Exercise 10: 🔨 Automatic analysis of text in R

Digital disconnection on Twitter

Open session slides

Download source file

Background

Definition

„Digital disconnection is a deliberate (i.e., chosen by the individual) form of non-use of devices, platforms, features, interactions, and/or messages that occurs with higher or lower frequencies, and for shorter or longer periods of time, after the initial adoption of these technologies, and with the aim of restoring or improving one’s perceived overuse, social interactions, psychological well-being, productivity, privacy and/or perceived usefulness“. (Nassen et al. 2023)

• Increasing trend towards more conscious use of digital media (devices), including (deliberate) non-use with the aim to restore or improve psychological well-being (among other factors)

• But how do “we” talk about digital detox/disconnection: 💊 drug, 👹 demon or 🍩 donut?

Todays’s data basis: Twitter dataset

• Collection of all tweets up to the beginning of 2023 that mention or discuss digital detox (and similar terms) on Twitter (not 𝕏)
• Initial query is searching for “digital detox”, “#digitaldetox”, “digital_detox”
• Access via official Academic-Twitter-API via academictwitteR (Barrie and Ho 2021) at the beginning of last year

Preparation

if (!require("pacman")) install.packages("pacman")
pacman::p_load(
  here, qs, # file management
  magrittr, janitor, # data wrangling
  easystats, sjmisc, # data analysis
  ggpubr, ggwordcloud, # visualization
  gt, gtExtras, # fancy tables
  tidytext, textdata, widyr, # tidy text processing
  quanteda, # quanteda text processing
  quanteda.textplots, 
  topicmodels, stm, 
  tidyverse # load last to avoid masking issues
  )

Import and process the data

Get twitter data

# Import raw data from local
tweets <- qs::qread(here("local_data/tweets-digital_detox.qs"))$raw %>% 
  janitor::clean_names()

# Initial data processing
tweets_correct <- tweets %>% 
  mutate(
    # reformat and create datetime variables
    across(created_at, ~ymd_hms(.)), # convert to dttm format
    year = year(created_at), 
    month = month(created_at), 
    day = day(created_at), 
    hour = hour(created_at),
    minute = minute(created_at),
    # create addtional variables
    retweet_dy = str_detect(text, "^RT"), # identify retweets
    detox_dy = str_detect(text, "#digitaldetox") 
  ) %>% 
  distinct(tweet_id, .keep_all = TRUE)

# Filter relevant tweets
tweets_detox <- tweets_correct %>% 
  filter(
    detox_dy == TRUE, # only tweets with #digitaldetox
    retweet_dy == FALSE, # no retweets
    lang == "en" # only english tweets
    )

DTM/DFM creation

tidytext

# Common HTML entities
remove_reg <- "&amp;|&lt;|&gt;"

# Create tidy data
tweets_tidy <- tweets_detox %>% 
  mutate(
    text = str_remove_all(text, remove_reg)) %>% 
    tidytext::unnest_tokens("text", text) %>% 
    filter(!text %in% tidytext::stop_words$word)

# Create summarized data
tweets_summarized <- tweets_tidy %>% 
  count(tweet_id, text) 

# Create DTM
tweets_dfm <- tweets_summarized %>% 
  cast_dfm(tweet_id, text, n)

# Preview
tweets_dfm

quanteda

# Create corpus
quanteda_corpus <- tweets_detox %>% 
  mutate(across(text, ~str_replace_all(., "#digitaldetox", ""))) %>% 
  select(-c(detox_dy, retweet_dy)) %>% 
  quanteda::corpus(
    docid_field = "tweet_id", 
    text_field = "text"
  )

# Tokenize
quanteda_token <- quanteda_corpus %>% 
  quanteda::tokens(
    remove_punct = TRUE,
    remove_symbols = TRUE, 
    remove_numbers = TRUE, 
    remove_url = TRUE, 
    split_tags = FALSE # keep hashtags and mentions
  ) %>% 
  quanteda::tokens_tolower() %>% 
  quanteda::tokens_remove(
    pattern = stopwords("en")
    )

