Sentiment Analysis on Music Streaming Services Reviews

library(tidytext)
library(readr)
library(dplyr)
library(anytime)

# Load manually scored review files here. ("../service_scored.csv")
service <- read_csv("https://raw.githubusercontent.com/seonnyseo/Streaming_Sentiment_Analysis/master/Data/amazon_scored.csv")
  
# I decided to consider ambiguous reviews as negative
service$score <- ifelse(service$score == 0, -1, 1)
  
# Split reviews by each word 
tidy_service <- service %>% unnest_tokens(word, review)
  
bing_sentiments <- tidy_service %>% inner_join(get_sentiments("bing"), by = "word")
bing_sentiments$score <- ifelse(bing_sentiments$sentiment == "negative", -1, 1)
bing_aggregate <- bing_sentiments %>% 
  select(review_id, score) %>% 
  group_by(review_id) %>% 
  summarise(bing_score = sum(score))
  
score_compare_service <- merge(x = service, y = bing_aggregate, all.x = TRUE, by = 'review_id')
score_compare_service[is.na(score_compare_service)] <- 0
  

# score_compare_service(review_id, date, review, score, bing_score)
# Pick 50 each time 
resampling <- function(service){
  
  neutral_average <- 0
  postive_average <- 0
  negative_average <- 0
  
  for(i in c(2000:3000)){
    set.seed(i)
    random_data <- service[sample(nrow(service), 50),]
    
    neutral_count <- sum(random_data$bing_score == 0)
    positive_neutral <- sum(random_data$bing_score == 0 & random_data$score == 1)
    negative_neutral <- sum(random_data$bing_score == 0 & random_data$score == -1)
    
    neutral_average <- neutral_average + neutral_count/50
    postive_average <- postive_average + positive_neutral/neutral_count
    negative_average <- negative_average + negative_neutral/neutral_count
  }
  cat(sprintf("Neutral : %.3f  Positive : %.3f  Negative : %.3f\n", 
              neutral_average/1000, postive_average/1000, negative_average/1000))
}
  
# Run resampling
resampling(score_compare_service)

## Neutral : 0.140  Positive : 0.710  Negative : 0.291

pre_process <- function(service)
{
  # PreProcessing
  service$review_id <- 1:nrow(service)
  service$date <- anydate(service$date)
  service <- service[c(3,1,2)]
  
  # Sentiment Score
  tidy_service <- service %>% unnest_tokens(word, review)
  
  # Edit Dictionary, I only add 4 words with sentiment at this time. This can be expanded later. 
  bing_edit <- rbind(get_sentiments("bing"), c("commercial", "negative"))
  bing_edit <- rbind(bing_edit, c("commercials", "negative"))
  bing_edit <- rbind(bing_edit, c("ad", "negative"))
  bing_edit <- rbind(bing_edit, c("ads", "negative"))
  bing_edit <- rbind(bing_edit, c("wish", "negative"))
  
  bing_sentiments <- tidy_service %>% inner_join(bing_edit, by = "word")
  
  bing_sentiments$score <- ifelse(bing_sentiments$sentiment == "negative", -1, 1)
  bing_aggregate <- bing_sentiments %>% select(review_id, score) %>% group_by(review_id) %>% summarise(bing_score = sum(score))
  
  service <- merge(x = service, y = bing_aggregate, all.x = TRUE, by = 'review_id')
  service[is.na(service)] <- 0
  service$bing_judgement <- ifelse(service$bing_score > 0, "positive", 
                                   ifelse(service$bing_score < 0, "negative", "neutral" ))
  
  return(service)
}

word_data <-function(service, start, end, word, sentiment){

  word <- tolower(word)
  
  # Filter Data between start date & end date
  extracted <- service[service$date >= start & service$date <= end,]
  # Filter Date that only contains word
  extracted <- extracted[grepl(word, tolower(extracted$review)),]
  
 
  set.seed(101)

  # Neutral / (Positive|Neutral) / (Negative/Neutral)
  # Pandora	0.14	0.54	0.46
  # Spotify	0.13	0.59	0.41
  # Amazon	0.14	0.70	0.30
  
  ifelse(service$service == "pandora", positive_weight <- 0.54,
         ifelse(service$service == "spotify", positive_weight <- 0.59, positive_weight <- 0.70))
    
  neutral_reviews <- extracted[extracted$bing_judgement == "neutral",] %>% select(review_id)
  positive_neutral <- neutral_reviews[sample(nrow(neutral_reviews), nrow(neutral_reviews) * positive_weight),]
  negative_neutral <- neutral_reviews[!(neutral_reviews$review_id %in% positive_neutral),]
    
  extracted$bing_judgement <- ifelse(extracted$review_id %in% positive_neutral, "positive",
                                      ifelse(extracted$review_id %in% negative_neutral, "negative",
                                            extracted$bing_judgement))
  
  
 
  extracted <- extracted[extracted$bing_judgement == sentiment,]
  
  return(extracted)
}

frequency_month <- function(service, start, end, word, sentiment){
  
  extracted <- word_data(service, start, end, word, sentiment)

  # Make year-month column
  extracted$year_month <- anydate(format(as.Date(extracted$date), "%Y-%m"))

  frequency_df <- extracted %>% group_by(year_month) %>% summarise(frequency = n())
  frequency_df <- frequency_df %>% pad(interval = 'month', start_val = anydate(start), end_val = anydate(end))
  frequency_df[is.na(frequency_df)] <- 0
  return (frequency_df)
}

word_graph <- function(service, word, start, end){

  positive = frequency_month(service, start, end, word, "positive")
  negative = frequency_month(service, start, end, word, "negative")
  
  positive$sentiment <- 'positive'
  negative$sentiment <- 'negative'
  
  frequency_df <- positive %>% full_join(negative)
  
  ret <- ggplot(frequency_df, aes(x = year_month)) +
        geom_line(aes(y = frequency, col = sentiment)) +
        theme(axis.title.x=element_blank())
  
  return(ret)
}