Tidy text analysis

# Tidy text analysis
### Devin Judge-Lord

---

## 1. Counting
![](https://media.giphy.com/media/3o7TKu5aIDY4tU3SXm/giphy.gif)

- text features `$\in$`\{all* words, some words, phrases, etc.\}

---

## 1. Counting

- text features `$\in$`\{all* words, some words, phrases, etc.\}

## 2. Matching
![](Figs/same.jpg)

- Exactly* the same string ("regular expressions," text reuse)

---

## 1. Counting
- text features `$\in$`\{all* words, some words, phrases, etc.\}

## 2. Matching
- Exactly* the same string ("regular expressions," text reuse)

## 3. Classifying
![](Figs/different.jpg)

---

## 1. Counting
- Text features `$\in$`\{all* words, some words, phrases, etc.\}

## 2. Matching
- Exactly* the same string ("regular expressions," text reuse)
- "Fuzzy" matching is exact matching a set of variants

## 3. Classifying
- Rules vs. probability
- Supervised vs. unsupervised

---

## Resources 
- [Tidy text class by Andrew Heiss](https://datavizf18.classes.andrewheiss.com/class/11-class/)
- [`tidytext`](https://cran.r-project.org/web/packages/tidytext/tidytext.pdf) package
- Regular expressions with `stringr`--[cheasheet](http://edrub.in/CheatSheets/cheatSheetStringr.pdf)
- [Tidy text analysis](https://www.tidytextmining.com/) including
[topic modeling](https://www.tidytextmining.com/topicmodeling.html) and
[tidy() for Structural Topic Models](https://juliasilge.github.io/tidytext/reference/stm_tidiers.html) from the `stm` package. More [here](https://rdrr.io/cran/tidytext/man/stm_tidiers.html).

## Reading 
- [Introduction to cluster analysis](https://eight2late.wordpress.com/2015/07/22/a-gentle-introduction-to-cluster-analysis-using-r/)
- Tidy Natural Language Processing with [`cleanNLP`](https://scholarship.richmond.edu/cgi/viewcontent.cgi?article=1195&context=mathcs-faculty-publications)
- [Text from audio](https://www.cambridge.org/core/journals/political-analysis/article/testing-the-validity-of-automatic-speech-recognition-for-political-text-analysis/E375085D96331A47E810C01AA6DB0A46)
- Here is an [ok blog post on visualizing qualitative data](https://depictdatastudio.com/how-to-visualize-qualitative-data/)--lmk if you find a better resource.

---

## Cool applications:
- [Gender tropes in film](https://pudding.cool/2017/08/screen-direction/)
- [Analysis of Trump's tweets confirms he writes only the (angrier) Android half](http://varianceexplained.org/r/trump-tweets/)
- [Every time Ford and Kavanaugh dodged a question, in one chart](https://www.vox.com/policy-and-politics/2018/9/28/17914308/kavanaugh-ford-question-dodge-hearing-chart)

---

# 1 Counting things in fancy ways.

With tidy text, counting words or phrases is simple:

- `unnest_tokens()` splits each response into tokens (by word by default, but we can also tokenize by phrases of length n, called n-grams).

- [optional] `anti_join(stop_words)` removes words that often have little meaning, like "a" and "the", called stop words. We can also do this with with `filter(!(word %in% stop_words$word))`

- `count()` how many times each word appears (`count(word)` is like `group_by(word) %>% summarize(n = n()) %>% ungroup()` )

---

### Word frequency

Responses to ANES question "What do you dislike about [Democrats/Republicans]?" (V161101, V161106)

```r
load(here("data/ANESdislikes.Rdata"))
d <- ANESdislikes
d[1,]
```

```
## # A tibble: 1 x 3
##   V160001 question       response                                          
##     <int> <chr>          <chr>                                             
## 1  300001 dislike_about… i see them trying take to much and give to the ho…
```

Tokenize by word

```r
words <-  unnest_tokens(d, word, response)
head(words)
```

```
## # A tibble: 6 x 3
##   V160001 question          word  
##     <int> <chr>             <chr> 
## 1  300001 dislike_about_DEM i     
## 2  300001 dislike_about_DEM see   
## 3  300001 dislike_about_DEM them  
## 4  300001 dislike_about_DEM trying
## 5  300001 dislike_about_DEM take  
## 6  300001 dislike_about_DEM to
```

---

### Replace meaningless words

```r
# 1. anti-join() a data frame with a column of unwanted words called 'word'
words %<>% anti_join(stop_words)

words %<>% filter(!str_detect(word, unwanted_words) ) 
```

