Week 10 Assignment
Taha Ahmad
2023-03-29
library(tidyverse)
library(janeaustenr)
library(tidytext)
library(gutenbergr)
library(wordcloud)
library(reshape2)
library(lexicon)Introduction
With the textbook Text Mining with R by by Julia Silge and David Robinson, we explore utilizing sentiment analysis on text. We begin with mimicking code examples present in the text, and then we extend it to utilize a different corpus and sentiment lexicons.
Example Code
As previously mentioned, the code in this section is sourced from Text Mining with R by by Julia Silge and David Robinson. We will be extending it later.
Loading Jane Austen
First we load in Jane Austen’s books through the janeaustenr package and transform them into a tidy format.
original_books <- austen_books() %>%
group_by(book) %>%
mutate(linenumber = row_number(),
chapter = cumsum(str_detect(text,
regex("^chapter [\\divxlc]",
ignore_case = TRUE)))) %>%
ungroup()Transforming Jane Austen
Then we tokenize our tidy dataframe and remove stop words which are not useful for analysis.
tidy_books <- original_books %>%
unnest_tokens(word, text)tidy_books <- tidy_books %>%
anti_join(stop_words, by = join_by(word))Analyzing Jane Austen
Here we begin the sentiment analysis initial example by a simple count of words that are joyous according to the nrc sentiment dataset within the book Emma.
Something interesting to note here is the differences between our code results and the textbook example. “Young” has been removed from the latest version of the nrc dataset and “good” has been removed for being a stop word.
nrc_joy <- get_sentiments("nrc") %>%
filter(sentiment == "joy")
tidy_books %>%
filter(book == "Emma") %>%
inner_join(nrc_joy, by = join_by(word)) %>%
count(word, sort = TRUE)## # A tibble: 297 × 2
## word n
## <chr> <int>
## 1 friend 166
## 2 hope 143
## 3 happy 125
## 4 love 117
## 5 deal 92
## 6 found 92
## 7 happiness 76
## 8 pretty 68
## 9 true 66
## 10 comfort 65
## # ℹ 287 more rowsNext we analyze the sentiment of the current point where we are at with each book to see how it fluctuates throughout. This is done with the bing sentiment dataset.
jane_austen_sentiment <- tidy_books %>%
inner_join(get_sentiments("bing"), by = join_by(word)) %>%
count(book, index = linenumber %/% 80, sentiment) %>%
pivot_wider(names_from = sentiment, values_from = n, values_fill = 0) %>%
mutate(sentiment = positive - negative)
ggplot(jane_austen_sentiment, aes(index, sentiment, fill = book)) +
geom_col(show.legend = FALSE) +
facet_wrap(~book, ncol = 2, scales = "free_x")
Afterwards, we want to focus on the differences between how each
sentiment dataset quantifies sentiment for a single book, “Pride and
Prejudice”.
pride_prejudice <- tidy_books %>%
filter(book == "Pride & Prejudice")
afinn <- pride_prejudice %>%
inner_join(get_sentiments("afinn"), by = join_by(word)) %>%
group_by(index = linenumber %/% 80) %>%
summarise(sentiment = sum(value)) %>%
mutate(method = "AFINN")
bing_and_nrc <- bind_rows(
pride_prejudice %>%
inner_join(get_sentiments("bing"), by = join_by(word)) %>%
mutate(method = "Bing et al."),
pride_prejudice %>%
inner_join(get_sentiments("nrc") %>%
filter(sentiment %in% c("positive",
"negative"))
, by = join_by(word)) %>%
mutate(method = "NRC")) %>%
count(method, index = linenumber %/% 80, sentiment) %>%
pivot_wider(names_from = sentiment,
values_from = n,
values_fill = 0) %>%
mutate(sentiment = positive - negative)
bind_rows(afinn,
bing_and_nrc) %>%
ggplot(aes(index, sentiment, fill = method)) +
geom_col(show.legend = FALSE) +
facet_wrap(~method, ncol = 1, scales = "free_y")
We can utilize a dataframe with combined word and sentiment to analyze
what is contributing to each sentiment more or less. This can then be
visualized through ggplot.
