sentimentr

sentimentr is designed to quickly calculate text polarity sentiment at the sentence level and optionally aggregate by rows or grouping variable(s).

sentimentr is a response to my own needs with sentiment detection that were not addressed by the current R tools. My own polarity function in the qdap package is slower on larger data sets. It is a dictionary lookup approach that tries to incorporate weighting for valence shifters (negation and amplifiers/deamplifiers). Matthew Jocker's created the syuzhet package that utilizes dictionary lookups for the Bing, NRC, and Afinn methods. He also utilizes a wrapper for the Stanford coreNLP which uses much more sophisticated analysis. Jocker's dictionary methods are fast but are more prone to error in the case of valence shifters. Jocker's addressed these critiques explaining that the method is good with regard to analyzing general sentiment in a piece of literature. He points to the accuracy of the Stanford detection as well. In my own work I need better accuracy than a simple dictionary lookup; something that considers valence shifters yet optimizes speed which the Stanford's parser does not. This leads to a trade off of speed vs. accuracy. The equation below describes the dictionary method of sentimentr that may give better results than a dictionary approach that does not consider valence shifters but will likely still be less accurate than Stanford's approach. Simply, sentimentr attempts to balance accuracy and speed.

Table of Contents

• [Functions](#functions)
• [The Equation](#the-equation)
• [Installation](#installation)
• [Examples](#examples)
  • [Plotting](#plotting)
    • [Plotting at Aggregated Sentiment](#plotting-at-aggregated-sentiment)
    • [Plotting at the Sentence Level](#plotting-at-the-sentence-level)
  • [Annie Swafford's Examples](#annie-swafford's-examples)
  • [Comparing sentimentr, syuzhet, and Stanford](#comparing-sentimentr-syuzhet-and-stanford)
  • [Text Highlighting](#text-highlighting)
• [Contact](#contact)

Functions

There are two main functions (top 2 in table below) in sentimentr with several helper functions summarized in the table below:

Function	Description
sentiment	Sentiment at the sentence level
sentiment_by	Aggregated sentiment by group(s)
uncombine	Extract sentence level sentiment from sentiment_by
get_sentences	Regex based string to sentence parser (or get sentences from sentiment/sentiment_by)
replace_emoticon	Replace emoticons with word equivalent
replace_grade	Replace gradess (e.g., "A+") with word equivalent
replace_rating	Replace ratings (e.g., "10 out of 10", "3 stars") with word equivalent
as_key	Coerce a data.frame lexicon to a polarity hash key
is_key	Check if an object is a hash key
update_key	Add/remove terms to/from a hash key
highlight	Highlight positive/negative sentences as an HTML document

The Equation

The equation used by the algorithm to assign value to polarity of each sentence fist utilizes the sentiment dictionary (Hu and Liu, 2004) to tag polarized words. Each paragraph (p_i = {s₁, s₂, ..., s_n}) composed of sentences, is broken into element sentences (s_i, j = {w₁, w₂, ..., w_n}) where w are the words within sentences. Each sentence (s_j) is broken into a an ordered bag of words. Punctuation is removed with the exception of pause punctuations (commas, colons, semicolons) which are considered a word within the sentence. I will denote pause words as c**w (comma words) for convenience. We can represent these words as an i,j,k notation as w_i, j, k. For example w_3, 2, 5 would be the fifth word of the second sentence of the third paragraph. While I use the term paragraph this merely represent a complete turn of talk. For example it may be a cell level response in a questionnaire composed of sentences.

The words in each sentence (w_i, j, k) are searched and compared to a modified version of Hu, M., & Liu, B.'s (2004) dictionary of polarized words. Positive (w_i, j, k⁺) and negative (w_i, j, k⁻) words are tagged with a +1 and −1 respectively (or other positive/negative weighting if the user provides the sentiment dictionary). I will denote polarized words as p**w for convenience. These will form a polar cluster (c_i, j, l) which is a subset of the a sentence (c_i, j, l ⊆ s_i, j).

