使用LDA分析新型冠狀病毒中文新聞資料

系統設置

Sys.setlocale(category = "LC_ALL", locale = "zh_TW.UTF-8") # For ubuntu

[1] "LC_CTYPE=zh_TW.UTF-8;LC_NUMERIC=C;LC_TIME=zh_TW.UTF-8;LC_COLLATE=zh_TW.UTF-8;LC_MONETARY=zh_TW.UTF-8;LC_MESSAGES=zh_HK.UTF-8;LC_PAPER=zh_TW.UTF-8;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=zh_TW.UTF-8;LC_IDENTIFICATION=C"

# Sys.setlocale("LC_CTYPE", "cht") # For windows.

安裝需要的packages

packages = c("readr","tm", "data.table", "dplyr", "stringr", "jiebaR", "tidytext", "ggplot2", "tidyr", "topicmodels", "LDAvis","data", "webshot","purrr","ramify","RColorBrewer", "htmlwidgets","servr")
existing = as.character(installed.packages()[,1])
for(pkg in packages[!(packages %in% existing)]) install.packages(pkg)

require(readr)
require(tm)
require(data.table)
require(dplyr)
require(stringr)
require(jiebaR)
require(udpipe)
require(tidytext)
require(ggplot2)
require(tidyr)
require(topicmodels)
require(LDAvis)
require(wordcloud2)
require(webshot)
require(htmlwidgets)
require(servr)
require(purrr)
require(ramify)
require(RColorBrewer)
mycolors <- colorRampPalette(brewer.pal(8, "Set3"))(20)

讀取資料

載入新聞資料

udn= read.csv("raw_data/udn.csv",stringsAsFactors = FALSE) %>% mutate(source = "udn")
ebc= read.csv("raw_data/ebc.csv",stringsAsFactors = FALSE)%>% mutate(source = "ebc")
chinatimes= read.csv("raw_data/chinatimes.csv",stringsAsFactors = FALSE)%>% mutate(source = "chinatimes")
news=rbind(udn,rbind(ebc,chinatimes))
news$artDate = as.Date(news$artDate)

資料描述

• 透過中山管院文字分析平台在聯合新聞網、東森新聞網、中國時報新聞網，搜尋關鍵字：[肺炎、新型冠狀病毒、冠狀病毒、武漢肺炎、防疫、疫情]，時間從2019-12-01~2020-05-01，六個月的資料。 總篇數為22803篇，其中聯合新聞網9128篇、東森新聞網2006篇、中國時報新聞網11669篇

news %>% 
  group_by(artDate) %>%
  summarise(count = n())%>%
  ggplot(aes(artDate,count))+
    geom_line(color="red")+
  geom_point()

可以觀察到資料主要分佈在 1月底之後

Tokenization

# 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)
    }
  })
}

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

將資料轉換為Document Term Matrix (DTM)

news_dtm <-tokens %>% cast_dtm(artUrl, word, count)
news_dtm

<<DocumentTermMatrix (documents: 22803, terms: 145608)>>
Non-/sparse entries: 3192554/3317106670
Sparsity           : 100%
Maximal term length: 16
Weighting          : term frequency (tf)

inspect(news_dtm[1:10,1:10])

<<DocumentTermMatrix (documents: 10, terms: 10)>>
Non-/sparse entries: 25/75
Sparsity           : 75%
Maximal term length: 4
Weighting          : term frequency (tf)
Sample             :
                                             Terms
Docs                                          癌症 癌症病人 安慰 半夜 辦法 爆發 北部 比較 表示 病床
  https://news.ebc.net.tw/news/article/189213    2        1    1    1    1    1    1    1    1    5
  https://news.ebc.net.tw/news/article/192073    0        0    0    0    0    2    0    0    1    0
  https://news.ebc.net.tw/news/article/192142    0        0    0    0    0    1    0    0    3    0
  https://news.ebc.net.tw/news/article/192157    0        0    0    0    0    0    0    0    1    0
  https://news.ebc.net.tw/news/article/192475    0        0    0    0    0    1    0    0    2    0
  https://news.ebc.net.tw/news/article/192528    0        0    0    0    0    0    0    0    0    0
  https://news.ebc.net.tw/news/article/192934    1        0    0    0    0    0    1    0    4    0
  https://news.ebc.net.tw/news/article/193023    0        0    0    0    0    1    0    0    1    0
  https://news.ebc.net.tw/news/article/193107    0        0    0    0    0    0    0    0    1    0
  https://news.ebc.net.tw/news/article/193183    0        0    0    0    0    0    0    3    3    0

查看DTM矩陣，可以發現是個稀疏矩陣。

建立LDA模型

lda <- LDA(news_dtm, k = 2, control = list(seed = 2020))