### Count

```r
## Count the number of times each word occurs in each group
words %<>% group_by(question) %>% count(word)

head(words)
```

```
## # A tibble: 6 x 3
## # Groups:   question [1]
##   question          word            n
##   <chr>             <chr>       <int>
## 1 dislike_about_DEM abandoning      1
## 2 dislike_about_DEM abaout          1
## 3 dislike_about_DEM abe             1
## 4 dislike_about_DEM abilities       1
## 5 dislike_about_DEM ability         5
## 6 dislike_about_DEM abomination     1
```

---

```r
top_n(words, 10) %>% ## Top 10 words in each group
  ggplot( aes(x = reorder(word, n), y = n) ) +
  geom_col() + 
  coord_flip() +
  facet_wrap("question", scales = "free_y", strip.position="top") + 
  labs(x = "Word", y = "Count")
```

---

```r
words %>%
  filter(question == "dislike_about_GOP") %>% 
  with(wordcloud(word, n, max.words = 50))
```

![GOP](Figs/ANES-wordcloud-2.png)

---

Word clouds only show word frequency, and font size is hard to compare visually. Nevertheless, [they may be useful](https://www.vis4.net/blog/2015/01/when-its-ok-to-use-word-clouds/) if all you care about is frequency.

---

Tokenize by word pair ("bi-gram")

```r
## Name the new column d$bigrams ("ngrams" where n = 2)
bigrams <-  unnest_tokens(d, bigram, response, token = "ngrams", n = 2) 
head(bigrams)
```

```
## # A tibble: 6 x 3
##   V160001 question          bigram     
##     <int> <chr>             <chr>      
## 1  300001 dislike_about_DEM i see      
## 2  300001 dislike_about_DEM see them   
## 3  300001 dislike_about_DEM them trying
## 4  300001 dislike_about_DEM trying take
## 5  300001 dislike_about_DEM take to    
## 6  300001 dislike_about_DEM to much
```

```r
## Count the number of times each bigram occurs in each group
bigrams %<>%
  group_by(question) %>%
  count(bigram) 
```

---

```r
## Top 10 bigrams in each group
top_n(bigrams, 10) %>% 
  ggplot( aes(x = reorder(bigram, n), y = n) ) +
  geom_col() + 
  coord_flip() +
  facet_wrap("question", scales = "free_y", strip.position="top") + 
  labs(x = "Bigram", y = "Count")
```

---

# 2 Matching

## Global alignment 
(more counting than matching)

Used in: overall document similarity

## Local alignment 
- Exact (`==`)
- Patterns (Regular expressions,  `stringr`)
- Fuzzy patterns

Used in: text-reuse/plagiarism, web scraping, data cleaning

---

## Resources

Matching usually involve doing the same thing lots of times, so we use [functions](https://nicercode.github.io/guides/functions/)
- [`purrr`](https://www.rstudio.com/resources/cheatsheets/#purrr) offers tools to apply functions to data (instead of writing loops where we have to index everything)

Web scraping also requires manipulating HTML. 
- [Khan Academy intro to HTML.](https://www.khanacademy.org/computing/computer-programming/html-css/intro-to-html/pt/html-basics)
- `rvest` offers tools for web scraping. Here is a [tutorial on Data Camp](https://www.datacamp.com/community/tutorials/r-web-scraping-rvest0)

Once we have scraped text, we need to clean it and extract the things we care about, both task for [RegEx]( https://www.rstudio.com/wp-content/uploads/2016/09/RegExCheatsheet.pdf)
- For tools to manipulate text, see the [stringr cheatsheet](https://www.rstudio.com/resources/cheatsheets/#stringr)
- Matt Denny has a [nice tutorial on cleaning text](http://www.mjdenny.com/Text_Processing_In_R.html)

---

## Global alignment

How similar are pairs of texts are overall (i.e., statistically)

1. Count things, usually word frequencies.
    - Recall the [congressional letters example](https://judgelord.github.io/PS811/text-legislator-letters.html)
    - The result is a high-dimension object. (If each word is a variable, we may have thousands of variables!)
2. Compare distributions
    - Cosign similarity scores (pairwise similarity score)
    - Dimensional scaling (reduce variation to one or two spatial dimensions)
    
---

---

## Local alignment

![](Figs/different.jpg) ![](Figs/things.jpg)

