Qualitative visualisation

Day 2 · 14:00

Complaints, exit interviews and app-store reviews are data. They are just not rectangular yet. Tokenise them, remove the theatre, and you can put voice-of-customer on the same page as NPS.

The original workshop used a corpus about a city. You will use 180 retail-bank complaint tickets — the sort of extract a CX team emails on a Friday. The temptation is a word cloud. Word clouds are decoration. EXCO needs ranked content words, split by channel, with the confidential tokens already gone.

One token per row

Julia Silge and David Robinson’s tidy text format is the same contract as yesterday: each observation (here, a word) is a row. unnest_tokens() splits a comment, drops punctuation, and lowercases in one step.

library(tidyverse)
library(tidytext)
 
voc <- read_csv("voc-complaints.csv")
 
tokens <- voc %>%
  unnest_tokens(word, comment) %>%
  count(word, sort = TRUE)
 
head(tokens, 10)

The top of that list will be *the, a, to, my*. That is not insight. It is English.

Stop-words — and your own

data("stop_words")
 
confidential <- tibble(word = c(
  "northridge", "bvn", "account", "nuban"
))
 
numbers <- tokens %>%
  filter(str_detect(word, "^[0-9]+$"))
 
clean <- tokens %>%
  anti_join(stop_words, by = "word") %>%
  anti_join(confidential, by = "word") %>%
  anti_join(numbers, by = "word") %>%
  filter(nchar(word) > 2)
 
clean %>% slice_head(n = 10)

After cleaning, the list starts to look like an operating agenda: app, branch, transfer, charges, agent, queue. ‘Resolved’ appearing at all is the only unambiguously positive token in the top ten — and it is a process word, not a brand word.

Split by channel, or you will mislead

voc %>%
  unnest_tokens(word, comment) %>%
  anti_join(stop_words, by = "word") %>%
  filter(nchar(word) > 2) %>%
  count(channel, word, sort = TRUE) %>%
  group_by(channel) %>%
  slice_head(n = 5)

A national cloud would have mixed ‘queue’ (branch) with ‘ussd’ (feature phone) and ‘app’ (mobile). Those are three different owners in the operating model. The chart has to respect the org chart.