Sentiment Analysis: From Jane Austen to The Office

Author

Kevin Villa

Published

November 3, 2026

This report has two parts. Part 1 reproduces the sentiment analysis from Chapter 2 of Text Mining with R (Silge and Robinson, 2017), which scores the novels of Jane Austen with three sentiment lexicons. Part 2 applies the same approach to a very different text, the complete scripts of the TV show The Office, and adds a fourth lexicon, Jockers-Rinker from the lexicon package.

Setup

library(tidytext)
library(janeaustenr)
library(dplyr)
library(stringr)
library(tidyr)
library(ggplot2)
library(wordcloud)
library(reshape2)
library(schrute)
library(lexicon)

The AFINN and NRC lexicons are downloaded through the textdata package. Each one asks you to accept its license the first time, so I ran get_sentiments("afinn") and get_sentiments("nrc") once in the console before rendering.

Part 1: Reproducing the Chapter 2 example

All code in Part 1 comes from Chapter 2, “Sentiment analysis with tidy data,” of Silge, J., & Robinson, D. (2017). Text Mining with R: A Tidy Approach. O’Reilly Media. The chapter is available at https://www.tidytextmining.com/sentiment.html, and the book’s source code is at https://github.com/dgrtwo/tidy-text-mining.

The sentiment lexicons

get_sentiments("afinn")
# A tibble: 2,477 × 2
   word       value
   <chr>      <dbl>
 1 abandon       -2
 2 abandoned     -2
 3 abandons      -2
 4 abducted      -2
 5 abduction     -2
 6 abductions    -2
 7 abhor         -3
 8 abhorred      -3
 9 abhorrent     -3
10 abhors        -3
# ℹ 2,467 more rows
get_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 rows
get_sentiments("nrc")
# A tibble: 13,872 × 2
   word        sentiment
   <chr>       <chr>    
 1 abacus      trust    
 2 abandon     fear     
 3 abandon     negative 
 4 abandon     sadness  
 5 abandoned   anger    
 6 abandoned   fear     
 7 abandoned   negative 
 8 abandoned   sadness  
 9 abandonment anger    
10 abandonment fear     
# ℹ 13,862 more rows

Sentiment analysis with inner join

tidy_books <- austen_books() %>%
  group_by(book) %>%
  mutate(
    linenumber = row_number(),
    chapter = cumsum(str_detect(text,
                                regex("^chapter [\\divxlc]",
                                      ignore_case = TRUE)))) %>%
  ungroup() %>%
  unnest_tokens(word, text)

nrc_joy <- get_sentiments("nrc") %>%
  filter(sentiment == "joy")

emma_joy <- tidy_books %>%
  filter(book == "Emma") %>%
  inner_join(nrc_joy) %>%
  count(word, sort = TRUE)

emma_joy
# A tibble: 301 × 2
   word          n
   <chr>     <int>
 1 good        359
 2 friend      166
 3 hope        143
 4 happy       125
 5 love        117
 6 deal         92
 7 found        92
 8 present      89
 9 kind         82
10 happiness    76
# ℹ 291 more rows
jane_austen_sentiment <- tidy_books %>%
  inner_join(get_sentiments("bing")) %>%
  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")

Comparing the three sentiment dictionaries

pride_prejudice <- tidy_books %>%
  filter(book == "Pride & Prejudice")

afinn <- pride_prejudice %>%
  inner_join(get_sentiments("afinn")) %>%
  group_by(index = linenumber %/% 80) %>%
  summarise(sentiment = sum(value)) %>%
  mutate(method = "AFINN")

bing_and_nrc <- bind_rows(
  pride_prejudice %>%
    inner_join(get_sentiments("bing")) %>%
    mutate(method = "Bing et al."),
  pride_prejudice %>%
    inner_join(get_sentiments("nrc") %>%
                 filter(sentiment %in% c("positive",
                                         "negative"))
    ) %>%
    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")

get_sentiments("nrc") %>%
  filter(sentiment %in% c("positive", "negative")) %>%
  count(sentiment)
# A tibble: 2 × 2
  sentiment     n
  <chr>     <int>
1 negative   3316
2 positive   2308
get_sentiments("bing") %>%
  count(sentiment)
# A tibble: 2 × 2
  sentiment     n
  <chr>     <int>
1 negative   4781
2 positive   2005