---

#### Exact (`==`) vs. patterns (`str_detect("regular expression")`)

```r
strings <- c("Elmo", "Zoe", "Thing 1", "Thing 2", 
             "are your politics more like Sesame Street or Dr Seuss?")

strings == "Elmo" # Exact match?
```

```
## [1]  TRUE FALSE FALSE FALSE FALSE
```

```r
strings == "Thing [0-9]" # Exact match?
```

```
## [1] FALSE FALSE FALSE FALSE FALSE
```

```r
str_detect(strings, "Thing [0-9]") # RegEx match?
```

```
## [1] FALSE FALSE  TRUE  TRUE FALSE
```

```r
str_extract_all(strings, "[A-Z][a-z]*", simplify = T) # stringr functions: str_....
```

```
##      [,1]     [,2]     [,3] [,4]   
## [1,] "Elmo"   ""       ""   ""     
## [2,] "Zoe"    ""       ""   ""     
## [3,] "Thing"  ""       ""   ""     
## [4,] "Thing"  ""       ""   ""     
## [5,] "Sesame" "Street" "Dr" "Seuss"
```

---

---

# Web scraping

`rvest`: `read_html()`, or

`rvest` + `httr`: `read_html( httr::GET() )`

99% of web scraping is finding the information we want and wrangling it into a data frame.

```r
html <- read_html("https://www.UN.org/") # The UN homepage
links <- html_nodes(html, "a") # "a" nodes are linked text
html_text(links)[2:7] # The text for links 2-7
```

```
## [1] "国际提高地雷意识和协助地雷行动日－4月4日"                                                                                             
## [2] "International Mine Awareness Day - 4 April"                                                                                           
## [3] "Journée internationale pour la sensibilisation au problème des mines et l’assistance à la lutte antimines - 4 avril"                  
## [4] "Международный день просвещения по вопросам минной опасности и помощи в деятельности, связанной с разминированием — 4 апреля"          
## [5] "Día Internacional de información sobre el peligro de las minas y de asistencia para las actividades relativas a las minas, 4 de abril"
## [6] "اليوم الدولي للتوعية بخطر الألغام\r\n\r\n4 نيسان/أبريل\r\n"
```

[Example: A list of companies lobbying the U.S. Federal Energy Regulatory Commission](https://judgelord.github.io/correspondence/functions/DOE_FERC-company-scraper.html)

[Example: A table including the text of scanned pdf letters from Members of Congress](https://judgelord.github.io/correspondence/functions/DOE_FERC-scraper.html)

---

## Text-reuse/plagerism

---

## Measuring similarity:
- Percent copied words, n-grams, sentences
- Percent aligned
- Global alignment (word frequencies) or relative topic proportions

## Measuring change across versions:
- Percent new words, n-grams, sentences
- Percent unaligned
- Change in global alignment (word frequencies) or topic proportions

---

### For example, changes in budget text and committee reports [[see more]()]:

![](Figs/reports.png)

---
### When similar words mean different things, edits matter:

2010: `"CLIMATE CHANGE The Committee continues to support the Administration's efforts to address climate change. Funding for its voluntary climate change programs are continued through this bill."`

2012: `"CLIMATE CHANGE This Committee remains skeptical of the Administration's efforts to re-package existing programs and to fund new ones in the name of climate change."`

---

### Edits may introduce new ideas in otherwise aligned texts.

The Smith-Waterman local alignment algorithm:

2017: `"FDA is implementing the Food Safety Modernization Act FSMA `and the  XXXXXXXXX motivation XXXXXX for XXXXX the generic drug labeling rule `and regulation of tobacco products"`

2018: `"FDA is implementing the Food Safety Modernization Act FSMA `XXX XXX addressing XXXXXXXXX opioid XXX abuse XXX XXXXXXX XXXX XXXXXXXX XXXX `and regulation of tobacco products"`

---

## [Smith-Waterman](http://fridolin-linder.com/resources/2016/03/30/local-alignment.html)
Tracing a path through the matrix of the two documents, maximizing matches (technically, points for sequences of matches and mismatches).

---

## N-gram window
Finding all exact matches of n words or more.

With a 4-gram window:

2010: "`CLIMATE CHANGE The Committee` continues to support `the Administration's efforts to` address climate change. Funding for its voluntary climate change programs are continued through this bill."`

2012: "`CLIMATE CHANGE This Committee` remains skeptical of `the Administration's efforts to` re-package existing programs and to fund new ones in the name of climate change."`

From this example: [Change and influence in budget texts]()

Another example: [Public comments on proposed regulations]()

---

![](https://media.giphy.com/media/14kkd3gt5FJt3a/giphy.gif)