Generally, lectures will be formatted with:
Data for this course will come in many forms, as language is inherently unstructured. We will mostly use tidy
format as defined:
However, we might define an observation as a frequency or proportion of observations or a specific word (rather than a participant or person in a study), etc. Learning how to structure the data for our analyses will be part of the goal of each lecture.
Tokens - meaningful unit of text, often a word, but could be phrases, documents, sentences, etc. Thus, to keep our data in tidy format, we can use one token per row, to treat each token as an observation. Later in the semester, we will use term-by-document matrices and corpus objects, which will be formatted differently.
The tidytext
package has many tools that we can use to help us analyze text information. Let’s load it and try sentiment analysis with the package.
library(tidytext)
When we read, we use our understanding of words to help determine meaning. Often the semanticity of words includes their emotional intent. Using information about the valence of words, we can determine if a text is positive or negative (or other emotional descriptors).
Run the code below to see the graphic. Make sure you’ve downloaded the picture and put it in the same folder as this assignment.
The graphic below shows how you might treat a research workflow using tidyverse
to analyze sentiment.
knitr::include_graphics("tidyflow-ch-2.png")
We are going to examine sentiment as a “sum of parts” - this approach means that we can look sum up the sentiments of individual words to represent the larger text.
head(sentiments)
## # A tibble: 6 x 4
## word sentiment lexicon score
## <chr> <chr> <chr> <int>
## 1 abacus trust nrc NA
## 2 abandon fear nrc NA
## 3 abandon negative nrc NA
## 4 abandon sadness nrc NA
## 5 abandoned anger nrc NA
## 6 abandoned fear nrc NA
The sentiments
dataset as part of tidytext
includes three sentiment lexicons: AFINN, bing, and nrc. There are several others we can use including one by Warriner et al., but these provide good coverage of common English words.
The dataset includes:
get_sentiments
as a quick subset function to only grab one dataset depending on what you are interested in.The limitation to these datasets is that we have to remember when and how they were validated. One thing we will discuss this semester is the fact the word meanings change over time, so we have to consider time period for each analysis.
Another limitation to this approach is that context is ignored (sometimes this approach is consider “bag-of-words” because words are just tossed into a bag and totalled up). Qualifiers like “no” and “aren’t” are not considered - additionally, sarcasm and idioms will not be captured.
For this analysis, we are going to explore Jane Austen novels. You will want to change the parameters of the analysis while exploring the functionality of the code. You should fill in the information where requested - look for instructions in ALL CAPS.
#load the libraries
library(janeaustenr)
library(dplyr)
##
## Attaching package: 'dplyr'
## The following objects are masked from 'package:stats':
##
## filter, lag
## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, union
library(stringr)
#specific to this package, pull jane austen books and create a tidy dataframe
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)
Specifically, here check out the unnest_tokens
function - word is the output column, while text is the input column.
WHAT DOES IT APPEAR THAT THE UNNEST_TOKENS FUNCTION DID? TRY RUNNING THE CODE WITH AND WITHOUT THE LAST LINE.
Answer: The UNNEST_TOKENS function split a column into tokens using the tokenizers package, splitting the table into one-token-per-row.
The code provided analyzes the “joy” sentiment in “Emma”.
nrc_joy <- get_sentiments("nrc") %>%
filter(sentiment == "joy")
tidy_books %>%
filter(book == "Mansfield Park") %>%
inner_join(nrc_joy ) %>% #this function merges sentiment with the Emma data
count(word, sort = TRUE) #makes a frequency table
## Joining, by = "word"
## # A tibble: 329 x 2
## word n
## <chr> <int>
## 1 good 326
## 2 young 153
## 3 hope 138
## 4 grant 134
## 5 love 124
## 6 happy 117
## 7 found 113
## 8 friend 98
## 9 mother 89
## 10 happiness 86
## # … with 319 more rows
EDIT THE CODE TO USE A DIFFERENT EMOTION AND NOVEL.
Answer: The code is edited to use Mansfield Park book and Anger.
WHAT ARE THE TOP WORDS IN YOUR EMOTION AND NOVEL?
Answer: The top words in Mansfield Park and Anger emotion is “feeling”, “ill”, “bad”, “words”, “evil”.
DO THERE APPEAR TO BE SOME WORDS THAT ARE SURPRISING TO YOU? (I.E. THEY DO NOT SEEM TO MATCH WHAT YOU MIGHT EXPECT TO FIND AS FREQUENT FOR THAT EMOTION)
Answer: The two words “feeling” and “words” do not sound like to represent anger emotion because they do not have any specific feeling associated with them.
We should consider the size of text to analyze for sentiment. If we use a whole document, the effects of sections of sentiment (like one sad chapter) may get washed out. However, you may not want to use small sentences because you might miss the larger structure of the text. The suggestion from the book is to use ~ 80 lines of text, and she is pretty smart, so let’s try that.
library(tidyr)
jane_austen_sentiment <- tidy_books %>%
inner_join(get_sentiments("bing")) %>%
count(book, index = linenumber %/% 120, sentiment) %>%
spread(sentiment, n, fill = 0) %>%
mutate(sentiment = positive - negative)
## Joining, by = "word"
WHAT DID THIS CODE APPEAR TO CREATE FOR US?
Answer: The code creats for us the negative score, postivie score, and overall sentiment score.
We can use ggplot2
to plot the sentiment across the predefined chunks of text. This plot is similar to a lexical dispersion plot, which shows the instances of a word across a text.
library(ggplot2)
## Registered S3 methods overwritten by 'ggplot2':
## method from
## [.quosures rlang
## c.quosures rlang
## print.quosures rlang
ggplot(jane_austen_sentiment, aes(index, sentiment, fill = book)) +
geom_col(show.legend = FALSE) +
facet_wrap(~book, ncol = 2, scales = "free_x") +
theme_bw()
EXAMINE THE GRAPH - WHAT BOOK APPEARS TO HAVE THE MOST POSITIVE INTERPRETATION? THE MOST NEGATIVE?
