经过研究表明，在旅行者的决策过程中，TripAdvisor（猫途鹰，全球旅游点评网）正变得越来越重要。然而，了解TripAdvisor评分与数千个评论文本中的每一个的细微差别是很有挑战性的。为了更彻底地了解酒店客人的评论是否会影响酒店的加班表现，我从TripAdvisor截取了一家酒店 - 希尔顿夏威夷度假村（Hilton Hawaiian Village）的所有英语评论 （Web抓取的细节和Python代码在文末）。

加载库

library(dplyr)

library(readr)

library(lubridate)

library(ggplot2)

library(tidytext)

library(tidyverse)

library(stringr)

library(tidyr)

library(scales)

library(broom)

library(purrr)

library(widyr)

library(igraph)

library(ggraph)

library(SnowballC)

library(wordcloud)

library(reshape2)

theme_set(theme_minimal())

数据

df <- read_csv("Hilton_Hawaiian_Village_Waikiki_Beach_Resort-Honolulu_Oahu_Hawaii__en.csv")

df <- df[complete.cases(df), ]

df$review_date <- as.Date(df$review_date, format = "%d-%B-%y")

dim(df); min(df$review_date); max(df$review_date)

在TripAdvisor上希尔顿夏威夷度假村共有13,701条评论，评论日期范围是2002-03-21到2018-08-02。

df %>%

  count(Week = round_date(review_date, "week")) %>%

  ggplot(aes(Week, n)) +

  geom_line() + 

  ggtitle('The Number of Reviews Per Week')

2014年底收到的每周评论数量最多。该酒店在那一周收到了70多条评论。

评论文本的文本挖掘

df <- tibble::rowid_to_column(df, "ID")

df <- df %>%

  mutate(review_date = as.POSIXct(review_date, origin = "1970-01-01"),month = round_date(review_date, "month"))

review_words <- df %>%

  distinct(review_body, .keep_all = TRUE) %>%

  unnest_tokens(word, review_body, drop = FALSE) %>%

  distinct(ID, word, .keep_all = TRUE) %>%

  anti_join(stop_words, by = "word") %>%

  filter(str_detect(word, "[^\\d]")) %>%

  group_by(word) %>%

  mutate(word_total = n()) %>%

  ungroup()

word_counts <- review_words %>%

  count(word, sort = TRUE)

word_counts %>%

  head(25) %>%

  mutate(word = reorder(word, n)) %>%

  ggplot(aes(word, n)) +

  geom_col(fill = "lightblue") +

  scale_y_continuous(labels = comma_format()) +

  coord_flip() +

  labs(title = "Most common words in review text 2002 to date",

       subtitle = "Among 13,701 reviews; stop words removed",

       y = "# of uses")

word_counts %>%

  head(25) %>%

  mutate(word = wordStem(word)) %>% 

  mutate(word = reorder(word, n)) %>%

  ggplot(aes(word, n)) +

  geom_col(fill = "lightblue") +

  scale_y_continuous(labels = comma_format()) +

  coord_flip() +

  labs(title = "Most common words in review text 2002 to date",

       subtitle = "Among 13,701 reviews; stop words removed and stemmed",

       y = "# of uses")

bigram

我们经常想要了解评论中单词之间的关系。在评论文本中，有哪些常见的单词序列？给定一些单词，哪些单词最有可能跟随在这个单词后面？哪些词关联最紧密？因此，许多有趣的文本分析都是基于这种关联。当我们检查两个连续单词的对时，它被称为“bigram”（二元语法）。

那么，这家酒店的评论中最常见的bigram评论是什么？

review_bigrams <- df %>%

  unnest_tokens(bigram, review_body, token = "ngrams", n = 2)

bigrams_separated <- review_bigrams %>%

  separate(bigram, c("word1", "word2"), sep = " ")

bigrams_filtered <- bigrams_separated %>%

  filter(!word1 %in% stop_words$word) %>%

  filter(!word2 %in% stop_words$word)

bigram_counts <- bigrams_filtered %>% 

  count(word1, word2, sort = TRUE)

bigrams_united <- bigrams_filtered %>%

  unite(bigram, word1, word2, sep = " ")

bigrams_united %>%

  count(bigram, sort = TRUE)