Most common positive and negative words

bing_word_counts <- tidy_books %>%
  inner_join(get_sentiments("bing")) %>%
  count(word, sentiment, sort = TRUE) %>%
  ungroup()

bing_word_counts
# A tibble: 2,585 × 3
   word     sentiment     n
   <chr>    <chr>     <int>
 1 miss     negative   1855
 2 well     positive   1523
 3 good     positive   1380
 4 great    positive    981
 5 like     positive    725
 6 better   positive    639
 7 enough   positive    613
 8 happy    positive    534
 9 love     positive    495
10 pleasure positive    462
# ℹ 2,575 more rows
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)

custom_stop_words <- bind_rows(tibble(word = c("miss"),
                                      lexicon = c("custom")),
                               stop_words)

custom_stop_words
# A tibble: 1,150 × 2
   word        lexicon
   <chr>       <chr>  
 1 miss        custom 
 2 a           SMART  
 3 a's         SMART  
 4 able        SMART  
 5 about       SMART  
 6 above       SMART  
 7 according   SMART  
 8 accordingly SMART  
 9 across      SMART  
10 actually    SMART  
# ℹ 1,140 more rows

Wordclouds

tidy_books %>%
  anti_join(stop_words) %>%
  count(word) %>%
  with(wordcloud(word, n, max.words = 100))

tidy_books %>%
  inner_join(get_sentiments("bing")) %>%
  count(word, sentiment, sort = TRUE) %>%
  acast(word ~ sentiment, value.var = "n", fill = 0) %>%
  comparison.cloud(colors = c("gray20", "gray80"),
                   max.words = 100)

Looking at units beyond just words

p_and_p_sentences <- tibble(text = prideprejudice) %>%
  unnest_tokens(sentence, text, token = "sentences")

p_and_p_sentences$sentence[2]
[1] "by jane austen"
austen_chapters <- austen_books() %>%
  group_by(book) %>%
  unnest_tokens(chapter, text, token = "regex",
                pattern = "Chapter|CHAPTER [\\dIVXLC]") %>%
  ungroup()

austen_chapters %>%
  group_by(book) %>%
  summarise(chapters = n())
# A tibble: 6 × 2
  book                chapters
  <fct>                  <int>
1 Sense & Sensibility       51
2 Pride & Prejudice         62
3 Mansfield Park            49
4 Emma                      56
5 Northanger Abbey          32
6 Persuasion                25
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      43           161  3405 0.0473
2 Pride & Prejudice        34           111  2104 0.0528
3 Mansfield Park           46           173  3685 0.0469
4 Emma                     15           151  3340 0.0452
5 Northanger Abbey         21           149  2982 0.0500
6 Persuasion                4            62  1807 0.0343

The most common joy words in Emma are “good,” “friend,” and “hope.” The three lexicons follow a similar path through Pride and Prejudice, even though their values are on different scales. The top negative word across the novels is “miss.” As the book points out, Bing reads “miss” as negative, but Austen mostly uses it as a title.

Part 2: Extending the analysis to The Office

A new corpus: The Office scripts

The schrute package (Lindblad) contains the complete scripts of The Office, with one row per line of dialogue. Each row records the season, the episode, and the character speaking. Spoken comedy is a useful contrast to Austen. The language is casual and full of filler words, and a lot of what is said is sarcastic.

office <- schrute::theoffice %>%
  mutate(season = as.integer(season),
         episode = as.integer(episode),
         character = str_trim(character)) %>%
  arrange(season, episode, index) %>%
  mutate(ep_num = dense_rank(season * 100 + episode))

tidy_office <- office %>%
  select(season, episode, ep_num, episode_name, character, text) %>%
  unnest_tokens(word, text)

office %>%
  summarise(lines = n(), episodes = n_distinct(ep_num),
            seasons = n_distinct(season))
# A tibble: 1 × 3
  lines episodes seasons
  <int>    <int>   <int>
1 55130      186       9
nrow(tidy_office)
[1] 571612

Here every episode is one unit of text, the way the book used 80-line sections of each novel.

