ex2.md

Slow Bernese?

Benjamin Gföhler (December 17, 2019)

As Switzerland is a federalistic country, Cantons (States) are a rather important feature. Accordingly, there is quite some competition between Cantons in all aspects of life. One very controversial point are languages. There are not just 4 different official languages, but also huge differences within the part that speaks the same language. One prominent common belief is that people from the Canton of Berne (the Capital) are extraordinary slow speakers - especially compared to people from Zurich (yes, I’m from Zurich). To test this common belief, we will use speeches from members of the Swiss Parliament, reaching back to 1999.

First we have to download the data, which takes quite some time…

df <- get_data(table = "Transcript", Language = "DE")

Exploring the Data

To get familiar with the data, we first check who spoke the most since 1999.

library(dplyr)

df %>%
  group_by(PersonNumber, SpeakerLastName, CouncilName) %>% 
  count() %>%
  ungroup() %>% 
  top_n(20, n) %>% 
  arrange(desc(n)) %>% 
  knitr::kable()

PersonNumber	SpeakerLastName	CouncilName	n
NA	NA	NA	43961
488	Leuthard	Bundesrat	3306
512	Sommaruga	Bundesrat	2796
3934	Widmer-Schlumpf	Bundesrat	2496
134	Leuenberger	Bundesrat	2417
277	Bieri	Ständerat	2274
370	Forster-Vannini	Ständerat	2256
50	Couchepin	Bundesrat	2050
372	Inderkum	Ständerat	2025
825	Germann	Ständerat	2022
540	Lombardi	Ständerat	1991
37	Büttiker	Ständerat	1940
1161	Berset	Bundesrat	1831
276	Brändli	Ständerat	1811
508	Schneider-Ammann	Bundesrat	1811
1124	de Buman	Nationalrat	1805
146	Maurer	Bundesrat	1776
398	Merz	Bundesrat	1770
3916	Hêche	Ständerat	1753
85	Frick	Ständerat	1735

This is still raw data and not very helpful for our case. First, lets drop all texts that are not speeches (Type1) and exclude all texts by presidents of the council, as they speak a lot more than the rest and on special occasions, which biases the analysis.

dfs <- df %>%
  filter(Type == 1) %>% 
  filter(!SpeakerLastName == "leer") %>% 
  filter(!SpeakerFunction %in% c("1VP-F", "1VP-M", "2VP-F", "2VP-M", "AP-M", "P-F", "P-M")) %>% 
  filter(!Function %in% c("1VP-M", "2VP-M", "P-F", "p-m", "P-m", "P-M", "P-MM"))

Looks a lot better now…

dfs %>%
  group_by(PersonNumber, SpeakerLastName, CouncilName) %>% 
  count() %>%
  ungroup() %>% 
  top_n(20, n) %>% 
  arrange(desc(n)) %>% 
  knitr::kable()

PersonNumber	SpeakerLastName	CouncilName	n
488	Leuthard	Bundesrat	3306
512	Sommaruga	Bundesrat	2796
3934	Widmer-Schlumpf	Bundesrat	2496
134	Leuenberger	Bundesrat	2417
50	Couchepin	Bundesrat	2050
1161	Berset	Bundesrat	1831
508	Schneider-Ammann	Bundesrat	1811
146	Maurer	Bundesrat	1776
398	Merz	Bundesrat	1770
487	Leutenegger Oberholzer	Nationalrat	1438
21	Blocher	Bundesrat	1198
466	Fetz	Ständerat	1116
1116	Burkhalter	Bundesrat	1077
57	David	Ständerat	942
58	Deiss	Bundesrat	929
429	Villiger	Bundesrat	924
379	Wicki	Ständerat	902
277	Bieri	Ständerat	815
1122	Fluri	Nationalrat	813
825	Germann	Ständerat	807

The magic is about to happen. After cleaning the text and free it from unnecessary remarks, we calculate the duration of the speech, count the number of words and divide the first by the latter to get speech rate words per minute. Lastly, we filter all entries that procude NAs in words per minute.

library(tokenizers)

dfs <- dfs %>%
  mutate(
    Text2 = swissparl::clean_text(Text, keep_round_brackets = F),
    SpeechDurationMin = as.numeric(difftime(End, Start, units = "mins")),
    NumOfWords = count_words(Text2),
    WordsPerMin = NumOfWords / SpeechDurationMin
    ) %>%
  filter(!is.na(WordsPerMin))

Let’s check the distribution of Number of Words and Speech Duration to see whether our calculations worked.

