使用主題模型分析新冠肺炎疫苗中文新聞資料

ch0.套件取得及資料載入

套件

library(data.table)
library(ggplot2)
library(dplyr)
library(jiebaR)
library(tidytext)
library(stringr)
library(tm)
library(topicmodels)
library(purrr)
require(RColorBrewer)
mycolors <- colorRampPalette(brewer.pal(8, "Set3"))(20)

資料描述

透過中山管院文字分析平台，載入聯合新聞網、蘋果新聞網、東森新聞網的新聞，搜尋關鍵字為「疫苗」，時間從2020/10/01到2021/05/09。

metadata <- fread("news_articleMetaData.csv", encoding = "UTF-8")

可以看到疫苗討論在2月過後的新聞報導數量增加

metadata %>% 
  mutate(artDate = as.Date(artDate)) %>%
  group_by(artDate) %>%
  summarise(count = n())%>%
  ggplot(aes(artDate,count))+
    geom_line(color="red")+
    geom_point()

#Ch1. Document Term Matrix (DTM)

資料前處理

使用默認參數初始化一個斷詞引擎

jieba_tokenizer = worker()
news_tokenizer <- function(t) {
  lapply(t, function(x) {
    if(nchar(x)>1){
      tokens <- segment(x, jieba_tokenizer)
      # 去掉字串長度爲1的詞彙
      tokens <- tokens[nchar(tokens)>1]
      return(tokens)
    }
  })
}

計算每篇文章各token出現次數

tokens <- metadata %>%
  unnest_tokens(word, sentence, token=news_tokenizer) %>%
  filter((!str_detect(word, regex("[0-9a-zA-Z]"))) | str_detect(word, regex("[Aa][Zz]"))) %>%
  count(artUrl, word) %>%
  rename(count=n)
tokens %>% head(20)

將資料轉換為Document Term Matrix (DTM)

dtm <-tokens %>% cast_dtm(artUrl, word, count)
dtm
inspect(dtm[1:10,1:10])

ch2. 主題模型

建立LDA模型

lda <- LDA(dtm, k = 2, control = list(seed = 2021))
# lda <- LDA(dtm, k = 2, control = list(seed = 2021,alpha = 2,delta=0.1),method = "Gibbs") #調整alpha即delta
lda

利用LDA模型建立phi矩陣

topics_words <- tidy(lda, matrix = "beta") # 注意，在tidy function裡面要使用"beta"來取出Phi矩陣。
colnames(topics_words) <- c("topic", "term", "phi")
topics_words

尋找Topic的代表字

terms依照各主題的phi值由大到小排序，列出前10大

topics_words %>%
  group_by(topic) %>%
  top_n(10, phi) %>%
  ungroup() %>%
  mutate(top_words = reorder_within(term,phi,topic)) %>%
  ggplot(aes(x = top_words, y = phi, fill = as.factor(topic))) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~ topic, scales = "free") +
  coord_flip() +
  scale_x_reordered()

#ch3. 尋找最佳主題數

建立更多主題的主題模型

嘗試2、4、6、10、15個主題數，將結果存起來，再做進一步分析。 此部分需要跑一段時間，已經將跑完的檔案存成ldas_result.rdata，可以直接載入

# ldas = c()
# topics = c(2,4,6,10,15)
# for(topic in topics){
#   start_time <- Sys.time()
#   lda <- LDA(dtm, k = topic, control = list(seed = 2021))
#   ldas =c(ldas,lda)
#   print(paste(topic ,paste("topic(s) and use time is ", Sys.time() -start_time)))
#   save(ldas,file = "ldas_result.rdata") # 將模型輸出成檔案
# }

載入每個主題的LDA結果

load("ldas_result.rdata")

透過perplexity找到最佳主題數

topics = c(2,4,6,10,15)
data_frame(k = topics, perplex = map_dbl(ldas, topicmodels::perplexity)) %>%
  ggplot(aes(k, perplex)) +
  geom_point() +
  geom_line() +
  labs(title = "Evaluating LDA topic models",
       subtitle = "Optimal number of topics (smaller is better)",
       x = "Number of topics",
       y = "Perplexity")

create LDAvis所需的json function 此function是將前面使用 “LDA function”所建立的model，轉換為“LDAVis”套件的input格式。

topicmodels_json_ldavis <- function(fitted, doc_term){
    require(LDAvis)
    require(slam)
  
    ###以下function 用來解決，主題數多會出現NA的問題
    ### 參考 https://github.com/cpsievert/LDAvis/commit/c7234d71168b1e946a361bc00593bc5c4bf8e57e
    ls_LDA = function (phi){
      jensenShannon <- function(x, y) {
        m <- 0.5 * (x + y)
        lhs <- ifelse(x == 0, 0, x * (log(x) - log(m+1e-16)))
        rhs <- ifelse(y == 0, 0, y * (log(y) - log(m+1e-16)))
        0.5 * sum(lhs) + 0.5 * sum(rhs)
      }
      dist.mat <- proxy::dist(x = phi, method = jensenShannon)
      pca.fit <- stats::cmdscale(dist.mat, k = 2)
      data.frame(x = pca.fit[, 1], y = pca.fit[, 2])
    }
  
      # Find required quantities
      phi <- as.matrix(posterior(fitted)$terms)
      theta <- as.matrix(posterior(fitted)$topics)
      vocab <- colnames(phi)
      term_freq <- slam::col_sums(doc_term)
  
      # Convert to json
      json_lda <- LDAvis::createJSON(phi = phi, theta = theta,
                                     vocab = vocab,
                                     doc.length = as.vector(table(doc_term$i)),
                                     term.frequency = term_freq, mds.method = ls_LDA)
  
      return(json_lda)
}

產生LDAvis結果

the_lda = ldas[[2]]
json_res <- topicmodels_json_ldavis(the_lda,dtm)
serVis(json_res,open.browser = T)