A fourth lexicon: Jockers-Rinker

The lexicon package (Rinker) includes hash_sentiment_jockers_rinker. This lexicon combines the Syuzhet lexicon (Jockers, 2017) with the Hu and Liu (2004) word list, which is the same source as Bing. Each word gets a graded score instead of just positive or negative, like AFINN, but the list is much larger. A few entries are two-word phrases. Because the text is split into single words, those can’t match, so I removed them.

jockers_rinker <- as_tibble(lexicon::hash_sentiment_jockers_rinker)
names(jockers_rinker) <- c("word", "value")
jockers_rinker <- jockers_rinker %>%
  filter(!str_detect(word, " "))

nrow(jockers_rinker)
[1] 11423
range(jockers_rinker$value)
[1] -2  1

After removing the phrases, Jockers-Rinker has 11,423 words, with scores from -2 to 1. AFINN has 2,477 words, with scores from -5 to 5.

How each lexicon scores a text

This function scores every unit of a text with all four lexicons. AFINN and Jockers-Rinker add up word scores. Bing and NRC count positive words minus negative words, the same way the book does. Each score is divided by the number of words in the unit and multiplied by 1,000, so long and short episodes can be compared fairly.

score_units <- function(tidy_text, unit) {
  words <- tidy_text %>% count({{ unit }}, name = "words")

  scored <- bind_rows(
    tidy_text %>%
      inner_join(get_sentiments("afinn"), by = "word") %>%
      group_by({{ unit }}) %>%
      summarise(score = sum(value)) %>%
      mutate(method = "AFINN"),
    tidy_text %>%
      inner_join(jockers_rinker, by = "word") %>%
      group_by({{ unit }}) %>%
      summarise(score = sum(value)) %>%
      mutate(method = "Jockers-Rinker"),
    tidy_text %>%
      inner_join(get_sentiments("bing"), by = "word") %>%
      count({{ unit }}, sentiment) %>%
      pivot_wider(names_from = sentiment, values_from = n, values_fill = 0) %>%
      mutate(score = positive - negative, method = "Bing et al.") %>%
      select(-positive, -negative),
    tidy_text %>%
      inner_join(get_sentiments("nrc") %>%
                   filter(sentiment %in% c("positive", "negative")),
                 by = "word") %>%
      count({{ unit }}, sentiment) %>%
      pivot_wider(names_from = sentiment, values_from = n, values_fill = 0) %>%
      mutate(score = positive - negative, method = "NRC") %>%
      select(-positive, -negative)
  )

  scored %>%
    left_join(words) %>%
    mutate(per_1000 = score / words * 1000)
}

office_scores <- score_units(tidy_office, ep_num) %>%
  left_join(distinct(office, ep_num, season, episode), by = "ep_num")

Sentiment across the series

office_scores %>%
  filter(method == "Bing et al.") %>%
  ggplot(aes(episode, per_1000, fill = per_1000 > 0)) +
  geom_col(show.legend = FALSE) +
  scale_fill_manual(values = c("TRUE" = "#2a78d6", "FALSE" = "#eb6834")) +
  scale_x_continuous(breaks = seq(0, 30, by = 5)) +
  facet_wrap(~season, ncol = 3, scales = "free_x",
             labeller = labeller(season = \(s) paste("Season", s))) +
  labs(title = "Net sentiment in each episode of The Office (Bing)",
       x = "Episode", y = "Net sentiment per 1,000 words") +
  theme_minimal()

season_avg <- office_scores %>%
  filter(method == "Bing et al.") %>%
  group_by(season) %>%
  summarise(avg_per_1000 = mean(per_1000), episodes = n())

knitr::kable(season_avg, digits = 1)
season avg_per_1000 episodes
1 27.8 6
2 23.7 22
3 20.9 23
4 22.8 14
5 20.7 26
6 19.3 24
7 22.0 24
8 21.7 24
9 19.4 23 
best_season  <- season_avg$season[which.max(season_avg$avg_per_1000)]
worst_season <- season_avg$season[which.min(season_avg$avg_per_1000)]

bing_episodes <- office_scores %>% filter(method == "Bing et al.")
pct_positive_eps <- mean(bing_episodes$per_1000 > 0)

austen_chunks <- jane_austen_sentiment
pct_positive_austen <- mean(austen_chunks$sentiment > 0)

shape <- if (pct_positive_eps > pct_positive_austen) {
  "The Office is more consistently positive. A sitcom resets every week instead of building one long story arc with dark stretches."
} else {
  "The Office is not more consistently positive than Austen, even though it is a comedy, which may come from words like \"funny\" and \"joke\" being scored as negative (see below)."
}

By the Bing lexicon, Season 1 is the most positive season on average and Season 6 is the least positive. 99% of episodes come out net positive. For comparison, 76% of the 80-line sections of the Austen novels are net positive.