Another anomaly from removing stop words can be seen here where “well”, “good”, and “like”, which are commonly used for their non-sentiment laden meanings, do not appear in positive sentiment.
bing_word_counts <- tidy_books %>%
inner_join(get_sentiments("bing"), by = join_by(word)) %>%
count(word, sentiment, sort = TRUE) %>%
ungroup()
bing_word_counts %>%
group_by(sentiment) %>%
slice_max(n, n = 10) %>%
ungroup() %>%
mutate(word = reorder(word, n)) %>%
ggplot(aes(n, word, fill = sentiment)) +
geom_col(show.legend = FALSE) +
facet_wrap(~sentiment, scales = "free_y") +
labs(x = "Contribution to sentiment",
y = NULL)
On the other hand miss which is the highest negative sentiment
contributor is likely used not to mean missing a shot, but the formal
title. Thus, it can be added to a custom_stop_words dataset for
removal.
custom_stop_words <- bind_rows(tibble(word = c("miss"),
lexicon = c("custom")),
stop_words)Pivoting to a new route of text analysis, we can consider building word clouds from the text we have.
Comparing our wordcloud to the example, we have to reduce the max words in order to fit each of the common words on screen. Additionally, the shape of the wordcloud seems to be random each time it is built.
tidy_books %>%
count(word) %>%
with(wordcloud(word, n, max.words = 75))
We can extend this wordcloud by visualizing which words are negative and
those that are positive.
tidy_books %>%
inner_join(get_sentiments("bing"), by = join_by(word)) %>%
count(word, sentiment, sort = TRUE) %>%
acast(word ~ sentiment, value.var = "n", fill = 0) %>%
comparison.cloud(colors = c("gray20", "gray80"),
max.words = 100)
Finally, we let’s take a look at answering the question of which chapter
is the most negative for each Jane Austen novel.
Note yet again, due to removing our stop words as part of the beginning step we get different chapters for half of the books and greater ratios!
bingnegative <- get_sentiments("bing") %>%
filter(sentiment == "negative")
wordcounts <- tidy_books %>%
group_by(book, chapter) %>%
summarize(words = n())
tidy_books %>%
semi_join(bingnegative) %>%
group_by(book, chapter) %>%
summarize(negativewords = n()) %>%
left_join(wordcounts, by = c("book", "chapter")) %>%
mutate(ratio = negativewords/words) %>%
filter(chapter != 0) %>%
slice_max(ratio, n = 1) %>%
ungroup()## # A tibble: 6 × 5
## book chapter negativewords words ratio
## <fct> <int> <int> <int> <dbl>
## 1 Sense & Sensibility 29 172 1135 0.152
## 2 Pride & Prejudice 34 108 646 0.167
## 3 Mansfield Park 45 132 884 0.149
## 4 Emma 15 147 1012 0.145
## 5 Northanger Abbey 27 55 337 0.163
## 6 Persuasion 21 215 1948 0.110Extension Code
Here we utilize different text and a new sentiment lexicon following the same general steps as the example code.
Loading A New Corpus
Let’s utilize the Project Gutenberg library to get our own corpus to analyze. I’m a particular fan of John Steinbeck’s Grapes of Wrath so I attempted to search for it, however Project Gutenberg doesn’t seem to have it just yet. Instead, we’ll go with Grapes of Wrath by Boyd Cable which is a work of fiction regarding a WWI soldier’s encounters.
Boyd Cable also has a decent amount of books which fall into the war category as well, but we’ll analyze 3 for better visualization down the line.
boyd_works <- gutenberg_works() %>%
filter(grepl("cable, boyd", author, ignore.case=TRUE))%>%
select(gutenberg_id) %>%
head(3) %>%
pull(gutenberg_id) %>%
gutenberg_download(meta_fields = "title") %>%
select(text, title)Then after downloading his works, we need to ensure that the chapter each book is captured in the dataframe. Unfortunately, only Grapes of Wrath follows a regex matchable chapter pattern. For the other two books we need to match on the chapter lists individually. Then we process the dataframe so we will have the rows show the line.
Transforming a New Corpus
grapes_book <- boyd_works %>%
filter(title == "Grapes of wrath") %>%
mutate(linenumber = row_number(),
chapter = cumsum(str_detect(text,
regex("^chapter [\\divxlc]",
ignore_case = TRUE))))
btn_chp <- c(
"THE ADVANCED TRENCHES",
"SHELLS",
"THE MINE",
"ARTILLERY SUPPORT",
"'NOTHING TO REPORT'",
"THE PROMISE OF SPRING",
"THE ADVANCE",
"A CONVERT TO CONSCRIPTION",
"'BUSINESS AS USUAL'",
"A HYMN OF HATE",
"THE COST",
"A SMOKER'S COMPANION",
"THE JOB OF THE AM. COL.",
"THE SIGNALLER'S DAY"
)
btn_book <- boyd_works %>%
filter(title == "Between the Lines") %>%
mutate(linenumber = row_number(),
chapter = cumsum(grepl(paste(btn_chp, collapse = "|"),text))) %>%
mutate(chapter = ifelse(chapter > 14, chapter-14, 0))
actn_chp <- c(
"IN ENEMY HANDS",
"A BENEVOLENT NEUTRAL",
"DRILL",
"A NIGHT PATROL",
"AS OTHERS SEE",
"THE FEAR OF FEAR",
"ANTI-AIRCRAFT",
"A FRAGMENT",
"AN OPEN TOWN",
"THE SIGNALERS",
"CONSCRIPT COURAGE",
"SMASHING THE COUNTER-ATTACK",
"A GENERAL ACTION",
"AT LAST"
)
actn_book <- boyd_works %>%
filter(title == "Action Front") %>%
mutate(linenumber = row_number(),
chapter = cumsum(grepl(paste(actn_chp, collapse = "|"),text))) %>%
mutate(chapter = ifelse(chapter > 14, chapter-14, 0))After preprocessing the data, we can take the combination of each dataframe to tokenize it and remove stop words.