Answer: Emma appears to be the most positive book; Mansfield Park appears to be the most negative book.
CHANGE THE NUMBER OF LINES TO SOMETHING SMALLER LIKE 10-20 OR MUCH LARGER LIKE 200 - RERUN THE CODE AND GRAPH. WHAT CHANGES DO YOU SEE? Answer: Changing to smaller lines of 15: smaller sentences are created, and the sentiment among books seems insignificant. Changing to bigger lines of 120: bigger sentences are created, and there seems less data points for us to tell differences among books.
The choice in lexicon might be based on word overlap (i.e. it has the words you need) or based on what you want to analyze. Because we have more than one, we can compare them directly.
pride_prejudice <- tidy_books %>%
filter(book == "Emma")
head(pride_prejudice)
## # A tibble: 6 x 4
## book linenumber chapter word
## <fct> <int> <int> <chr>
## 1 Emma 1 0 emma
## 2 Emma 3 0 by
## 3 Emma 3 0 jane
## 4 Emma 3 0 austen
## 5 Emma 8 0 volume
## 6 Emma 8 0 i
THIS EXAMPLE IS FOR PRIDE & PREJUDICE. CHANGE THE CODE HERE TO USE A DIFFERENT BOOK.
Answer: The code is changed to use Emma to do analysis.
This code pulls each sentiment and merges it with the book chosen above. The plot at the end compares each of the methods. Here it is most appropriate to use the positive and negative categories from NRC to match the bing and AFINN datasets. Otherwise we might not be comparing the same ideas.
afinn <- pride_prejudice %>%
inner_join(get_sentiments("afinn")) %>%
group_by(index = linenumber %/% 80) %>%
summarise(sentiment = sum(score)) %>%
mutate(method = "AFINN")
## Joining, by = "word"
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) %>%
spread(sentiment, n, fill = 0) %>%
mutate(sentiment = positive - negative)
## Joining, by = "word"
## Joining, by = "word"
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")
EXAMINING YOUR BOOK - DO THE THREE SOURCES APPEAR TO AGREE? WHAT ARE THE MAJOR DIFFERENCES OR SIMILARITIES?
Answer: The three books do not appear to agree. The major difference is that NRC appears to be more positive than the analysis by AFINN/Bing.
Let’s figure out the most common positive and negative words across all of Jane Austen’s texts.
bing_word_counts <- tidy_books %>%
filter(book == "Emma") %>%
inner_join(get_sentiments("bing")) %>%
count(word, sentiment, sort = TRUE) %>%
ungroup()
## Joining, by = "word"
head(bing_word_counts)
## # A tibble: 6 x 3
## word sentiment n
## <chr> <chr> <int>
## 1 miss negative 599
## 2 well positive 401
## 3 good positive 359
## 4 great positive 264
## 5 like positive 200
## 6 better positive 173
LOOK AT THE TOP WORDS HERE. WHY MIGHT A FEW OF THESE BE PROBLEMATIC/MISINTERPRETED? THINK ABOUT THE STYLE OF WRITING FOR THESE NOVELS.
Answer: The sentiment of miss looks problematic because “miss” appears to be a positive emotion. Because when Jane is writing the books, miss represent people rather than feelings.
Let’s plot that analysis to easier viewing:
bing_word_counts %>%
group_by(sentiment) %>%
top_n(10) %>%
ungroup() %>%
mutate(word = reorder(word, n)) %>%
ggplot(aes(word, n, fill = sentiment)) +
geom_col(show.legend = FALSE) +
facet_wrap(~sentiment, scales = "free_y") +
labs(y = "Contribution to sentiment",
x = NULL) +
coord_flip() +
theme_bw()
## Selecting by n
CHANGE THE CODE ABOVE TO REFLECT ONLY ONE OF THE NOVELS IN THE DATASET. WHAT DO YOU OBSERVE ABOUT THE MOST USED POSITIVE AND NEGATIVE WORDS?
Answer: the book is changed to Emma book The most used postive word is well; The most used negative word is miss.
Word clouds are a popular visualization tool for text analysis. We can use the wordcloud
library to help us create those plots. This analysis ignores stop words which are common words that appear a lot like “the, an, of, into”.
library(wordcloud)
## Loading required package: RColorBrewer
tidy_books %>%
filter(book == "Emma") %>%
anti_join(stop_words) %>%
count(word) %>%
with(wordcloud(word, n, max.words = 50))
## Joining, by = "word"
It might be more interesting though to compare positive versus negative in the same plot:
library(reshape2)
##
## Attaching package: 'reshape2'
## The following object is masked from 'package:tidyr':
##
## smiths
tidy_books %>%
filter(book == "Emma") %>%
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 = 50)
## Joining, by = "word"
EDIT THE ABOVE CODE TO ONLY INCLUDE ONE OF THE BOOKS. WHAT DO YOU FIND TO BE THE MOST POSITIVE AND NEGATIVE WORDS IN YOUR BOOK (SHOULD MATCH ABOVE).
Answer: Same as the answer above, the book is changed to Emma only, and I observed that the most positive word is well, and the most negative word is miss.
You now have the skills to explore a set of text for positive and negative sentiment! You can apply these ideas to many types of text. In a future session, we will explore Twitter word usage and sentiment.
To turn in this assignment, hit KNIT at the top. You will submit the report in html/pdf/word format (default is html) on Moodle for credit. Be sure you have answered the questions. Great job!