Comparing the four lexicons

office_scores %>%
  ggplot(aes(ep_num, per_1000, fill = method)) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~method, ncol = 1, scales = "free_y") +
  labs(title = "Four lexicons scoring every episode of The Office",
       x = "Episode (in series order)", y = "Sentiment per 1,000 words") +
  theme_minimal()

To see how closely the lexicons agree, I correlated their episode scores. I did the same for the 80-line sections of Pride and Prejudice, so the two texts can be compared directly.

lexicon_cor <- function(scores, unit) {
  scores %>%
    select({{ unit }}, method, per_1000) %>%
    pivot_wider(names_from = method, values_from = per_1000) %>%
    select(-{{ unit }}) %>%
    cor(use = "pairwise.complete.obs")
}

pp_scores <- score_units(pride_prejudice %>% mutate(index = linenumber %/% 80), index)

office_cor <- lexicon_cor(office_scores, ep_num)
pp_cor     <- lexicon_cor(pp_scores, index)

knitr::kable(office_cor, digits = 2, caption = "The Office: correlation between lexicons (by episode)")
The Office: correlation between lexicons (by episode)
AFINN Jockers-Rinker Bing et al. NRC
AFINN 1.00 0.81 0.79 0.60
Jockers-Rinker 0.81 1.00 0.84 0.74
Bing et al. 0.79 0.84 1.00 0.61
NRC 0.60 0.74 0.61 1.00 
knitr::kable(pp_cor, digits = 2, caption = "Pride and Prejudice: correlation between lexicons (by 80-line section)")
Pride and Prejudice: correlation between lexicons (by 80-line section)
AFINN Jockers-Rinker Bing et al. NRC
AFINN 1.00 0.84 0.81 0.72
Jockers-Rinker 0.84 1.00 0.90 0.85
Bing et al. 0.81 0.90 1.00 0.73
NRC 0.72 0.85 0.73 1.00 
avg_cor <- function(m) mean(m[upper.tri(m)])

agreement <- if (avg_cor(office_cor) >= 0.5) {
  "The book's point still holds for The Office: the lexicons give different sizes of sentiment but mostly agree on which episodes are more positive or more negative."
} else {
  "Unlike in the book, the lexicons often disagree about which episodes are more positive, which suggests that word lists built for other kinds of writing fit casual dialogue less well."
}

The average correlation between lexicons is 0.73 for The Office and 0.81 for Pride and Prejudice.

How much of each text the lexicons can score

A lexicon can only score words it contains. This table shows the share of all words in each text that each lexicon matched.

coverage <- function(tidy_text, corpus) {
  tibble(
    corpus = corpus,
    AFINN = mean(tidy_text$word %in% get_sentiments("afinn")$word),
    `Bing et al.` = mean(tidy_text$word %in% get_sentiments("bing")$word),
    NRC = mean(tidy_text$word %in% (get_sentiments("nrc") %>%
                 filter(sentiment %in% c("positive", "negative")))$word),
    `Jockers-Rinker` = mean(tidy_text$word %in% jockers_rinker$word)
  )
}

coverage_table <- bind_rows(
  coverage(tidy_books, "Jane Austen (all six novels)"),
  coverage(tidy_office, "The Office (all episodes)")
)

coverage_table %>%
  mutate(across(-corpus, \(x) scales::percent(x, accuracy = 0.1))) %>%
  knitr::kable()
corpus AFINN Bing et al. NRC Jockers-Rinker
Jane Austen (all six novels) 6.5% 7.2% 8.6% 12.2%
The Office (all episodes) 8.4% 7.0% 6.3% 11.0%

Most common positive and negative words

office_bing_counts <- tidy_office %>%
  inner_join(get_sentiments("bing"), by = "word") %>%
  count(word, sentiment, sort = TRUE)

office_bing_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) +
  scale_fill_manual(values = c(positive = "#2a78d6", negative = "#eb6834")) +
  facet_wrap(~sentiment, scales = "free_y") +
  labs(title = "Words that drive sentiment in The Office (Bing)",
       x = "Contribution to sentiment", y = NULL) +
  theme_minimal()

top_pos <- office_bing_counts %>% filter(sentiment == "positive") %>% slice_head(n = 3)
top_neg <- office_bing_counts %>% filter(sentiment == "negative") %>% slice_head(n = 3)

The top positive words in The Office are “like,” “right,” and “well,” and the top negative words are “sorry,” “bad,” and “wrong.”