最常见的bigram是“rainbow tower”，其次是“hawaiian village”。

我们可以在单词网络中可视化bigram：

review_subject <- df %>% 

  unnest_tokens(word, review_body) %>% 

  anti_join(stop_words)

my_stopwords <- data_frame(word = c(as.character(1:10)))

review_subject <- review_subject %>% 

  anti_join(my_stopwords)

title_word_pairs <- review_subject %>% 

  pairwise_count(word, ID, sort = TRUE, upper = FALSE)

set.seed(1234)

title_word_pairs %>%

  filter(n >= 1000) %>%

  graph_from_data_frame() %>%

  ggraph(layout = "fr") +

  geom_edge_link(aes(edge_alpha = n, edge_width = n), edge_colour = "cyan4") +

  geom_node_point(size = 5) +

  geom_node_text(aes(label = name), repel = TRUE, 

                 point.padding = unit(0.2, "lines")) +

  ggtitle('Word network in TripAdvisor reviews')

  theme_void()

上面显示了TripAdvisor评论中常见的bigram组合，显示了至少出现了1000次且不是停用词的单词。

网络图显示了前几个词（“hawaiian ”，“village ”，“ocean ”和“view ”）之间的紧密联系。然而，我们在网络中并没有看到清晰的聚类结构。

Trigram

Bigram有时是不够的，让我们看看希尔顿夏威夷度假村在TripAdvisor评论中最常见的trigram（三元语法）？

review_trigrams <- df %>%

  unnest_tokens(trigram, review_body, token = "ngrams", n = 3)



trigrams_separated <- review_trigrams %>%

  separate(trigram, c("word1", "word2", "word3"), sep = " ")



trigrams_filtered <- trigrams_separated %>%

  filter(!word1 %in% stop_words$word) %>%

  filter(!word2 %in% stop_words$word) %>%

  filter(!word3 %in% stop_words$word)



trigram_counts <- trigrams_filtered %>% 

  count(word1, word2, word3, sort = TRUE)



trigrams_united <- trigrams_filtered %>%

  unite(trigram, word1, word2, word3, sep = " ")



trigrams_united %>%

  count(trigram, sort = TRUE)

最常见的trigram 是“hilton hawaiian village”，其次是“hilton hawaiian village”，依此类推。

评论中的重要的词汇趋势

随着时间的推移，哪些词语和话题变得更频繁（或者更频繁）了？这些可以让我们了解酒店不断变化的生态系统，例如服务，翻新，问题解决，让我们预测哪些话题的关联词将继续增长。

我们需要了解的问题是：在TripAdvisor评论中，随着时间的推移，哪些词的频率在增加？

reviews_per_month <- df %>%

  group_by(month) %>%

  summarize(month_total = n())

word_month_counts <- review_words %>%

  filter(word_total >= 1000) %>%

  count(word, month) %>%

  complete(word, month, fill = list(n = 0)) %>%

  inner_join(reviews_per_month, by = "month") %>%

  mutate(percent = n / month_total) %>%

  mutate(year = year(month) + yday(month) / 365)

mod <- ~ glm(cbind(n, month_total - n) ~ year, ., family = "binomial")

slopes <- word_month_counts %>%

  nest(-word) %>%

  mutate(model = map(data, mod)) %>%

  unnest(map(model, tidy)) %>%

  filter(term == "year") %>%

  arrange(desc(estimate))

slopes %>%

  head(9) %>%

  inner_join(word_month_counts, by = "word") %>%

  mutate(word = reorder(word, -estimate)) %>%

  ggplot(aes(month, n / month_total, color = word)) +

  geom_line(show.legend = FALSE) +

  scale_y_continuous(labels = percent_format()) +

  facet_wrap(~ word, scales = "free_y") +

  expand_limits(y = 0) +

  labs(x = "Year",

       y = "Percentage of reviews containing this word",

       title = "9 fastest growing words in TripAdvisor reviews",

       subtitle = "Judged by growth rate over 15 years")

在2010年之前，我们可以看到关于“friday fireworks”和“lagoon”的讨论高峰。像“resort fee”和“busy”这样的词在2005年之前增长最快。

在评论中，哪些词的频率在下降？

word_month_counts %>%

  filter(word %in% c("service", "food")) %>%

  ggplot(aes(month, n / month_total, color = word)) +

  geom_line(size = 1, alpha = .8) +

  scale_y_continuous(labels = percent_format()) +

  expand_limits(y = 0) +

  labs(x = "Year",

       y = "Percentage of reviews containing this term", title = "service vs food in terms of reviewers interest")

服务和食品都是2010年之前的主要话题。关于服务和食品的讨论在2003年左右的数据开始时达到顶峰，在2005年之后一直呈下降趋势，偶尔出现高峰。

情绪分析

情感分析广泛应用于客户反馈，需要分析的有：评论和调查结果，在线和社交媒体。它适用于从营销到客户服务以及临床医学的各种应用。

在我们的案例中，我们的目的是确定评论者（即酒店客人）对他过去对酒店的体验的看法。这种可能是判断或评价。

评论中最常见的正面和负面词汇。

reviews <- df %>% 

  filter(!is.na(review_body)) %>% 

  select(ID, review_body) %>% 

  group_by(row_number()) %>% 

  ungroup()

tidy_reviews <- reviews %>%

  unnest_tokens(word, review_body)

tidy_reviews <- tidy_reviews %>%

  anti_join(stop_words)



bing_word_counts <- tidy_reviews %>%

  inner_join(get_sentiments("bing")) %>%

  count(word, sentiment, sort = TRUE) %>%

  ungroup()