The polarized context cluster (c_i, j, l) of words is pulled from around the polarized word (p**w) and defaults to 4 words before and two words after p**w to be considered as valence shifters. The cluster can be represented as (c_i, j, l = {p**w_{i, j, k − n**b}, ..., p**w_i, j, k, ..., p**w_{i, j, k − n**a}}), where n**b & n**a are the parameters n.before and n.after set by the user. The words in this polarized context cluster are tagged as neutral (w_i, j, k⁰), negator (w_i, j, kⁿ), amplifier (w_i, j, k^a), or de-amplifier (w_i, j, k^d). Neutral words hold no value in the equation but do affect word count (n). Each polarized word is then weighted (w) based on the weights from the polarity_dt argument and then further weighted by the function and number of the valence shifters directly surrounding the positive or negative word (p**w). Pause (c**w) locations (punctuation that denotes a pause including commas, colons, and semicolons) are indexed and considered in calculating the upper and lower bounds in the polarized context cluster. This is because these marks indicate a change in thought and words prior are not necessarily connected with words after these punctuation marks. The lower bound of the polarized context cluster is constrained to max{p**w_{i, j, k − n**b}, 1, max{c**w_i, j, k < p**w_i, j, k}} and the upper bound is constrained to min{p**w_{i, j, k + n**a}, w_i, j**n, min{c**w_i, j, k > p**w_i, j, k}} where w_i, j**n is the number of words in the sentence.

The core value in the cluster, the polarized word is acted upon by valence shifters. Amplifiers increase the polarity by 1.8 (.8 is the default weight (z)). Amplifiers (w_i, j, k^a) become de-amplifiers if the context cluster contains an odd number of negators (w_i, j, kⁿ). De-amplifiers work to decrease the polarity. Negation (w_i, j, kⁿ) acts on amplifiers/de-amplifiers as discussed but also flip the sign of the polarized word. Negation is determined by raising −1 to the power of the number of negators (w_i, j, kⁿ) plus 2. Simply, this is a result of a belief that two negatives equal a positive, 3 negatives a negative, and so on.

The "but" conjunctions (i.e., 'but', 'however', and 'although') also weight the context cluster. A but conjunction before the polarized word (w_{but conjunction}, ..., w_i, j, k^p) up-weights the cluster by 1 + z₂ * {|w_{but conjunction}|,...,w_i, j, k^p} (.85 is the default weight (z₂) where |w_{but conjunction}| are the number of but conjunctions before the polarized word). A but conjunction after the polarized word down-weights the cluster by 1 + {w_i, j, k^p, ..., |w_{but conjunction}|* − 1}*z₂. This corresponds to the belief that a but makes the next clause of greater values while lowering the value placed on the prior clause.

The researcher may provide a weight (z) to be utilized with amplifiers/de-amplifiers (default is .8; de-amplifier weight is constrained to −1 lower bound). Last, these weighted context clusters (c_i, j, l) are summed (c′_i, j) and divided by the square root of the word count (√w_i, j**n) yielding an unbounded polarity score (δ_i, j) for each sentence.

δ_i**j = c'_i**j/√w_ijn

Where:

c′_i, j = ∑((1 + w_amp + w_deamp)⋅w_i, j, k^p(−1)^{2 + w_neg})

w_amp = ∑(w_neg ⋅ (z ⋅ w_i, j, k^a))

w_deamp = max(w_deamp′, −1)

w_deamp′ = ∑(z(−w_neg ⋅ w_i, j, k^a + w_i, j, k^d))

w_b = 1 + z₂ * w_b′

w_b′ = ∑(|w_{but conjunction}|,...,w_i, j, k^p, w_i, j, k^p, ..., |w_{but conjunction}|* − 1)

w_neg = (∑w_i, j, kⁿ ) mod 2

To get the mean of all sentences (s_i, j) within a paragraph (p_i) simply take the average sentiment score p_{i, δi, j} = 1/n ⋅ ∑ δ_i, j.

Installation

To download the development version of sentimentr:

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/sentimentr")

Examples

if (!require("pacman")) install.packages("pacman")
pacman::p_load(sentimentr)

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'
)
sentiment(mytext)

##    element_id sentence_id word_count  sentiment
## 1:          1           1          4  0.5000000
## 2:          1           2          6 -2.6781088
## 3:          2           1          5  0.4472136
## 4:          3           1          5  0.8049845
## 5:          3           2          4  0.0000000

To aggregate by element (column cell or vector element) use sentiment_by with by = NULL.

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'
)
sentiment_by(mytext)

##    element_id word_count       sd ave_sentiment
## 1:          1         10 2.247262    -1.0890544
## 2:          2          5       NA     0.4472136
## 3:          3          9 0.569210     0.4024922

To aggregate by grouping variables use sentiment_by using the by argument.