# Packages
library(ggplot2)
library(hrbrthemes)

# Scatterplot
dfs %>% 
  ggplot(aes(NumOfWords, SpeechDurationMin)) +
  geom_point() +
  theme_ipsum_rc()

Seems like there are some unrealistic values. Let’s have a closer look at the fastest and slowest speaker and their speeches.

# Fastest
dfs %>% 
  arrange(desc(WordsPerMin)) %>% 
  select(IdSubject, NumOfWords, SpeechDurationMin, WordsPerMin) %>% 
  slice(1:20) %>% 
  knitr::kable()

IdSubject	NumOfWords	SpeechDurationMin	WordsPerMin
15063	484	0.0182667	26496.351
21019	419	0.0330667	12671.371
21019	426	0.0565000	7539.823
46255	638	0.1692667	3769.201
2405	776	0.2359500	3288.832
608	108	0.0338833	3187.408
21019	77	0.0294500	2614.601
21019	64	0.0250000	2560.000
25254	224	0.0912167	2455.692
18048	215	0.1023333	2100.977
4914	237	0.1223833	1936.538
6935	96	0.0500000	1920.000
22738	189	0.0987667	1913.601
21019	58	0.0333333	1740.000
6505	562	0.3388333	1658.633
34607	815	0.5068333	1608.024
217	399	0.2551667	1563.684
1491	385	0.2557167	1505.573
22762	261	0.1767833	1476.384
21019	59	0.0419333	1406.995

# Slowest
dfs %>% 
  arrange(WordsPerMin) %>% 
  select(IdSubject, NumOfWords, SpeechDurationMin, WordsPerMin) %>% 
  slice(1:20) %>% 
  knitr::kable()

IdSubject	NumOfWords	SpeechDurationMin	WordsPerMin
213	53	-31.3060000	-1.6929662
46124	12	9687.8827333	0.0012387
46915	6	4301.4335500	0.0013949
38661	3	1373.3121667	0.0021845
36014	4	1018.2860500	0.0039282
40152	12	1319.8989333	0.0090916
46124	7	408.3947167	0.0171403
37202	3	61.9845000	0.0483992
46125	2	40.9251000	0.0488698
17163	4	28.7405500	0.1391762
43274	1	3.5910500	0.2784701
37201	1	3.4094500	0.2933024
3692	1	2.5661667	0.3896863
43197	1	1.8215667	0.5489780
1491	3	3.9611667	0.7573526
16776	47	52.1074000	0.9019832
2665	3	3.2271167	0.9296224
17657	33	30.0277833	1.0989822
42351	1	0.7871667	1.2703790
1115	4	3.0943333	1.2926856

Seems like keeping track of the speaking time was the pièce de résistance. What irony! In a country famous for its high end, reliable watches…

Filtering outliers based on Interquartile Range

MinSpeed <- quantile(dfs$WordsPerMin, 0.25) - 1.5 * IQR(dfs$WordsPerMin)
MaxSpeed <- quantile(dfs$WordsPerMin, 0.75) + 1.5 * IQR(dfs$WordsPerMin)
 
dfs2 <- dfs %>% 
  filter(WordsPerMin > MinSpeed) %>% 
  filter(WordsPerMin < MaxSpeed)

Re-examination of the distribution:

dfs2 %>% 
  filter(MeetingCouncilAbbreviation == "N") %>% 
  ggplot(aes(NumOfWords, SpeechDurationMin)) +
  geom_point() +
  theme_ipsum_rc()

Looks a lot better now…

Language-Specific Analysis

Let’s have a look at the distribution of speeches over different languages:

# Speechs per Language
dfs2 %>% 
  group_by(LanguageOfText) %>% 
  count() %>% 
  mutate(Share = n/nrow(dfs2)) %>% 
  knitr::kable()

LanguageOfText	n	Share
DE	110719	0.7558333
FR	34430	0.2350395
IT	1176	0.0080281
NA	161	0.0010991

The use of italian and french mirrors their respective share of the Swiss population almost perfectly!

Next step: delete NAs and look at the speech rate per language.

library(ggridges)

dfs2 %>% 
  filter(!is.na(LanguageOfText)) %>% 
  ggplot(aes(x = WordsPerMin, y = LanguageOfText)) +
  geom_density_ridges(alpha = 0.8) +
  theme_ipsum_rc()

As expected, slow german-speaking members, fast speaking representatives from the french part - and please note this nice normal distribution for german speeches!

Speed on Individual Level

Now, let’s have a look at speeds on the individual level.

In a first step, we calculate the average speed per person. To create greater reliability, we exclude individuals with less than 10 speeches in a given language. To find the fastet and slowest speaking individuals we look at their deviation from the mean.

