plot the co-occurences. 

The connection between "ich" and "kalt" looks interesting and might be opposed to "ihr" and "heiß", but we are talking about very low absolute frequencies. 

       feature frequency rank docfreq group

1         kalt       122    1        19   all

2          ich        47    2         25   all

3         heiß       20    3           5   all

4           du        18    4           9   all

5          ihr         18    4          7   all

6         seid        18    4          5   all

7  deutschland  16    7          4   all

8        amour     12    8          3   all

9         ende      10    9          7   all

10        lust       10    9         10   all

Anyway, it might be interesting drawing similar context plots for "ich", "du", "Liebe" and "Lust". 

The context of the "personal pronoun" (in fact it is a deictic, an indicating, term (Coseriu)) "ich" seemingly does not reveal surprises. "ich", "du", "Lust", "Liebe" again.

But where does the adjective "gut" come from? 

Focussing on words of interest, the importance of "gut" becomes clearer:

Partly, this may be caused by the repetitions in the famous song "Du riechst so gut". But the link between "ich" and "gut" seems to be stronger. Indeed, the 43 appearances of "gut" are distributed among 17 different songs. Without "riechst" (31 occurrences!), within the context of "gut", we get: 

Maybe a topic analysis would be helpful here. But the result of a simple and fast approach does not  look really helpful.  With a simple try, the output looks like 

      topic1     topic2    topic3      topic4   

 [1,] "ich"      "tut"     "ich"       "ich"    

 [2,] "du"       "ich"     "du"        "puppe"  

 [3,] "ihr"      "hört"    "gut"       "ab"     

 [4,] "kind"     "schreit" "riechst"   "kopf"   

 [5,] "schonste" "tu"      "steig"     "ja"     

 [6,] "blut"     "weh"     "geh"       "reiß"   

 [7,] "giftig"   "leid"    "hinterher" "geht"   

 [8,] "dass"     "himmel"  "finde"     "beiß"   

 [9,] "seh"      "gesicht" "gleich"    "hals"   

[10,] "ach"      "ende"    "warte"     "dam-dam"

Topic analysis should be done more extensively. 

Reading words like "ach", "schrei", "giftig", "Leid": We might consider a sentiment analysis. The triangle "ich", "du" and "gut" probably does not give the definitive analytical insight.

For now, a context search for "Lust" and "Liebe" has to be enough. 

The complete Lust plot

But, talking about contexts, "Lust" can be eliminated. Only with the keywords:

In (very low) numbers: 

Feature co-occurrence matrix

                  ich          du          kalt       gut

    ich         33477     259        2590     259

With the keyword "Liebe", again "ich" is central:

If I had lemmatized the tokens, "lieben" and "liebe" would have been one point. 

Without "Lust" and "Liebe":

Nice, is it not? In numbers:

features       du        ihr   ich      gut     kalt

    du           179     352   1179   18      72

    ihr            0        214   804     12      56

    ich            0           0    23710 3973 1697

    gut            0           0     0        225   154

    kalt           0           0     0        0        224

In the field of "Liebe", "ich" is good, not you. 

In the field of "Lust", "ich" is "kalt", while "du" does only rarely appear. Is this Rammstein?

A last glimpse at "du": 207 entries, and still "du hast" is dominating? Only 13 "du" are found in the famous song. 

There is something interesting to be investigated. For now, the "du" context without "du":

In numbers?

features   ich      ihr       gut         kalt     heiß

    ich      17023 2960   17760   6845     10915

In the field of "du", "ich" becomes "gut" and "heiß". Reflect. 

technically

everything is done with R quanteda here, see Benoit et.al.

context search

kontext1 <- kwic(toks_fund, "kalt", case_insensitive = TRUE, valuetype = "glob", window = 10)

context into matrix

dopo1 <- kontext1$post

dopom1 <- quanteda::tokens(dopo1)%>%

          dfm()

pre1 <- kontext1$pre

prem1 <- quanteda::tokens(pre1)%>%

        dfm()

#now bind the two matrices and, because of overlapping #columns, compress

umfeld <- cbind(dopom1, prem1)%>%

    dfm_compress()

relative frequency

clean_rel <- dfm_weight(umfeld, scheme = "prop")

textstat_frequency(clean_rel, n=10)

Topic analysis trial

for (k in 3:17) {

fund_lda <- textmodel_lda(umfeld, k)

print(terms(fund_lda, 10))}