boyd_token <- rbind(grapes_book, btn_book, actn_book) %>%
unnest_tokens(word, text) %>%
mutate(word = str_extract(word, "[a-z']+")) %>%
anti_join(stop_words, by = join_by(word)) %>%
select(book = title, chapter, linenumber, word)A New Sentiment Dataset
Scouring the internet for an additional unigram sentiment dataset to work with, we end up utilizing the sentiment dataset used within the syuzhet package. However, since the package does not provide direct access to the dataset from R based on my attempts to utilize it, we use the lexicon package to load it in.
The syuzhet sentiment dataset classifies the positivity of unigrams on a sliding scale from -1 to 1.
syuzhet <- lexicon::key_sentiment_jockers %>%
select(word, sentiment = value)
head(syuzhet)## word sentiment
## 1 abandon -0.75
## 2 abandoned -0.50
## 3 abandoner -0.25
## 4 abandonment -0.25
## 5 abandons -1.00
## 6 abducted -1.00Analyzing Grapes of Wrath
Here we begin the sentiment analysis initial example by a simple count of words that are positive according to being a positive value within the syuzhet dataset.
The characteristics of a war story are fully on display here with “forward”, “sir”, “advance”, and other strategic action terms showing up.
boyd_token %>%
filter(book == "Grapes of wrath") %>%
inner_join(syuzhet %>% filter(sentiment > 0), by = join_by(word)) %>%
count(word, sort = TRUE)## # A tibble: 550 × 2
## word n
## <chr> <int>
## 1 forward 57
## 2 chance 37
## 3 found 29
## 4 sir 27
## 5 ready 21
## 6 straight 21
## 7 action 18
## 8 advance 17
## 9 flying 17
## 10 shelter 17
## # ℹ 540 more rowsget_sentiments("bing")## # A tibble: 6,786 × 2
## word sentiment
## <chr> <chr>
## 1 2-faces negative
## 2 abnormal negative
## 3 abolish negative
## 4 abominable negative
## 5 abominably negative
## 6 abominate negative
## 7 abomination negative
## 8 abort negative
## 9 aborted negative
## 10 aborts negative
## # ℹ 6,776 more rowsNext we analyze the sentiment of the current point where we are at with each book to see how it fluctuates throughout. This is done with the syuzhet sentiment dataset again. Perhaps not so surprisingly, as war fiction the sentiment that we see throughout this book is overwhelmingly negative. However, we do see glimpses of hope near the start, middle, and end as a trend. As a purely negative novel does not attract many readers.
boyd_sentiment <- boyd_token %>%
inner_join(syuzhet, by = join_by(word)) %>%
group_by(book, index = linenumber %/% 80) %>%
summarise(sentiment = sum(sentiment), .groups = 'drop')
ggplot(boyd_sentiment, aes(index, sentiment, fill = book)) +
geom_col(show.legend = FALSE) +
facet_wrap(~book, ncol = 2, scales = "free_x")
Afterwards, we want to focus on the differences between how each
sentiment dataset quantifies sentiment for a single book, “Grapes of
Wrath”. With this visualization we can see that the Syuzhet sentiment
set is comparible to NRC in the amount of variation of sentiment that is
possible. This is because of the break down of each individual value
into 8 separate stages of sentiment.