(out <- with(presidential_debates_2012, sentiment_by(dialogue, list(person, time))))

##        person   time word_count        sd ave_sentiment
##  1:    LEHRER time 1        765 0.3474359    0.10958889
##  2:     OBAMA time 1       3598 0.4406033    0.10915755
##  3:     OBAMA time 3       7241 0.4121806    0.09922089
##  4:     OBAMA time 2       7476 0.3829126    0.09089932
##  5:    ROMNEY time 3       8302 0.3883042    0.07834848
##  6: SCHIEFFER time 3       1445 0.3773859    0.06669193
##  7:    ROMNEY time 1       4085 0.3550998    0.06391677
##  8:   CROWLEY time 2       1672 0.2144920    0.05560477
##  9:    ROMNEY time 2       7534 0.3246025    0.04589893
## 10:  QUESTION time 2        583 0.3255268    0.03334828

Plotting

Plotting at Aggregated Sentiment

plot(out)

Plotting at the Sentence Level

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(out))

Annie Swafford's Examples

Annie Swafford critiqued Jocker's approach to sentiment and gave the following examples of sentences (ase for Annie Swafford example). Here I test each of Jocker's 3 dictionary approaches (Bing, NRC, Afinn), his Stanford wrapper (note I use my own GitHub Stanford wrapper package based off of Jocker's approach as it works more reliably on my own Windows machine), and my own algorithm with both the default Hu & Liu (2004) polarity lexicon as well as Baccianella, Esuli and Sebastiani's (2010) SentiWord lexicon.

if (!require("pacman")) install.packages("pacman")
pacman::p_load_gh("trinker/sentimentr", "trinker/stansent")
pacman::p_load(syuzhet, qdap, microbenchmark)

ase <- c(
    "I haven't been sad in a long time.",
    "I am extremely happy today.",
    "It's a good day.",
    "But suddenly I'm only a little bit happy.",
    "Then I'm not happy at all.",
    "In fact, I am now the least happy person on the planet.",
    "There is no happiness left in me.",
    "Wait, it's returned!",
    "I don't feel so bad after all!"
)

syuzhet <- setNames(as.data.frame(lapply(c("bing", "afinn", "nrc"),
    function(x) get_sentiment(ase, method=x))), c("bing", "afinn", "nrc"))


left_just(data.frame(
    stanford = sentiment_stanford(ase)[["sentiment"]],
    hu_liu = round(sentiment(ase, question.weight = 0)[["sentiment"]], 2),
    sentiword = round(sentiment(ase, sentiword, question.weight = 0)[["sentiment"]], 2),    
    syuzhet,
    sentences = ase,
    stringsAsFactors = FALSE
), "sentences")

  stanford hu_liu sentiword bing afinn nrc
1     -0.5   0.35      0.18   -1    -2   0
2        1    0.8      0.65    1     3   1
3      0.5    0.5      0.32    1     3   1
4     -0.5      0         0    1     3   1
5     -0.5  -0.41     -0.56    1     3   1
6     -0.5   0.06      0.05    1     3   1
7     -0.5  -0.38     -0.05    1     2   1
8        0      0     -0.14    0     0  -1
9     -0.5   0.38      0.24   -1    -3  -1
  sentences                                              
1 I haven't been sad in a long time.                     
2 I am extremely happy today.                            
3 It's a good day.                                       
4 But suddenly I'm only a little bit happy.              
5 Then I'm not happy at all.                             
6 In fact, I am now the least happy person on the planet.
7 There is no happiness left in me.                      
8 Wait, it's returned!                                   
9 I don't feel so bad after all!                         

Also of interest is the computational time used by each of these methods. To demonstrate this I increased Annie's examples by 100 replications and microbenchmark on a few iterations (Stanford takes so long I didn't extend to more). Note that if a text needs to be broken into sentence parts syuzhet has the get_sentences function that uses the openNLP package, this is a time expensive task. sentimentr uses a much faster regex based approach that is nearly as accurate in parsing sentences with a much lower computational time. We see that Stanford takes the longest time while sentimentr and syuzhet are comparable depending upon lexicon used.

ase_100 <- rep(ase, 100)

stanford <- function() {sentiment_stanford(ase_100)}

sentimentr_hu_liu <- function() sentiment(ase_100)
sentimentr_sentiword <- function() sentiment(ase_100, sentiword) 
    
syuzhet_binn <- function() get_sentiment(ase_100, method="bing")
syuzhet_nrc <- function() get_sentiment(ase_100, method="nrc")
syuzhet_afinn <- function() get_sentiment(ase_100, method="afinn")
     
microbenchmark(
    stanford(),
    sentimentr_hu_liu(),
    sentimentr_sentiword(),
    syuzhet_binn(), 
    syuzhet_nrc(),
    syuzhet_afinn(),
    times = 3
)