# Speed per councillor
dfs3 <- dfs2 %>%
  
  # Mean per language
  filter(!is.na(LanguageOfText)) %>% 
  group_by(LanguageOfText) %>% 
  mutate(MeanLanguage = mean(WordsPerMin)) %>% 
  
  # Minimal number of speeches per language
  group_by(PersonNumber, SpeakerLastName, CantonAbbreviation, LanguageOfText) %>%
  mutate(NumberOfSpeeches = n()) %>% 
  ungroup() %>% 
  filter(NumberOfSpeeches > 10) %>% 
  
  # Deviation from mean
  group_by(PersonNumber, SpeakerLastName, CantonAbbreviation, LanguageOfText) %>%
  summarise(
    MeanWordsPerMin = mean(WordsPerMin),
    DevFromLanguageMean = MeanWordsPerMin - mean(MeanLanguage)
    )
  
# Fastest vs. slowest
bind_rows(
  dfs3 %>% 
    group_by(LanguageOfText) %>% 
    top_n(5, MeanWordsPerMin),
  dfs3 %>% 
    group_by(LanguageOfText) %>% 
    top_n(-5, MeanWordsPerMin)
  ) %>% 
  mutate(label = paste0(SpeakerLastName, " (", CantonAbbreviation, ")")) %>% 
  ggplot(aes(reorder(label, DevFromLanguageMean), DevFromLanguageMean)) +
  geom_col() +
  coord_flip() +
  scale_y_continuous(limits = c(-40, 40)) +
  facet_wrap(.~LanguageOfText, scales = "free_y") +
  labs(
    title = "Fastest and Slowest Speakers per Language" ,
    x = "",
    y = "Deviation from Average Speed (Words/Min)"
    ) +
  theme_ipsum_rc() +
  theme(panel.grid.minor = element_blank())

Cantonal Differences

That’s the fun part, how slow are the bernese actually?

German

dfs2 %>%
  filter(LanguageOfText == "DE") %>% 
  filter(!is.na(CantonAbbreviation)) %>%
  group_by(CantonAbbreviation) %>% 
  filter(n() > 20) %>% 
  mutate(CantonMedian = median(WordsPerMin)) %>% 
  ungroup() %>% 
  ggplot(aes(x = WordsPerMin, y = reorder(CantonAbbreviation, CantonMedian))) +
  stat_density_ridges(
    quantile_lines = TRUE, 
    quantiles = 2, 
    fill = hrbrthemes::ft_cols$peach, 
    alpha = 0.8
    ) +
  labs(
    title = "Mean Speech Rate of Councillors per Canton (German speeches)" ,
    x = "Words/Min"
    ) +
  theme_ipsum_rc() +
  theme(axis.title.y = element_blank())

Bummer! The Bernese representative are about average and even faster speaking than their colleagues from Zurich. However, in parliament all representatives speak in standard german and not in their respective dialects, which probably distorts the values somewhat in favour of the bernese.

French

dfs2 %>%
  filter(LanguageOfText == "FR") %>% 
  filter(!is.na(CantonAbbreviation)) %>%
  group_by(CantonAbbreviation) %>%
  filter(n() > 20) %>% 
  mutate(CantonMedian = median(WordsPerMin)) %>% 
  ungroup() %>% 
  ggplot(aes(x = WordsPerMin, y = reorder(CantonAbbreviation, CantonMedian))) +
  stat_density_ridges(
    quantile_lines = TRUE, 
    quantiles = 2, 
    fill = "dodgerblue3", 
    alpha = 0.8
    ) +
  labs(
    title = "Mean Speech Rate of Councillors per Canton (French speeches)" ,
    x = "Words/Min"
    ) +
  theme_ipsum_rc() +
  theme(axis.title.y = element_blank())

Too bad, even in french the bernese councillors are faster than their counterparts from Zurich.

Italian

dfs2 %>%
  filter(LanguageOfText == "IT") %>% 
  filter(!is.na(CantonAbbreviation)) %>%
  group_by(CantonAbbreviation) %>%
  filter(n() > 20) %>% 
  mutate(CantonMedian = median(WordsPerMin)) %>% 
  ungroup() %>% 
  ggplot(aes(x = WordsPerMin, y = reorder(CantonAbbreviation, CantonMedian))) +
  stat_density_ridges(
    quantile_lines = TRUE, 
    quantiles = 2, 
    fill = "green", 
    alpha = 0.8
    ) +
  labs(
    title = "Mean Speech Rate of Councillors per Canton (Italian speeches)" ,
    x = "Words/Min"
    ) +
  theme_ipsum_rc() +
  theme(axis.title.y = element_blank())

And finally, they are also using Italian and Zurich’s representatives not - show-offs! ;)