Text Analysis
Steven Liu
May 2019
1 Word Frequency Summary
Load data
source("R/utils.R")
theme_set(theme_light())
tw <- readRDS("data/tweets_simp_522.rds")
users <- readRDS("data/users_simp_521.rds")1.1 Top Hashtags (other than #Monsanto)
ht.top <- tw %>%
related_hashtags_freq(upper = T) %>%
sort() %>%
as.data.frame() %>%
filter(tags.all != "MONSANTO") %>%
top_n(15, wt = Freq)
ht.top %>%
ggplot(aes(tags.all, Freq)) +
geom_col() +
labs(x = "", y = "Number of appearances") +
coord_flip()
1.1.1 Hashtag cooccurrences
ht.net <-
readRDS("data/ht_adj_mat.rds") %>%
graph_from_adjacency_matrix(mode = "undirected", weighted = T, diag = F)
htop <- induced_subgraph(ht.net, as.vector(ht.top$tags.all))
graph_attr(htop, "layout") <- layout_in_circle
plot(
htop,
vertex.shape = "none",
edge.color = "orange",
edge.width = E(ht.net)$weight / 3,
vertex.label.dist = 0,
vertex.label.color = "steel blue",
vertex.label.font = 1.5,
vertex.label.cex = .7,
vertex.color = "gray50"
)
- The edge weight indicates how frequently a pair of hashtags would appear at the same time.
1.2 Basic Word Count
1.2.1 For all tweets
tw.words <- tw %>% clean_tweets()
tw.words %>%
plot_word_freq()
1.2.1.1 Would cloud for all tweets
words.count <- tw.words %>% count(word, sort = T)
wordcloud(words = words.count$word, freq = words.count$n,
random.order = F, max.words = 200, rot.per = 0.15,
colors = brewer.pal(8, "Dark2"))
1.2.2 For all tweets except tweets from bots
tw %>%
filter(!is.bot) %>%
clean_tweets() %>%
plot_word_freq()
1.2.3 For all unique tweets except tweets from bots
tw %>%
filter(!is.bot) %>%
distinct(text) %>%
clean_tweets() %>%
plot_word_freq()
1.2.4 For all retweets except tweets from bots
tw %>%
filter(is_retweet & !is.bot) %>%
clean_tweets() %>%
plot_word_freq()
1.2.5 For all tweets by bots
tw %>%
filter(is.bot) %>%
clean_tweets() %>%
plot_word_freq()
2 Sentiment Extraction
2.1 Unigram-based Analysis
2.1.1 Most common positive and negative words
words.bing <- tw.words %>%
inner_join(get_sentiments("bing"), by = "word")
words.bing %>%
count(word, sentiment, sort = T) %>%
group_by(sentiment) %>%
top_n(10, wt = n) %>%
ungroup() %>%
mutate(word = reorder(word, n)) %>%
ggplot(aes(word, n, fill = sentiment)) +
geom_col(show.legend = F) +
facet_wrap(~sentiment, scales = "free_y") +
labs(y = "Contribution to sentiment", x = NULL) +
coord_flip()
2.1.2 Sentiment Wordclouds
words.bing %>%
count(word, sentiment, sort = TRUE) %>%
reshape2::acast(word ~ sentiment, value.var = "n", fill = 0) %>%
comparison.cloud(colors = c("gray20", "gray80"), max.words = 150)
2.2 Sentence-based Analysis
The sentence-based sentiment analysis is pre-performed using sentimentr package. Each tweet was assigned a polarity score (positive/negative) by sentence-based sentiment mining algorithm. The dictionary Jockers (2017) is used for our analysis.
senti_score <-
tw$text %>%
sentimentr::get_sentences() %>%
sentimentr::sentiment_by(polarity_dt = lexicon::hash_sentiment_jockers_rinker)
tw.ex <- mutate(tw, senti = senti_score$ave_sentiment) 2.2.1 Hitogram of sentiment score for all tweets
plot_density(tw.ex, "senti") +
labs(x = "sentiment score")
- Retweet may not mean endorsement.
2.2.2 Hitogram of sentiment score for all unique tweets
tw.ex %>%
distinct(text, .keep_all = T) %>%
plot_density("senti") +
labs(x = "sentiment score")
2.2.3 Hitogram of sentiment score of all users
users.ex <-
tw.ex %>%
group_by(user_id) %>%
summarise(senti.mean = mean(senti), senti.sd = sd(senti)) %>%
inner_join(users, by = "user_id")
plot_density(users.ex, "senti.mean", bins = 40) +
labs(x = "sentiment score")
2.2.4 Sentiment variation over time
tw.ex %>%
group_by(date = lubridate::floor_date(created_at, unit = "day")) %>%
summarise(senti.mean = mean(senti), senti.sd = sd(senti)) %>%
ggplot(aes(x = date, y = senti.mean)) +
geom_ribbon(aes(ymin = senti.mean - senti.sd, ymax = senti.mean + senti.sd),fill = "grey70") +
geom_line() +
scale_x_datetime(date_breaks = "1 month", date_labels = "%b") +
labs(x = "", y = "avg. sentiment score")
2.2.5 Comparison of sentiment score between user groups
2.2.5.1 Bots vs. Normal users
ggplot(data = users.ex, aes(x = is.bot, y = senti.mean)) +
geom_boxplot() +
labs(x = "", y = "sentiment score") +
coord_flip() +
scale_x_discrete(labels = c("users", "bots"))
t.test(filter(users.ex, is.bot)$senti.mean, filter(users.ex, !is.bot)$senti.mean)##
## Welch Two Sample t-test
##
## data: filter(users.ex, is.bot)$senti.mean and filter(users.ex, !is.bot)$senti.mean
## t = 6.6519, df = 452.66, p-value = 8.368e-11
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## 0.02016328 0.03707310
## sample estimates:
## mean of x mean of y
## -0.1232990 -0.1519172The difference of mean sentiment score between the group “Bots” and the group “Normal users” is significant at a = 0.001, where the tweets posted/retweeted by bots tend to have more neutral sentiments.