Unit: milliseconds
                   expr        min         lq       mean     median
             stanford() 24007.9133 25026.5316 25501.0191 26045.1499
    sentimentr_hu_liu()   257.8798   260.6083   292.9230   263.3368
 sentimentr_sentiword()   977.4298   988.2982  1009.3642   999.1666
         syuzhet_binn()   374.4916   386.0058   389.8711   397.5200
          syuzhet_nrc()   861.2786   877.5804   923.3852   893.8822
        syuzhet_afinn()   159.7357   160.2946   168.4575   160.8536
         uq        max neval
 26247.5720 26449.9941     3
   310.4446   357.5525     3
  1025.3314  1051.4962     3
   397.5609   397.6017     3
   954.4385  1014.9948     3
   172.8184   184.7832     3

Comparing sentimentr, syuzhet, and Stanford

The accuracy of an algorithm weighs heavily into the decision as to what approach to take in sentiment detection. Both syuzhet and sentimentr provide multiple dictionaries with a general algorithm to compute sentiment scores. syuzhet provides 3 approaches while sentimentr provides 2, but can be extended easily using the 3 dictionaries from the syuzhet package. The follow visualization provides the accuracy of these approaches in comparison to Stanford's Java based implementation of sentiment detection. The visualization is generated from testing on three reviews data sets from Kotzias, Denil, De Freitas, & Smyth (2015). These authors utilized the three 1000 element data sets from:

• amazon.com
• imdb.com
• yelp.com

The data sets are hand scored as either positive or negative. The testing here merely matches the sign of the algorithm to the human coded output to determine accuracy rates.

• Kotzias, D., Denil, M., De Freitas, N., & Smyth,P. (2015). From group to individual labels using deep features. Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 597-606. http://mdenil.com/media/papers/2015-deep-multi-instance-learning.pdf

The bar graph on the left shows the accuracy rates for the various sentiment set-ups in the three review contexts. The rank plot on the right shows how the rankings for the methods varied across the three review contexts (note that rank ties were determined by ties.method = "first").

The take away here seems that, unsurprisingly, Stanford's algorithm consistently out scores syuzhet and sentimentr. The sentimentr approach loaded with the hu_lu dictionary is a top pick for speed and accuracy. The bing dictionary also performs well within both the syuzhet and sentimentr algorithms. Generally, the sentimentr algorithm out performs syuzhet when their dictonaries are comparable.

It is important to point out that this is a small sample data set that covers a narrow range of uses for sentiment detection. Jocker's syuzhet was designed to be applied across book chunks and it is, to some extent, unfair to test it out of this context. Still this initial analysis provides a guide that may be of use for selecting the sentiment detection set up most applicable to the reader's needs.

The reader may access the R script used to generate this visual via:

testing <- system.file("sentiment_testing/sentiment_testing.R", package = "sentimentr")
file.copy(testing, getwd())

In the figure below we compare raw table counts as a heat map, plotting the predicted values from the various algorithms on the x axis versus the human scored values on the y axis.

Across all three contexts, notice that the Stanford coreNLP algorithm is better at:

• Detecting negative sentiment as negative
• Discrimination (i.e., reducing neutral assignments)

The Bing, Hu & Lu, and Afinn dictionaries all do well with regard to not assigning negative scores to positive statements, but perform less well in the reverse, often assigning positive scores to negative statements. We can now see that the reason for the NRC's poorer performance in accuracy rate above is its inability to discriminate. The Sentiword dictionary does well at discriminating (like Stanford's coreNLP) but does not perform well. We can deduce to things from this observation:

1. Larger dictionaries discriminate better (Sentiword [n = 20,102] vs. Hu & Lu [n = 6,827])
2. The Sentiword dictionary may have words with reversed polarities

A reworking of the Sentiword dictionary may yield better results for a dictionary lookup approach to sentiment detection, potentially, improving on discrimination and accuracy.

The reader may access the R script used to generate this visual via:

testing2 <- system.file("sentiment_testing/raw_results.R", package = "sentimentr")
file.copy(testing2, getwd())

Text Highlighting

The user may wish to see the output from sentiment_by line by line with positive/negative sentences highlighted. The 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_by(review, number)))

Contact

You are welcome to:

• submit suggestions and bug-reports at: https://github.com/trinker/sentimentr/issues
• send a pull request on: https://github.com/trinker/sentimentr/
• compose a friendly e-mail to: [email protected]