index.Rmd

---
title: "Text Analysis in R"
subtitle: "Special Christmas Edition"
author: "Emil Hvitfeldt"
date: "2018-10-29"
output:
  xaringan::moon_reader:
    css: ["default", "theme.css"]
    lib_dir: libs
    nature:
      beforeInit: "macros.js"
      highlightStyle: github
      highlightLines: true
      countIncrementalSlides: false
      titleSlideClass: [center, middle]
---

```{r include=FALSE}
library(knitr)
hook_output <- knit_hooks$get("output")
knit_hooks$set(output = function(x, options) {
  lines <- options$output.lines
  if (is.null(lines)) {
    return(hook_output(x, options))  # pass to default hook
  }
  x <- unlist(strsplit(x, "\n"))
  more <- "..."
  if (length(lines) == 1) {        # first n lines
    if (length(x) > lines) {
      # truncate the output, but add ....
      x <- c(head(x, lines), more)
    }
  } else {
    x <- c(more, x[lines], more)
  }
  # paste these lines together
  x <- paste(c(x, ""), collapse = "\n")
  hook_output(x, options)
})
library(tidyr)
``` 

# What is Natural Language Processing (NLP)

Using computers to extract insights and make decision based on human languages.

- Information extraction
- Machine translation
- Speech processing
- Image understanding

???

https://www.ling.upenn.edu/~beatrice/humor/headlines.html

Not Computational linguistics which is using computers to reason about human langauges

---

# Plan of Action


## Text mining

We will be doing text mining as exploratory data analysis

## Modeling

Apply some simple models to make decisions only based on text

---

background-image: url("http://3.bp.blogspot.com/-FDjkfNyHOCA/UNNVcB9eBsI/AAAAAAAABp4/fivfG6senDU/s1600/fir-tree1.jpg")
background-position: 90% 35%
background-size: 20% 30%

# The Data

## Text mining

The Fir Tree  
H.C. Andersen  
21 December 1844  
EK (**e**ventyr**k**ode / fairly tale code) = 26

Fairly early, one of first works displaying pessimism.


???

Types of data

- Strings
- Document term matrix
- Corpus

---

# The Data

## Modeling

Movie reviews from [IMDb.com](IMDb.com)

2 Movies

---

.pull-left[
![](https://m.media-amazon.com/images/M/MV5BMzUxNzkzMzQtYjIxZC00NzU0LThkYTQtZjNhNTljMTA1MDA1L2ltYWdlL2ltYWdlXkEyXkFqcGdeQXVyMTMxODk2OTU@._V1_.jpg)
]

.pull-right[
![](https://m.media-amazon.com/images/M/MV5BNWNiNTczNzEtMjQyZC00MjFmLTkzMDMtODk4ZGMyZmE0N2E4XkEyXkFqcGdeQXVyMTMxODk2OTU@._V1_.jpg)
]

---

# Finding gold

.medium[
```{r}
#devtools::install_github("emilhvitfeldt/hcandersenr")
library(hcandersenr)
```
]

Includes most of H.C. Andersens 157 fairly tales in 5 languages (Danish, German, English, Spanish, French).

--

.medium[
```{r}
library(dplyr)
library(tidyr)
tree <- hcandersen_en %>%
  filter(book == "The fir tree") %>%
  select(text)
```
]

---

# The Text

.smallish[
```{r}
tree
```
]

---

.pull-left[
```{r}
library(tidytext)

unnest_tokens(tree, word, text) #<<
```
]

.pull-right[
Observational unit: 

- Document
- Sentence
- Word
- Letter
]

---

.pull-left[
```{r}
library(tidytext)

unnest_tokens(tree, word, text) #<<
```
]

.pull-right[
Observational unit: 

- ~~Document~~
- ~~Sentence~~
- **Word**
- ~~Letter~~

Word tokens are default in `unnest_tokens()`
]

---

```{r}
library(tidytext)

unnest_tokens(tree, word, text) %>%
  count(word, sort = TRUE) #<<
```

---

```{r, highlight.output=c(4, 5, 7, 8, 9, 10, 11, 12, 13)}
library(tidytext)

unnest_tokens(tree, word, text) %>%
  count(word, sort = TRUE) #<<
```

--

These words don't give us much context

---

# Stop words

Stop words or "non-context" words are words that doens't add much to the text (filler that make sentences work).

```{r, output.lines = 9}
stop_words$word
```

---

# Stop words

Stop words or "non-context" words are words that doens't add much to the text (filler that make sentences work).

Don't remove stop words willy nilly!  

Pre-constructed word list might not work in your domain  

Creating your own word list is hard...

???

Lean
computer
old

---

```{r}
unnest_tokens(tree, word, text) %>%
  inner_join(stop_words, by = "word") %>% #<<
  count(word, sort = TRUE) %>%
  top_n(50, n) %>%
  pull(word)
```

Look at the words you remove (easy)  

or know you stop word by heart (hard!!!)