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") +

  labs(y = "Contribution to sentiment", x = NULL) +

  coord_flip() + 

  ggtitle('Words that contribute to positive and negative sentiment in the reviews')

contributions <- tidy_reviews %>%

  inner_join(get_sentiments("afinn"), by = "word") %>%

  group_by(word) %>%

  summarize(occurences = n(),

            contribution = sum(score))

contributions %>%

  top_n(25, abs(contribution)) %>%

  mutate(word = reorder(word, contribution)) %>%

  ggplot(aes(word, contribution, fill = contribution > 0)) +

  ggtitle('Words with the greatest contributions to positive/negative 

          sentiment in reviews') +

  geom_col(show.legend = FALSE) +

  coord_flip()

有趣的是，“diamond ”（diamond head）被归类为积极的情绪。

这里有一个可能出现的问题，例如，“clean”，在不通的上下文，如前面带有“not”，则会产生负面情绪。事实上，在大多数unigram（一元模型）会有这个否定的问题。所以我们需要进行下一步：

使用Bigrams在情感分析中提供语境

我们想知道单词前面有“not”这样的单词的频率。

bigrams_separated %>%

  filter(word1 == "not") %>%

  count(word1, word2, sort = TRUE)

AFINN <- get_sentiments("afinn")

not_words <- bigrams_separated %>%

  filter(word1 == "not") %>%

  inner_join(AFINN, by = c(word2 = "word")) %>%

  count(word2, score, sort = TRUE) %>%

  ungroup()



not_words

这告诉我们，在数据中，跟随“not”的最常见的情感关联词是“worth”，而跟随“not”的第二个常见情感关联词是“recommend”，这通常得分为2分。

那么，在我们的数据中，哪些词在错误的方向上做了最大的“贡献”呢？

not_words %>%

  mutate(contribution = n * score) %>%

  arrange(desc(abs(contribution))) %>%

  head(20) %>%

  mutate(word2 = reorder(word2, contribution)) %>%

  ggplot(aes(word2, n * score, fill = n * score > 0)) +

  geom_col(show.legend = FALSE) +

  xlab("Words preceded by \"not\"") +

  ylab("Sentiment score * number of occurrences") +

  ggtitle('The 20 words preceded by "not" that had the greatest contribution to 

          sentiment scores, positive or negative direction') +

  coord_flip()

“not worth”，“not great”，“not good”，“not recommend”和“not like”的最大的错误识别原因，这使得文本看起来比实际上更积极。

除了“not”之外，还有其他词语否定后续词语，例如“no”，“never”和“without”。

negation_words <- c("not", "no", "never", "without")



negated_words <- bigrams_separated %>%

  filter(word1 %in% negation_words) %>%

  inner_join(AFINN, by = c(word2 = "word")) %>%

  count(word1, word2, score, sort = TRUE) %>%

  ungroup()



negated_words %>%

  mutate(contribution = n * score,

         word2 = reorder(paste(word2, word1, sep = "__"), contribution)) %>%

  group_by(word1) %>%

  top_n(12, abs(contribution)) %>%

  ggplot(aes(word2, contribution, fill = n * score > 0)) +

  geom_col(show.legend = FALSE) +

  facet_wrap(~ word1, scales = "free") +

  scale_x_discrete(labels = function(x) gsub("__.+$", "", x)) +

  xlab("Words preceded by negation term") +

  ylab("Sentiment score * # of occurrences") +

  ggtitle('The most common positive or negative words to follow negations 

          such as "no", "not", "never" and "without"') +

  coord_flip()

看起来把一个词误认为是正面情绪的最大来源是“not worth/great/good/recommend”，而错误分类的负面情绪的最大来源是“not bad”和“no problem”。

最后，让我们找出最正面和最负面的评论。

sentiment_messages <- tidy_reviews %>%

  inner_join(get_sentiments("afinn"), by = "word") %>%

  group_by(ID) %>%

  summarize(sentiment = mean(score),

            words = n()) %>%

  ungroup() %>%

  filter(words >= 5)

sentiment_messages %>%

  arrange(desc(sentiment))

df [which（df $ ID == 2363），] $ review_body [1]

sentiment_messages %>%

  arrange(sentiment)

df [which（df $ ID == 3748），] $ review_body [1]