# Convert to Document-Feature-Matrix (DFM)
quanteda_dfm <- quanteda_token %>% 
  quanteda::dfm()

# Preview
quanteda_dfm

Get topic model (data)

# TPM data
stm_results <- qs::qread(here("local_data/stm_results.qs"))

# Base data with topics 
tweets_detox_topics <- qs::qread(here("local_data/tweets-digital-detox-topics.qs"))

📋 Exercises

Objective of this exercise

• Brief review of the contents of the last session
• Teaching the basic steps for creating and analyzing document feature matrices and stm topic models

Attention

• Before you start working on the exercise, please make sure to render all the chunks of the section Preparation. You can do this by using the “Run all chunks above”-button of the next chunk.
• You can choose to solve some exercises using either the tidytext or quanteda functions. As the work steps differ, please select the tab with your preferred package.
• When in doubt, use the showcase (.qmd or .html) to look at the code chunks used to produce the output of the slides.

📋 Exercise 1: Hashtag co-occurence

Objective(s)

Recreate the network plot from the session without the token #digitaldetox

with tidytext

• Create new dataset tweets_dfm_cleaned
  • Based on the dataset tweets_detox,
    1. Use mutate() and create the variables
       • text that removes the hashtag "#digitaldetox" using str_remove_all() and
       • hashtag that extracts hashtags from the variable text with the help of str_extract_all() (and the pattern "#\\S+").
    2. Tokenize the data by using only unnest() on the variable hashtag.
    3. Summarize occurrences using count(tweet_id, hashtags).
    4. Convert to DFM using cast_dfm(tweet_id, hashtags, n).
  • Save this transformation by creating a new dataset with the name tweets_dfm_cleaned.
• Create new dataset top_hashtags_tidy
  • Based on the dataset tweets_detox,
    1. Repeat steps 1. & 2. from before.
    2. Summarize occurrences using count(hashtags, sort = TRUE).
    3. Extract top 50 hashtags by using slice_head(n = 50).
    4. Convert to string vector by using pull() .
  • Save this transformation by creating a new dataset with the name top_hashtags_tidy.
• Visualize co-occurence
  • Based on the dataset tweets_dfm_cleaned,
    1. Transform data to feature co-occurrence matrix [FCM] using quanteda::fcm()
    2. Select relevant hashtags using quanteda::fcm_select(pattern = top_hashtags_tidy, case_insensitive = FALSE).
    3. Visualize using quanteda.textplots::textplot_network()
• Interpret and compare results
  • Analyze patterns or connections among top hashtags.
  • Discuss insights from the visualization, especially in comparison to the visualization of the slides/showcase.

with quanteda

• Create new dataset quanteda_dfm_cleaned:
  • Based on the dataset quanteda_dfm,
    • Use quanteda::dfm_select(pattern = "#*") to create a DFM containing only hashtags.
    • Use quanteda::dfm_remove(pattern = "#digitaldetox") to removes the hashtag "#digitaldetox".
  • Save this transformation by creating a new dataset with the name quanteda_dfm_cleaned.
• Create new dataset top_hashtags_quanteda
  • Based on the dataset quanteda_dfm_cleaned,
    • Use topfeatures(50) o extract the top 50 most common hashtags.
    • Use names() to only store the names (not the values).
  • Save this transformation by creating a new dataset with the name top_hashtags_quanteda.
• Visualize co-occurence
  • Based on the dataset quanteda_dfm_cleaned,
    1. Transform data to feature co-occurrence matrix [FCM] using quanteda::fcm()
    2. Select relevant hashtags using quanteda::fcm_select(pattern = top_hashtags_tidy, case_insensitive = FALSE).
    3. Visualize using quanteda.textplots::textplot_network() .
• Interpret and compare results
  • Analyze patterns or connections among top hashtags.
  • Discuss insights from the visualization, especially in comparison to the visualization of the slides/showcase.

# Create new dataset tweets_dfm_cleaned | quanteda_dfm_cleaned

# Create new dataset top_hashtags_quanteda |top_hashtags_quanteda

# Visualize co-occurence

📋 Exercise 2: Understanding Topic 9

Objective(s)

Take a closer look at another topic from the topic model used in the session by examining the most representative tweets and the users who post the most tweets on that topic.

