The code below is cited from Chapter
2 of Welcome to Text Mining with R: A Tidy Approach by
Julia Silge and David Robertson, which shows how the
tidytext library can be used to evaluate the sentiment of
text. The example I have chosen shows how the nrc lexicon
can be used in order to get the most commonly used words in Jane Austen
novels that are associated with joy (text data used comes from the
janeaustenr package).
# clean the Jane Austen text data
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)
# get those words from the nrc lexicon mapped to "joy"
nrc_joy <- get_sentiments("nrc") %>%
filter(sentiment == "joy")
# join the lexicon mappings to the Jane Austen book data.
tidy_books %>%
filter(book == "Emma") %>%
inner_join(nrc_joy) %>%
count(word, sort = TRUE)## Joining, by = "word"## # 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
## # … with 291 more rowsAs expected, the words shown above all exhibit a certain sense of joy.
The example above uses the nrc lexicon in order to
perform this analysis, which maintains a dictionary of words that are
mapped to different feelings or emotions. The
get_sentiments() function can be used to show these
mappings:
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
## # … with 13,862 more rowsIn addition to nrc there are numerous other lexicons
that can be used, each providing a slightly different methodology to
perform sentiment analysis.
This section will use a different lexicon to analyze a number of
Tweets that have been written concerning different entities. The data
was pulled from Kaggle
and was uploaded to Github.
Each row of the data set pertains to a tweet about a different entity
(i.e. Microsoft, Verizon, HomeDepot). The data has also been labelled to
include the sentiment of each tweet (positive, negative, neutral) so
that we can check the quality of our sentiment analysis. The data is
downloaded and stored as a dataframe tweets in the code
chunk below:
link <- getURL('https://raw.githubusercontent.com/williamzjasmine/CUNY_SPS_DS/master/DATA_607/Homeworks/HW7/twitter_entity.csv')
tweets <- read_csv(link, na=c("", "NA"))## Rows: 74682 Columns: 4
## ── Column specification ────────────────────────────────────────────────────────
## Delimiter: ","
## chr (3): entity, sentiment, tweet_content
## dbl (1): tweet_id
##
## ℹ Use `spec()` to retrieve the full column specification for this data.
## ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.dim(tweets)## [1] 74682 4Based on the output above, we see that the tweets
dataframe contains 74,682 tweets.
Before beginning the sentiment analysis, there are a number of data
cleaning steps that need to be performed. To limit the scope of the
data, the code chunk below filters tweets to include only
those tweets concerning Facebook, Google, and Amazon. It also limits the
only sentiments to “positive”, “neutral”, or “negative” (excluding
“irrelevant”).
tweets <- tweets %>%
filter(entity %in% c('Amazon', 'Google', 'Facebook')) %>%
filter(sentiment %in% c('Positive', 'Neutral', 'Negative')) Because the data appears to include multiple tweets for the same
tweet_id, the code below picks the first row for each
tweet_id:
tweets <- tweets[!duplicated(tweets$tweet_id),]
dim(tweets)## [1] 930 4The output above shows that we have shrunk down the
tweets data frame to now only include 930 tweets concerning
three different entities. This will make the following sentiment
analysis much more manageable.
Next, before tokenizing our data, the cell below pulls out the
tweet_id and sentiment values for each row so
that we can have them to compare to once the sentiment analysis is
complete.
sentiments <- select(tweets, tweet_id, entity, sentiment)
sentiments <- sentiments %>%
mutate(sentiment = tolower(sentiment))
tweets <- select(tweets, -sentiment, -entity)Lastly, we need to modify the data set so that it contains a single
row for each word present in the tweet (tokenization). This is done in
the cell below using the unnest_tokens function:
tweets <- unnest_tokens(tweets, words, tweet_content)
colnames(tweets) <- c('tweet_id', 'word')
head(tweets)## # A tibble: 6 × 2
## tweet_id word
## <dbl> <chr>
## 1 1 amazon
## 2 1 wtf
## 3 2 iâ
## 4 2 m
## 5 2 really
## 6 2 disappointedThe output above represents the final dataframe that we can now use to perform a sentiment analysis.
The lexicon that will be used for this example is the
afinn lexicon, which contains 2,477 words rated on a scale
from -5 to 5. In this case, the scale represents word’s
positivity/negativity, with higher scores indicating a more positive
sentiment. Some of these mappings can be seen below, which are stored in
the dataframe afinn:
afinn <- get_sentiments("afinn")
head(afinn)## # A tibble: 6 × 2
## word value
## <chr> <dbl>
## 1 abandon -2
## 2 abandoned -2
## 3 abandons -2
## 4 abducted -2
## 5 abduction -2
## 6 abductions -2The cell below joins the afinn data frame to our
tweets data set in an attempt to get a sentiment score for
each word.
tweets <- tweets %>%
left_join(afinn, by='word') %>%
mutate(value_clean = ifelse(is.na(value), 0, value))To determine the sentiment of the entire tweet, we can regroup our
tweets dataframe and sum the sentiment scores of each
word:
tweets <- tweets %>%
group_by(tweet_id) %>%
summarise(
score = sum(value_clean),
num_words = n(),
num_found_words = sum(ifelse(is.na(value) == FALSE, 1, 0))
)Lastly, given a sentiment score \(s\) and the number of scored words \(N\), we can categorize the tweet as negative, neutral, or positive via the following conditions:
This is performed in R in the cell below:
tweets <- tweets %>%
mutate(sentiment_pred =
case_when(
score / num_found_words < -1 ~ "negative",
score / num_found_words > 1 ~ "positive",
TRUE ~ "neutral"
)
)
head(tweets)## # A tibble: 6 × 5
## tweet_id score num_words num_found_words sentiment_pred
## <dbl> <dbl> <int> <dbl> <chr>
## 1 1 -4 2 1 negative
## 2 2 -4 44 2 negative
## 3 3 -2 30 2 neutral
## 4 4 -3 15 1 negative
## 5 6 1 3 2 neutral
## 6 7 0 11 0 neutralFinally, we can check these newly made predictions to the ones that initially came with the data set:
tweets <- tweets %>%
left_join(sentiments, by='tweet_id') %>%
mutate(correct = ifelse(sentiment_pred == sentiment, TRUE, FALSE))
table(tweets$correct)##
## FALSE TRUE
## 519 409Based on the output above, we can see that overall the sentiment analysis didn’t do too great, but did outperform chance: it was correct about 44% of the time. This number gets better when you remove the neutral tweets and redo our original scoring by simply using a sum to determine if a tweet had a positive or negative sentiment:
tweets <- tweets %>%
filter(sentiment != 'neutral') %>%
mutate(
correct_no_neutral = ifelse(
(score > 0 & sentiment == "positive") |
(score < 0 & sentiment == 'negative'), TRUE, FALSE)
)
table(tweets$correct_no_neutral)##
## FALSE TRUE
## 158 298When ignoring the neutral tweets, the sentiment analysis performed was correct a respectable 65% of the time.
Obviously this is not the most accurate sentiment analysis performed, and there are a number of improvements that could likely improve the final accuracy score. However, the preceding section provides a framework for how one might go about analyzing text in this way.