\(\phi\) Matrix

查看\(\phi\) matrix (topic * term)

topics <- tidy(lda, matrix = "beta") # 注意，在tidy function裡面要使用"beta"來取出Phi矩陣。
topics

從topics中可以得到特定主題生成特定詞彙的概率。

尋找Topic的代表字

top_terms <- topics %>%
  group_by(topic) %>%
  top_n(10, beta) %>%
  ungroup() %>%
  arrange(topic, -beta)


top_terms %>%
  mutate(term = reorder(term, beta)) %>%
  ggplot(aes(term, beta, fill = factor(topic))) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~ topic, scales = "free") +
  coord_flip()

透過上方的圖，感覺兩個主題看起來差不多，沒有看出兩者的差異，嘗試看看分多一點topics

更多主題

• 嘗試2,3,10,25,36主題數，將結果存起來，再做進一步分析

# ldas = c()
# topics = c(2,3,10,25,36)
# for(topic in topics){
#   start_time <- Sys.time()
#   lda <- LDA(news_dtm, k = topic, control = list(seed = 2020))
#   ldas =c(ldas,lda)
#   print(paste(topic ,paste("topic(s) and use time is ", Sys.time() -start_time)))
#   save(ldas,file = "ldas_result")
# }

這邊要跑N個小時，已將主題結果存在lda_result

載入每個主題的LDA結果

load("ldas_result")

結果存在ldas這個變數裡面，同學可以透過ldas[[x]] 來找到想要的主題數

透過perplexity找到最佳主題數

topics = c(2,3,10,25,36)
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")

`data_frame()` is deprecated as of tibble 1.1.0.
Please use `tibble()` instead.
This warning is displayed once every 8 hours.
Call `lifecycle::last_warnings()` to see where this warning was generated.

perplexity 越小越好，但是太小的話，主題數會分太細。通常會找一個主題數適當，且perplexity比較低的主題

補充 ldatuning

• 參考 https://rpubs.com/siri/ldatuning

• Minimization: Arun2010、CaoJuan2009

• Maximization: Deveaud2014、Griffiths2004

if(!('ldatuning' %in% existing)){install.packages(ldatuning)}
library("ldatuning")
result <- FindTopicsNumber(
  news_dtm,
  topics = topics,
  metrics = c("Griffiths2004", "CaoJuan2009", "Arun2010", "Deveaud2014"),
  method = "Gibbs",
  control = list(seed = 2020),
  mc.cores = 2L,
  verbose = TRUE
)
FindTopicsNumber_plot(result)

這邊也要跑N個小時，可以參考上面的連結了解如何使用

LDAvis

產生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

• 可以透過 github的方式，線上demo你的LDAvis結果 範例：https://234gts868.github.io/LDA/3_ldavis/#topic=0&lambda=0.5&term=

# 設置alpha及delta參數
#devotion_lda_removed <- LDA(devotion_dtm_removed, k = 4, method = "Gibbs", control = list(seed = 1234, alpha = 2, delta= 0.1))

####### 以下用來產生ldavis的檔案，可以之後用來在local端、放在網路上打開 ##########
# for(lda in ldas){
#   
#   k = lda@k ## lda 主題數
#   if(k==2){next}
#   json_res <- topicmodels_json_ldavis(lda,news_dtm)
#   # serVis(json_res,open.browser = T)
#   lda_dir =  paste0(k,"_ldavis")
#   if(!dir.exists(lda_dir)){ dir.create("./",lda_dir)}
#   
#   serVis(json_res, out.dir =lda_dir, open.browser = F)
#   
#   writeLines(iconv(readLines(paste0(lda_dir,"/lda.json")), to = "UTF8"))
# }

topic_10 = ldas[[3]]
json_res <- topicmodels_json_ldavis(topic_10,news_dtm)

serVis(json_res,open.browser = T)

# 如果無法開啟視窗(windows用戶)可執行這段
# serVis(json_res, out.dir = "vis", open.browser = T)
# writeLines(iconv(readLines("./vis/lda.json"), to = "UTF8"))

LDA後續分析

• 根據前面的探索之後，我們對於資料有更加了解，並且看完每個主題數的LDAvis之後，選定主題數10的結果來作後續的分析

news_lda = ldas[[3]] ## 選定topic 為10 的結果

topics <- tidy(news_lda, matrix = "beta") # 注意，在tidy function裡面要使用"beta"來取出Phi矩陣。
topics

尋找Topic的代表字

• 整理出每一個Topic中生成概率最高的10個詞彙。

top_terms <- topics %>%
  group_by(topic) %>%
  top_n(10, beta) %>%
  ungroup() %>%
  arrange(topic, -beta)


top_terms %>%
  mutate(term = reorder(term, beta)) %>%
  ggplot(aes(term, beta, fill = factor(topic))) +
  geom_col(show.legend = FALSE) +
  scale_fill_manual(values=mycolors)+
  facet_wrap(~ topic, scales = "free") +
  coord_flip()