grapes_wrath <- boyd_token %>%
filter(book == "Grapes of wrath")
syuzhet_join <- grapes_wrath %>%
inner_join(syuzhet, by = join_by(word)) %>%
group_by(index = linenumber %/% 80) %>%
summarise(sentiment = sum(sentiment), .groups = 'drop') %>%
mutate(method = "Syuzhet")
afinn <- grapes_wrath %>%
inner_join(get_sentiments("afinn"), by = join_by(word)) %>%
group_by(index = linenumber %/% 80) %>%
summarise(sentiment = sum(value)) %>%
mutate(method = "AFINN")
bing_and_nrc <- bind_rows(
grapes_wrath %>%
inner_join(get_sentiments("bing"), by = join_by(word)) %>%
mutate(method = "Bing et al."),
grapes_wrath %>%
inner_join(get_sentiments("nrc") %>%
filter(sentiment %in% c("positive",
"negative"))
, by = join_by(word)) %>%
mutate(method = "NRC")) %>%
count(method, index = linenumber %/% 80, sentiment) %>%
pivot_wider(names_from = sentiment,
values_from = n,
values_fill = 0) %>%
mutate(sentiment = positive - negative)
bind_rows(afinn,
bing_and_nrc, syuzhet_join) %>%
ggplot(aes(index, sentiment, fill = method)) +
geom_col(show.legend = FALSE) +
facet_wrap(~method, ncol = 1, scales = "free_y")
We can utilize a dataframe with combined word and sentiment to analyze
what is contributing to each sentiment more or less. This can then be
visualized through ggplot.
Here we can observe a quirk of sentiment analysis utilizing the syuzheit library for this corpus. Mentioning a gun, rifle, or attack might be negative in sentiment for many other stories. However, in a war story weapons would tend to be more neutral statements while attacks could be positive.
syuzhet_sentiment_counts <- boyd_token %>%
inner_join(syuzhet, by = join_by(word)) %>%
group_by(word) %>%
summarise(sentiment = sum(sentiment), .groups = 'drop') %>%
mutate(type = ifelse(sentiment>0, "positive", "negative"))
syuzhet_sentiment_counts %>%
group_by(type) %>%
slice_max(abs(sentiment), n = 10) %>%
ungroup() %>%
mutate(word = reorder(word, abs(sentiment))) %>%
ggplot(aes(sentiment, word, fill = sentiment)) +
geom_col(show.legend = FALSE) +
facet_wrap(~type, scales = "free_y") +
labs(x = "Contribution to sentiment",
y = NULL)
We could remove gun and rifle by setting our custom stop words.
custom_stop_words <- bind_rows(tibble(word = c("fire","gun"),
lexicon = c("custom")),
stop_words)Pivoting to a new route of text analysis, we can consider building word clouds from the text we have.
As we know trench is the highest contributor to negative sentiment and since it is referred to so often, it might be wise to add it to the stop list as well. However, I have reservations about this as despite being a common setting for war novels, being in trenches is definitely a negative thing.
boyd_token %>%
count(word) %>%
with(wordcloud(word, n, max.words = 75))
We can extend this word cloud by visualizing which words are negative
and those that are positive.
boyd_token %>%
inner_join(syuzhet, by = join_by(word)) %>%
group_by(word) %>%
summarise(sentiment = sum(sentiment), .groups = 'drop') %>%
mutate(type = ifelse(sentiment>0, "positive", "negative")) %>%
acast(word ~ type, value.var = "sentiment", fill = 0) %>%
comparison.cloud(colors = c("gray20", "gray80"),
max.words = 100)
Finally, we let’s take a look at answering the question of which chapter
is the most negative for each Boyd Cable novel.
“Between the Lines” has the most negative ratio towards the start of the novel. While “Action Front” and “Grapes of Wrath” both get more negative towards the end.
syuzhet_negative <- syuzhet %>%
filter(sentiment < 0)
wordcounts <- boyd_token %>%
group_by(book, chapter) %>%
summarize(words = n())
boyd_token %>%
semi_join(syuzhet_negative) %>%
group_by(book, chapter) %>%
summarize(negativewords = n()) %>%
left_join(wordcounts, by = c("book", "chapter")) %>%
mutate(ratio = negativewords/words) %>%
filter(chapter != 0) %>%
slice_max(ratio, n = 1) %>%
ungroup()## # A tibble: 3 × 5
## book chapter negativewords words ratio
## <chr> <dbl> <int> <int> <dbl>
## 1 Action Front 11 303 1214 0.250
## 2 Between the Lines 1 224 971 0.231
## 3 Grapes of wrath 14 173 714 0.242Conclusions
Sentiment mining from text is a good way to process data that might otherwise not be perfectly quantifiable. However, this naive version of sentiment mining we have gone through in this assignment has many weak points. For example, any negation words such as “not” are not kept in mind. We also saw with the Boyd Cable novels that the existing sentiment dictionaries we have might not be perfectly suited for charting the general sentiment in a war novel as many war related terms are detected as negative within these dictionaries. The corpus you are analyzing must be considered when choosing a proper sentiment dictionary for sentiment mining.
If I were to extend this assignment further, I would like to attempt utilizing bigram analysis to encapsulate negation words and compare the sentiment to this unigram based analysis.