Fixing words that don’t fit the show

The book found that “miss” is a problem for Austen. The Office has the same kind of problem with different words.

  • Filler words scored as positive: Bing scores “like,” “well,” “right,” and “pretty” as positive. In conversation they are mostly filler (“it’s, like, fine”) or hedges (“pretty bad”).
  • Comedy words scored as negative: Bing scores “funny” and “joke” as negative, because it reads “funny” as meaning strange. In a comedy, both words are usually neutral or positive.
office_stop_words <- tibble(word = c("like", "well", "right", "pretty", "funny", "joke"))

office_stop_words %>%
  inner_join(office_bing_counts, by = "word") %>%
  knitr::kable()
word sentiment n
like positive 2986
well positive 2284
right positive 2419
pretty positive 314
funny negative 203
joke negative 119 
before_after <- bind_rows(
  tidy_office %>% mutate(version = "All words"),
  tidy_office %>% anti_join(office_stop_words, by = "word") %>%
    mutate(version = "Filler and comedy words removed")
) %>%
  inner_join(get_sentiments("bing"), by = "word") %>%
  count(version, sentiment) %>%
  pivot_wider(names_from = sentiment, values_from = n) %>%
  mutate(positive_share = positive / (positive + negative))

knitr::kable(before_after, digits = 3)
version negative positive positive_share
All words 13872 26120 0.653
Filler and comedy words removed 13550 18117 0.572

Removing these six words changes the positive share of Bing matches from 65.3% to 57.2%.

The most positive characters

main_characters <- tidy_office %>%
  count(character, name = "words") %>%
  filter(words >= 10000)

character_sentiment <- tidy_office %>%
  semi_join(main_characters, by = "character") %>%
  inner_join(jockers_rinker, by = "word") %>%
  group_by(character) %>%
  summarise(score = sum(value)) %>%
  left_join(main_characters, by = "character") %>%
  mutate(per_1000 = score / words * 1000) %>%
  arrange(desc(per_1000))

ggplot(character_sentiment, aes(per_1000, reorder(character, per_1000))) +
  geom_col(fill = "#2a78d6", width = 0.7) +
  labs(title = "Sentiment by character (Jockers-Rinker)",
       subtitle = "Characters with at least 10,000 spoken words",
       x = "Sentiment per 1,000 words", y = NULL) +
  theme_minimal()

Among the characters with at least 10,000 spoken words, Jim is the most positive by the Jockers-Rinker lexicon and Dwight is the least positive.

How the results differ from the original example

  • Shape over time: Austen’s novels move between positive and negative as the plot develops, and the book’s plots show dips at the dramatic moments. In The Office, 99% of episodes are net positive, compared with 76% of the Austen sections. The Office is more consistently positive. A sitcom resets every week instead of building one long story arc with dark stretches.
  • Coverage: Jockers-Rinker matched 11.0% of the words in The Office, compared with 8.4% for AFINN and 7.0% for Bing. For Austen the same three numbers are 12.2%, 6.5%, and 7.2%.
  • Agreement between lexicons: The average correlation between lexicons is 0.73 for The Office and 0.81 for Pride and Prejudice. The book’s point still holds for The Office: the lexicons give different sizes of sentiment but mostly agree on which episodes are more positive or more negative.
  • Domain words: For Austen, the problem word was “miss.” For The Office, filler words (“like,” “well,” “right”) and comedy words (“funny,” “joke”) are scored in ways that don’t match how they are used. Removing them changed the positive share of Bing matches from 65.3% to 57.2%.
  • Limits: Every lexicon here scores single words. Sarcasm, which The Office relies on, and phrases like “not good” are scored as if they were sincere. A sentence-level tool like sentimentr would handle negation better.

References