產生LDAvis檔案，存至local端

serVis(json_res, out.dir = "vis", open.browser = T)
writeLines(iconv(readLines("./vis/lda.json"), to = "UTF8"))

ch4. LDA分析

選定4個主題數的主題模型

the_lda = ldas[[2]] ## 選定topic 為 4 的結果

topics_words <- tidy(the_lda, matrix = "beta") # 注意，在tidy function裡面要使用"beta"來取出Phi矩陣。
colnames(topics_words) <- c("topic", "term", "phi")
topics_words %>% arrange(desc(phi)) %>% head(10)

terms依照各主題的phi值由大到小排序

topics_words %>%
  group_by(topic) %>%
  top_n(10, phi) %>%
  ungroup() %>%
  ggplot(aes(x = reorder_within(term,phi,topic), y = phi, fill = as.factor(topic))) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~ topic, scales = "free") +
  coord_flip() +
  scale_x_reordered()

去除共通詞彙

removed_word = c("肺炎","新冠","疫苗","接種","目前","表示","沒有")

topics_words %>%
  filter(!term  %in% removed_word) %>%
  group_by(topic) %>%
  top_n(10, phi) %>%
  ungroup() %>%
  ggplot(aes(x = reorder_within(term,phi,topic), y = phi, fill = as.factor(topic))) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~ topic, scales = "free") +
  coord_flip() +
  scale_x_reordered()

主題命名

topics_name = c("AZ疫苗","台灣疫苗施打","疫苗研發進度","輝瑞疫苗")

Document 主題分佈

# for every document we have a probability distribution of its contained topics
tmResult <- posterior(the_lda)
doc_pro <- tmResult$topics
document_topics <- doc_pro[metadata$artUrl,]
document_topics_df =data.frame(document_topics)
colnames(document_topics_df) = topics_name
rownames(document_topics_df) = NULL
news_topic = cbind(metadata,document_topics_df)

現在我們看每一篇的文章分佈了！

查看特定主題的文章

• 透過找到特定文章的分佈進行排序之後，可以看到此主題的比重高的文章在討論什麼。

news_topic %>%
  arrange(desc(`AZ疫苗`)) %>%head(10) 

了解主題在時間的變化

news_topic %>% 
  mutate(artDate = as.Date(artDate)) %>%
  group_by(artDate = format(artDate,'%Y%m')) %>%
  summarise_if(is.numeric, sum, na.rm = TRUE) %>%
  melt(id.vars = "artDate")%>%
  ggplot( aes(x=artDate, y=value, fill=variable)) + 
  geom_bar(stat = "identity") + ylab("value") + 
  scale_fill_manual(values=mycolors[c(1,5,8,12)])+
  theme(axis.text.x = element_text(angle = 90, hjust = 1))

去除筆數少月份

news_topic %>%
  mutate(artDate = as.Date(artDate)) %>% 
  filter( !format(artDate,'%Y%m') %in% c(202011,202105))%>%
  group_by(artDate = format(artDate,'%Y%m')) %>%
  summarise_if(is.numeric, sum, na.rm = TRUE) %>%
  melt(id.vars = "artDate")%>%
  ggplot( aes(x=artDate, y=value, fill=variable)) + 
  geom_bar(stat = "identity") + ylab("value") + 
    scale_fill_manual(values=mycolors[c(1,5,8,12)])+
    theme(axis.text.x = element_text(angle = 90, hjust = 1))

以比例了解主題時間變化

news_topic %>%
  mutate(artDate = as.Date(artDate)) %>% 
  filter( !format(artDate,'%Y%m') %in% c(202011,202105))%>%
  group_by(artDate = format(artDate,'%Y%m')) %>%
  summarise_if(is.numeric, sum, na.rm = TRUE) %>%
  melt(id.vars = "artDate")%>%
  group_by(artDate)%>%
  mutate(total_value =sum(value))%>%
  ggplot( aes(x=artDate, y=value/total_value, fill=variable)) + 
  geom_bar(stat = "identity") + ylab("proportion") + 
    scale_fill_manual(values=mycolors[c(1,5,8,12)])+
    theme(axis.text.x = element_text(angle = 90, hjust = 1))

補充 - 不同訓練LDA模型套件

參考 http://text2vec.org/topic_modeling.html#latent_dirichlet_allocation

library(text2vec)
library(udpipe)
tokens <- metadata %>%
  unnest_tokens(word, sentence, token=news_tokenizer) %>%
  filter(!str_detect(word, regex("[0-9a-zA-Z]"))| str_detect(word, regex("[Aa][Zz]")))

建立DTM matrix

dtf <- document_term_frequencies(tokens, document = "artUrl", term = "word")
dtm <- document_term_matrix(x = dtf)
dtm_clean <- dtm_remove_lowfreq(dtm, minfreq = 30)
dim(dtm_clean)

LDA 模型

set.seed(2019)

topic_n = 4

lda_model =text2vec::LDA$new(n_topics = topic_n,doc_topic_prior = 0.1, topic_word_prior = 0.001)
doc_topic_distr =lda_model$fit_transform(dtm_clean, n_iter = 1000, convergence_tol = 1e-5,check_convergence_every_n = 100)

這個比topicmodels的package跑快超多倍

一樣可以用LDAvis的套件來看

lda_model$get_top_words(n = 10, lambda = 0.5) ## 查看 前10主題字
lda_model$plot()
# lda_model$plot(out.dir ="lda_result", open.browser = TRUE)