可以看到topic都被一開始所使用的搜尋關鍵字影響看不出每一群的差異。

• 移除常出現、跨主題共享的詞彙。

remove_word = c("疫情","武漢","防疫","確診","防疫","肺炎","冠狀病毒","病毒","新冠")
top_terms <- topics %>%
  filter(!term  %in% remove_word)%>%
  group_by(topic) %>%
  top_n(10, beta) %>%
  ungroup() %>%
  arrange(topic, -beta)


top_terms %>%
  mutate(term = reorder(term, beta)) %>%
  ggplot(aes(term, beta, fill = factor(topic))) +
  geom_col(show.legend = FALSE) +
  scale_fill_manual(values=mycolors)+
  facet_wrap(~ topic, scales = "free") +
  coord_flip()

可以看出每個主題主要在討論什麼了！

主題命名

topic_name = c("國際關係",'服務業、舉辦活動','醫院、病例','學校','政府、企業','報導、感染人數','經濟、市場','研發、疫苗','None','中央指揮中心')

Document 主題分佈

# for every document we have a probability distribution of its contained topics
tmResult <- posterior(news_lda)
doc_pro <- tmResult$topics 
dim(doc_pro)               # nDocs(DTM) distributions over K topics

[1] 22803    10

每篇文章都有topic的分佈，所以22803筆的文章*10個主題

cbind Document 主題分佈

# get document topic proportions 
document_topics <- doc_pro[news$artUrl,]
document_topics_df =data.frame(document_topics)
colnames(document_topics_df) = topic_name
rownames(document_topics_df) = NULL
news_topic = cbind(news,document_topics_df)
# news_topic %>% head(10)

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(`國際關係`)) %>%head(10) 

可以看到國際關係這個主題主要討論 美中關係、台灣與WHO，以及其他的國際關係

了解主題在時間的變化

news_topic[,c(7:16)] =sapply(news_topic[,c(7:16)] , as.numeric)
news_topic %>% 
  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)+
  theme(axis.text.x = element_text(angle = 90, hjust = 1))

NA
NA

由於我們的資料，12、5月資料太少，所以將這兩個月去除，以及None的主題

去除筆數少月份、及None的主題

news_topic %>%
  filter( !format(artDate,'%Y%m') %in% c(201912,202005))%>%
  dplyr::select(-None)%>%
  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)+
  theme(axis.text.x = element_text(angle = 90, hjust = 1))

可以看出每個月的聲量，但是不能很清楚出每個月的比例

以比例了解主題時間變化

news_topic %>%
  filter( !format(artDate,'%Y%m') %in% c(201912,202005))%>%
  dplyr::select(-None)%>%
  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)+
  theme(axis.text.x = element_text(angle = 90, hjust = 1))

現在我們可以看到每個月主題的佔比了！

補充

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

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

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)

[1] 22803 13264

LDA 模型

set.seed(2019)

topic_n = 10

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

      [,1]   [,2]     [,3]   [,4]       [,5]   [,6]   [,7]   [,8]   [,9]   [,10] 
 [1,] "警方" "業者"   "醫院" "中國"     "經濟" "防疫" "企業" "專題" "病毒" "自己"
 [2,] "透過" "人潮"   "確診" "美國"     "市場" "口罩" "申請" "數據" "研究" "網友"
 [3,] "活動" "遊客"   "症狀" "大陸"     "成長" "居家" "員工" "圖表" "可能" "口罩"
 [4,] "訊息" "餐廳"   "武漢" "疫情"     "影響" "檢疫" "勞工" "確診" "疫苗" "大家"
 [5,] "捐贈" "飯店"   "患者" "武漢"     "投資" "中央" "取消" "台灣" "認為" "因為"
 [6,] "分局" "連假"   "感染" "防控"     "生產" "市府" "方案" "肺炎" "專家" "沒有"
 [7,] "網路" "民眾"   "病例" "國家"     "營收" "學校" "補助" "疫情" "藥物" "現在"
 [8,] "學習" "觀光"   "肺炎" "北京"     "指數" "人員" "銀行" "新冠" "一個" "他們"
 [9,] "員警" "酒店"   "出現" "川普"     "衝擊" "消毒" "貸款" "案例" "新冠" "我們"
[10,] "平台" "消費者" "通報" "冠狀病毒" "億元" "市長" "計畫" "圖解" "可以" "真的"

lda_model$plot()

To stop the server, run servr::daemon_stop(1) or restart your R session
Serving the directory /tmp/Rtmpu71thE/file7a3f5167e183 at http://127.0.0.1:4321

# lda_model$plot(out.dir ="lda_result", open.browser = TRUE)