---

```{r}
unnest_tokens(tree, word, text) %>%
  anti_join(stop_words, by = "word") %>% #<<
  count(word, sort = TRUE)
```

---

```{r, eval=FALSE}
library(ggplot2)
unnest_tokens(tree, word, text) %>%
  anti_join(stop_words, by = "word") %>%
  count(word, sort = TRUE) %>%
  top_n(10, n) %>%
  ggplot(aes(reorder(word, n), n)) +
  geom_col() +
  coord_flip() +
  theme_light() +
  labs(x = "Times",
       y = "Word",
       title = "Word frequency in 'The Fur Tree'")
```

---

```{r echo=FALSE, fig.asp=0.618, fig.width=5, dpi=300, out.width='100%'}
library(ggplot2)
unnest_tokens(tree, word, text) %>%
  anti_join(stop_words, by = "word") %>%
  count(word, sort = TRUE) %>%
  top_n(10, n) %>%
  ggplot(aes(reorder(word, n), n)) +
  geom_col() +
  coord_flip() +
  theme_light() +
  labs(x = "Times",
       y = "Word",
       title = "Word frequency in 'The Fur Tree'")
```

---

```{r, eval=FALSE}
unnest_tokens(tree, word, text) %>%
  mutate(pos = row_number(),
         place = word == "story") %>%
  filter(place) %>%
  ggplot(aes(pos, place)) +
  geom_point() +
  theme_light() +
  labs(x = "Position",
       y = "",
       title = "Occurence plot of word 'story'")
```

---

```{r, echo=FALSE, out.width = '90%', fig.asp=0.618, fig.width=4, dpi=300}
unnest_tokens(tree, word, text) %>%
  mutate(pos = row_number(),
         place = word == "story") %>%
  filter(place) %>%
  ggplot(aes(pos, place)) +
  geom_point() +
  theme_light() +
  labs(x = "Position",
       y = "",
       title = "Occurence plot of word 'story'")
```

---

```{r, eval=FALSE}
unnest_tokens(tree, word, text) %>%
  mutate(pos = row_number(),
         place = case_when(word == "story" ~ "story",
                           word %in% c("tree", "trees") ~ "tree",
                           word == "mice" ~ "mice",
                           word == "children" ~ "children",
                           word == "forest" ~ "forest",
                           TRUE ~ NA_character_)) %>%
  drop_na() %>%
  ggplot(aes(pos, place, color = place)) +
  geom_point() +
  theme_light() +
  guides(color = "none") +
  labs(x = "Position",
       y = "Word",
       title = "Occurence plot of 'The Fur Tree'")
```

---

```{r, echo=FALSE, out.width = '90%', fig.asp=0.618, fig.width=4, dpi=300}
unnest_tokens(tree, word, text) %>%
  mutate(pos = row_number(),
         place = case_when(word == "story" ~ "story",
                           word %in% c("tree", "trees") ~ "tree",
                           word == "mice" ~ "mice",
                           word == "children" ~ "children",
                           word == "forest" ~ "forest",
                           TRUE ~ NA_character_)) %>%
  drop_na() %>%
  ggplot(aes(pos, place, color = place)) +
  geom_jitter(width = 0, height = 0.2) +
  theme_light() +
  guides(color = "none") +
  labs(x = "Position",
       y = "Word",
       title = "Occurence plot of 'The Fur Tree'")
```

---

## Going to the movies

Scraped review (scraping_reviews.Rmd)

```{r, message=FALSE}
library(readr)
library(tidyr)
reviews <- read_csv("review_data.csv") %>%
  select(movie, rating, review) %>%
  drop_na()
```

.smallish[
```{r, echo=FALSE}
reviews
```
]

---

```{r, out.width = '90%', fig.asp=0.618, fig.width=4, dpi=300}
ggplot(reviews, aes(as.factor(rating), 1, fill = movie)) +
  geom_col() +
  facet_grid(~ movie) +
  theme_minimal() +
  labs(x = "Rating",
       y = "Count") +
  guides(fill = "none")
```

---

## tidymodels

```{r}
library(tidymodels)
#devtools::install_github("tidymodels/textrecipes")
library(textrecipes)
```

`tidymodels` is a "meta-package" for modeling and statistical analysis that share the underlying design philosophy, grammar, and data structures of the tidyverse.

`textrecipes` is an addition to the recipes package providing text processing capabilities (coming to CRAN any day)

---

.medium[
```{r}
set.seed(2018)

split <- reviews %>%
  mutate(good = factor(rating > 5, labels = c("bad", "good"))) %>%
  select(good, text = review) %>%
  initial_split(props = 7 / 10)

review_train <- training(split)
review_test  <- testing(split)
```
]

Splitting data into training and testing set

Next we specify a preprocesing step using recipes

---

## What do we measure?

![](https://user-images.githubusercontent.com/6179259/47669547-78c9f180-dbee-11e8-85e8-e01cb4cbe46d.png)

---

.medium[
```{r, eval=FALSE}
review_rec <- recipe(good ~ ., data = review_train) %>%
  step_tokenize(text) %>%
  step_tokenfilter(text, max_tokens = 500) %>%
  step_tfidf(text) %>%
  prep(training = review_train)

review_rec
```
]

```{r, echo=FALSE}
review_rec <- recipe(good ~ ., data = review_train) %>%
  step_tokenize(text) %>%
  step_tokenfilter(text, max_tokens = 500) %>%
  step_tf(text) %>%
  prep(training = review_train)

review_rec
```

---

```{r}
# Processed data
train_data <- juice(review_rec)
test_data  <- bake(review_rec, review_test)
```

```{r}
train_data
```

---

```{r}
review_model <- logistic_reg() %>%
  set_engine("glm")
```

```{r, message=FALSE, warning=FALSE}
review_fit <- review_model %>%
  fit(good ~ ., data = train_data)
```

```{r}
test_results <- review_test %>%
  bind_cols(
   predict(review_fit, test_data)
  ) 

test_results %>% accuracy(truth = good, estimate = .pred_class)
```

(This is not an example of finished classification. The Accuracy is not that good, but the general steps you would follow.)