2.2.5.2 Active users vs. Less active users
We treat the users that contributed > 7 tweets (1 sd. above the mean) as avtive users.
ggplot(data = users.ex, aes(x = ntweets > 7, y = senti.mean)) +
geom_boxplot() +
labs(x = "", y = "sentiment score") +
coord_flip() +
scale_x_discrete(labels = c("less", "more"))
t.test(filter(users.ex, ntweets > 7)$senti.mean, filter(users.ex, ntweets <= 7)$senti.mean)##
## Welch Two Sample t-test
##
## data: filter(users.ex, ntweets > 7)$senti.mean and filter(users.ex, ntweets <= 7)$senti.mean
## t = 13.575, df = 1376.3, p-value < 2.2e-16
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## 0.03065838 0.04101555
## sample estimates:
## mean of x mean of y
## -0.1165339 -0.1523709The difference of mean sentiment score between the group “Active users” and the group “Less active users” is significant at a = 0.001, where the tweets posted/retweeted by active users tend to have more neutral sentiments. Note that the bots we identified is also a subset of active users.
2.2.5.3 Creator vs. ReTweeter vs. Mixed behaviour users
Quick definition: A creator is who never retweets, a retweeter is who always retweets.
users.ex <- mutate(users.ex,
role = ifelse(ncreate / ntweets == 0, "retweeter",
ifelse(ncreate / ntweets == 1, "creator", "mix")))
ggplot(data = users.ex, aes(x = role, y = senti.mean)) +
geom_boxplot() +
labs(x = "", y = "sentiment score") +
coord_flip() 
role.aov <- aov(senti.mean ~ role, data = users.ex)
summary(role.aov)## Df Sum Sq Mean Sq F value Pr(>F)
## role 2 11.6 5.786 180.1 <2e-16 ***
## Residuals 66637 2141.1 0.032
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1TukeyHSD(role.aov)## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = senti.mean ~ role, data = users.ex)
##
## $role
## diff lwr upr p adj
## mix-creator 0.026947064 0.018405989 0.03548814 0.0000000
## retweeter-creator 0.030367237 0.026607380 0.03412709 0.0000000
## retweeter-mix 0.003420173 -0.004725811 0.01156616 0.5869729The difference in mean sentiment score among the three groups is significant (p < 2e-16). Furthermore, our post-hoc test shows that the difference between the pair mix-creator and the pair retweeter-creator is significant at any confidence level, but there is no significant difference between the pair retweeter-mix. We conclude that the tweets posted by the “creators” tend to have a more negative sentiment.
2.2.5.4 Compare by gender group
ggplot(data = users.ex, aes(x = gender, y = senti.mean)) +
geom_boxplot() +
labs(x = "", y = "sentiment score") +
coord_flip() 
role.aov <- aov(senti.mean ~ gender, data = users.ex)
summary(role.aov)## Df Sum Sq Mean Sq F value Pr(>F)
## gender 3 3.2 1.0775 33.41 <2e-16 ***
## Residuals 66636 2149.4 0.0323
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1TukeyHSD(role.aov)## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = senti.mean ~ gender, data = users.ex)
##
## $gender
## diff lwr upr p adj
## MALE-FEMALE -0.0156877802 -0.020724003 -1.065156e-02 0.0000000
## UNISEX-FEMALE -0.0085301278 -0.016866177 -1.940784e-04 0.0425731
## UNKNOWN-FEMALE -0.0165390657 -0.021045649 -1.203248e-02 0.0000000
## UNISEX-MALE 0.0071576524 -0.001135818 1.545112e-02 0.1185321
## UNKNOWN-MALE -0.0008512856 -0.005278613 3.576042e-03 0.9604574
## UNKNOWN-UNISEX -0.0080089380 -0.015991883 -2.599348e-05 0.0489000The difference in mean sentiment score among gender groups is significant (p < 2e-16). Furthermore, our post-hoc test shows that the difference between the pair male-female and the pair unknown-female is significant at any confidence level, which indicates that the tweets posted by the (known) female users tend to have a more neutral sentiment.
3 Emotion Extraction
We use nrc lexicon for emotion analysis. It categorizes words in a binary fashion (“yes”/“no”) into categories of 8 basic emotions: anger, anticipation, disgust, fear, joy, sadness, surprise, trust and 2 sentiments: positive/negative. So each tweet is associated with a 10-dimensional vector, which stores the count of the words belong to these 10 categories. Here we only look at the first 8 dimensions (emotion part). We then adopt the k-means algorithm to cluster the latent “emotion groups” among the Twitter discussion, and the result is visualized by the Euler diagram. The whole procedure is encapsulated in one simple function plot_nrc().
3.1 For all unique tweets
set.seed(1000) # to fix the plot layout
nrc <- fread("data/nrc_uniq.csv")
nrc %>%
as.matrix() %>%
drop_lastN(2) %>% # ignore sentiment attribute
extract_emo() %>%
euler_km() %>%
plot()
- The dominant emotion among the whole discussion is “fear”, and also its composition with “anger” and “sadness”.