stansent wraps Stanford's coreNLP sentiment
tagger in a way that makes the
process easier to get set up. The output is designed to look and behave
like the objects from the
sentimentr package.
Plotting and the sentimentr::highlight functionality will work similar
to the sentiment/sentiment_by objects from sentimentr. This
requires less learning to work between the two packages.
In addition to sentimentr and stansent, Matthew Jocker's has created the syuzhet package that utilizes dictionary lookups for the Bing, NRC, and Afinn methods. Similarly, Subhasree Bose has contributed RSentiment which utilizes dictionary lookup that atempts to address negation and sarcasm. Click here for a comparison between stansent, sentimentr, syuzhet, and RSentiment. Note the accuracy and run times of the packages.
To download the development version of stansent:
Download the zip
ball or tar
ball, decompress
and run R CMD INSTALL on it, or use the pacman package to install
the development version:
if (!require("pacman")) install.packages("pacman")
pacman::p_load_gh("trinker/coreNLPsetup", "trinker/stansent")
After installing use the following to ensure Java and coreNLP are installed correctly:
check_setup()
to make sure your Java version is of the right version and coreNLP is set up in the right location.
There are two main functions in sentimentr with a few helper functions. The main functions, task category, & descriptions are summarized in the table below:
| Function | Function | Description |
|---|---|---|
sentiment_stanford |
sentiment | Sentiment at the sentence level |
sentiment_stanford_by |
sentiment | Aggregated sentiment by group(s) |
uncombine |
reshaping | Extract sentence level sentiment from sentiment_by |
get_sentences |
reshaping | Regex based string to sentence parser (or get sentences from sentiment/sentiment_by) |
highlight |
Highlight positive/negative sentences as an HTML document | |
check_setup |
initial set-up | Make sure Java and coreNLP are set up correctly |
You are welcome to:
- submit suggestions and bug-reports at: https://github.com/trinker/stansent/issues
- send a pull request on: https://github.com/trinker/stansent/
- compose a friendly e-mail to: tyler.rinker@gmail.com
if (!require("pacman")) install.packages("pacman")
pacman::p_load_gh(c("trinker/stansent", "trinker/sentimentr"))
pacman::p_load(dplyr)
mytext <- c(
'do you like it? But I hate really bad dogs',
'I am the best friend.',
'Do you really like it? I\'m not a fan'
)
data(presidential_debates_2012, cannon_reviews)
set.seed(100)
dat <- presidential_debates_2012[sample(1:nrow(presidential_debates_2012), 100), ]
out1 <- sentiment_stanford(mytext)
out1[["text"]] <- unlist(get_sentences(out1))
out1
## element_id sentence_id word_count sentiment text
## 1: 1 1 4 0.0 do you like it?
## 2: 1 2 6 -0.5 But I hate really bad dogs
## 3: 2 1 5 0.5 I am the best friend.
## 4: 3 1 5 0.0 Do you really like it?
## 5: 3 2 4 -0.5 I'm not a fan
To aggregate by element (column cell or vector element) use
sentiment_stanford_by with by = NULL.
out2 <- sentiment_stanford_by(mytext)
out2[["text"]] <- mytext
out2
## element_id word_count sd ave_sentiment
## 1: 1 10 0.3535534 -0.25
## 2: 2 5 NA 0.50
## 3: 3 9 0.3535534 -0.25
## text
## 1: do you like it? But I hate really bad dogs
## 2: I am the best friend.
## 3: Do you really like it? I'm not a fan
To aggregate by grouping variables use sentiment_by using the by
argument.
(out3 <- with(dat, sentiment_stanford_by(dialogue, list(person, time))))
## person time word_count sd ave_sentiment
## 1: OBAMA time 2 207 0.4042260 0.1493099
## 2: OBAMA time 1 34 0.7071068 0.0000000
## 3: LEHRER time 1 2 NA 0.0000000
## 4: QUESTION time 2 7 0.7071068 0.0000000
## 5: SCHIEFFER time 3 47 0.5000000 0.0000000
## 6: OBAMA time 3 129 0.4166667 -0.1393260
## 7: CROWLEY time 2 72 0.4166667 -0.1393260
## 8: ROMNEY time 3 321 0.3746794 -0.1508172
## 9: ROMNEY time 2 323 0.3875534 -0.2293311
## 10: ROMNEY time 1 95 0.2236068 -0.4138598
Note that the Stanford coreNLP functionality takes considerable time to
compute (~14.5 seconds to compute out above). The output from
sentiment_stanford/sentiment_stanford_by can be recycled inside of
sentiment_stanford_by, reusing the raw scoring to save the new call to
Java.
with(dat, sentiment_stanford_by(out3, list(role, time)))
## role time word_count sd ave_sentiment
## 1: candidate time 1 129 0.3933979 -0.29271628
## 2: candidate time 2 530 0.4154046 -0.06751165
## 3: candidate time 3 450 0.3796283 -0.15455530
## 4: moderator time 1 2 NA 0.00000000
## 5: moderator time 2 72 0.4166667 -0.13932602
## 6: moderator time 3 47 0.5000000 0.00000000
## 7: other time 2 7 0.7071068 0.00000000
The possible sentiment values in the output are {-1, -0.5, 0, 0.5, 1}. The raw number of occurrences as each sentiment level are plotted as a bubble version of Cleveland's dot plot. The red cross represents the mean sentiment score (grouping variables are ordered by this by default).
plot(out3)
The plot method for the class sentiment uses syuzhet's
get_transformed_values combined with ggplot2 to make a reasonable,
smoothed plot for the duration of the text based on percentage, allowing
for comparison between plots of different texts. This plot gives the
overall shape of the text's sentiment. The user can see
syuzhet::get_transformed_values for more details.
plot(uncombine(out3))
The user may wish to see the output from sentiment_stanford_by line by
line with positive/negative sentences highlighted. The
sentimentr::highlight function wraps a sentiment_by output to
produces a highlighted HTML file (positive = green; negative = pink).
Here we look at three random reviews from Hu and Liu's (2004) Cannon G3
Camera Amazon product reviews.
set.seed(2)
highlight(with(subset(cannon_reviews, number %in% sample(unique(number), 3)), sentiment_stanford_by(review, number)))