2.1: Explore top tweets

• Based on the dataset tweets_detox_topics
  • Use filter() to only analyze tweets that belong to topic 9. Use the variable top_topic for filtering.
  • Arrange the selected tweets in descending order based on top_gamma values using arrange(-top_gamma).
  • Extract the top 10 tweets using slice_head(n = 10).
  • Select only relevant columns (tweet_id, user_username, created_at, text, top_gamma) using select().
  • Create a tabular presentation of the selected tweets using gt() from the gt package.
• Interpret and compare results (with a partner)

2.1 Explore top users

• Based on the dataset tweets_detox_topics
  • Use filter() to only analyze tweets that belong to topic 9. Use the variable top_topic for filtering.
  • Count the number of tweets per user using count(user_username, sort = TRUE).
  • Calculate the proportion of tweets for each user by creating a new column prop using mutate(prop = round(n/sum(n)*100, 2)).
  • Extract the top 15 users with the highest engagement using slice_head(n = 15).
  • Create a tabular presentation of the selected tweets using gt() from the gt package.
• Interpret and compare results (with a partner)

📋 Exercise 3: Expolore different topic model

Objective(s)

In the session, three models were considered for closer examination. Choose one of the other topics and recreate all the steps for the initial exploration of the topic model (as in the session).

3.1 Initial exploration

• Explore a different topic model (estimation)
  • Based on the dataset stm_results,
    • Use the filter(k == X) to select a different topic model (estimation) of your choice (by defining X).
      Tip: Use stm_results$k to see available options for X
    • Use pull(mdl) %>% .[[1]] to select the topic model (estimation) with the specified number of topics.
  • Save this transformation by creating a new dataset with the name tpm_new.
• Visual exploration of tpm_new
  • Visualize the summary of the selected topic model using plot(type = "summary") .

# Explore a different topic model (estimation)
tpm_new <- stm_results |>
   filter(k == ) |> 
   pull(mdl) %>% .[[1]]

# Visual exploration of tpm_new

3.2 Document/Word-topic relations

• Calculate mean gamma values
  • based on the dataset tpm_new
    • Use tidy(matrix = "gamma") on tpm_new to get the gamma values.
    • Calculate the mean gamma values for each topic using group_by() and summarise().
    • Use arrange() to sort topics (descending) by gamma
  • Save this transformation by creating a new dataset with the name top_gamma_new.
• Identify top terms of each topic
  • based on the dataset tpm_new
    • Use tidy(matrix = "beta") on tpm_new to get the beta values.
    • Arrange terms within each topic in descending order based on beta values using group_by(), arrange(-beta), and top_n(10, wt = beta).
    • Select relevant columns (topic, term) using select().
    • Summarize the top 10 terms for each topic using summarise(terms_beta = toString(term), .groups = "drop").
  • Save this transformation by creating a new dataset with the name top_beta_new.
• Combine top topics and top terms
  • Join the dataframes top_gamma_new and top_beta_new based on the “topic” column using left_join().
  • Within mutate(), - Adjusted topic names by topic = paste0("Topic ", topic) and - Reorder the dataset with topic = reorder(topic, gamma)
  • Save this transformation by creating a new dataset with the name top_topics_terms_new.
• Preview the results
  • Display a table preview of the top topics and terms with rounded gamma values using gt()

# Calculate mean gamma values

# Identify top terms of each topic

# Combine top topics and top terms

# Preview the esults

References

Barrie, Christopher, and Justin Ho. 2021. “academictwitteR: An r Package to Access the Twitter Academic Research Product Track V2 API Endpoint.” Journal of Open Source Software 6 (62): 3272. https://doi.org/10.21105/joss.03272.

Nassen, Lise-Marie, Heidi Vandebosch, Karolien Poels, and Kathrin Karsay. 2023. “Opt-Out, Abstain, Unplug. A Systematic Review of the Voluntary Digital Disconnection Literature.” Telematics and Informatics 81 (June): 101980. https://doi.org/10.1016/j.tele.2023.101980.