###基本介紹 + 目的:使用coreNLP與sentimentr分析twitter上台灣與台積電的文字資料 + 概述:隨著中美貿易站開始，台灣的台積電在全球晶片產業上佔有舉足輕重的地位。希望利用 推特的文章分析了解全球民眾對“台灣及台積電”的認知與相關情緒。 + 資料來源:Twitter，4/3~4/11，5000筆，English

1. coreNLP

安裝package

packages = c("dplyr","ggplot2","rtweet" ,"xml2", "httr", "jsonlite", "data.tree", "NLP", "igraph","sentimentr","tidytext","wordcloud2","DiagrammeR","dplyr")
existing = as.character(installed.packages()[,1])
for(pkg in packages[!(packages %in% existing)]) install.packages(pkg)
library(wordcloud2)
library(ggplot2)
library(scales)
library(rtweet)
library(dplyr)
## 
## Attaching package: 'dplyr'
## The following objects are masked from 'package:stats':
## 
##     filter, lag
## The following objects are masked from 'package:base':
## 
##     intersect, setdiff, setequal, union
library(xml2)
library(httr)
library(jsonlite)
## 
## Attaching package: 'jsonlite'
## The following object is masked from 'package:rtweet':
## 
##     flatten
library(magrittr)
library(data.tree)
library(tidytext)
library(stringr)
library(DiagrammeR)
library(magrittr)
load("coreNLP_all.RData")

1.1 資料收集：tweets

(1). Twitter API設定 透過rtweet抓取tweets

app = 'heminghsin'
consumer_key = 'kjJGO9cTWdoCG9BHgBwfFtcfi'
consumer_secret = 'zjamiuhUuWZjjbsi01Jlg38uwVyXeQpiuFkUDk6QhiSIC379UO'
access_token = '1380751997343698949-qYevxQu1xqqP4dYegz4VeZUzSASi44'
access_secret = 'Lnbft6o7jK2Du8X9087qTFRRhs3UA2coYJSRi4KnrbR9d'
twitter_token <- create_token(app,consumer_key, consumer_secret,
                    access_token, access_secret,set_renv = FALSE)
#Consumer Keys:知道你的身分
#Authentication Tokens:認證給你的授權

(2). 設定關鍵字抓tweets

# 查詢關鍵字
key = c("#Taiwan")
context = "TSMC"
q = paste(c(key,context),collapse=" OR ")   
# 查詢字詞 "#Taiwan AND TSMC"
# 為了避免只下#Taiwan和TSMC同時出現的條件下資料筆數過少選擇OR條件

#抓5000筆 不抓轉推
tweets = search_tweets(q,lang="en",n=5000,include_rts = FALSE,token = twitter_token)

(3). tweets內容清理

## 用於資料清理
clean = function(txt) {
  txt = iconv(txt, "latin1", "ASCII", sub="") #改變字的encoding
  txt = gsub("(@|#)\\w+", "", txt) #去除@或#後有數字,字母,底線 (標記人名或hashtag)
  txt = gsub("(http|https)://.*", "", txt) #去除網址(.:任意字元，*:0次以上)
  txt = gsub("[ \t]{2,}", "", txt) #去除兩個以上空格或tab
  txt = gsub("\\n"," ",txt) #去除換行
  txt = gsub("\\s+"," ",txt) #去除一個或多個空格(+:一次以上)
  txt = gsub("^\\s+|\\s+$","",txt) #去除開頭/結尾有一個或多個空格
  txt = gsub("&.*;","",txt) #去除html特殊字元編碼
  txt = gsub("[^a-zA-Z0-9?!. ']","",txt) #除了字母,數字空白?!.的都去掉(表情符號去掉)
  txt }


tweets$text = clean(tweets$text)  #text套用資料清理

df = data.frame()
  
df = rbind(df,tweets)  # transfer to data frame

df = df[!duplicated(df[,"status_id"]),]  #去除重複的tweets
head(df)
## # A tibble: 6 x 90
##   user_id   status_id   created_at          screen_name  text            source 
##   <chr>     <chr>       <dttm>              <chr>        <chr>           <chr>  
## 1 12640270… 1381643805… 2021-04-12 16:22:07 hongkongers… It is importan… Twitte…
## 2 12640270… 1380470283… 2021-04-09 10:38:58 hongkongers… will continue … Twitte…
## 3 12640270… 1380698251… 2021-04-10 01:44:49 hongkongers… is a responsib… Twitte…
## 4 12640270… 1380209386… 2021-04-08 17:22:15 hongkongers… is consistentl… Twitte…
## 5 12640270… 1381563373… 2021-04-12 11:02:31 hongkongers… Agree with .an… Twitte…
## 6 12640270… 1380145360… 2021-04-08 13:07:50 hongkongers… The Train Cras… Twitte…
## # … with 84 more variables: display_text_width <dbl>,
## #   reply_to_status_id <chr>, reply_to_user_id <chr>,
## #   reply_to_screen_name <chr>, is_quote <lgl>, is_retweet <lgl>,
## #   favorite_count <int>, retweet_count <int>, quote_count <int>,
## #   reply_count <int>, hashtags <list>, symbols <list>, urls_url <list>,
## #   urls_t.co <list>, urls_expanded_url <list>, media_url <list>,
## #   media_t.co <list>, media_expanded_url <list>, media_type <list>,
## #   ext_media_url <list>, ext_media_t.co <list>, ext_media_expanded_url <list>,
## #   ext_media_type <chr>, mentions_user_id <list>, mentions_screen_name <list>,
## #   lang <chr>, quoted_status_id <chr>, quoted_text <chr>,
## #   quoted_created_at <dttm>, quoted_source <chr>, quoted_favorite_count <int>,
## #   quoted_retweet_count <int>, quoted_user_id <chr>, quoted_screen_name <chr>,
## #   quoted_name <chr>, quoted_followers_count <int>,
## #   quoted_friends_count <int>, quoted_statuses_count <int>,
## #   quoted_location <chr>, quoted_description <chr>, quoted_verified <lgl>,
## #   retweet_status_id <chr>, retweet_text <chr>, retweet_created_at <dttm>,
## #   retweet_source <chr>, retweet_favorite_count <int>,
## #   retweet_retweet_count <int>, retweet_user_id <chr>,
## #   retweet_screen_name <chr>, retweet_name <chr>,
## #   retweet_followers_count <int>, retweet_friends_count <int>,
## #   retweet_statuses_count <int>, retweet_location <chr>,
## #   retweet_description <chr>, retweet_verified <lgl>, place_url <chr>,
## #   place_name <chr>, place_full_name <chr>, place_type <chr>, country <chr>,
## #   country_code <chr>, geo_coords <list>, coords_coords <list>,
## #   bbox_coords <list>, status_url <chr>, name <chr>, location <chr>,
## #   description <chr>, url <chr>, protected <lgl>, followers_count <int>,
## #   friends_count <int>, listed_count <int>, statuses_count <int>,
## #   favourites_count <int>, account_created_at <dttm>, verified <lgl>,
## #   profile_url <chr>, profile_expanded_url <chr>, account_lang <lgl>,
## #   profile_banner_url <chr>, profile_background_url <chr>,
## #   profile_image_url <chr>

df共有90個欄位，但我們在這裡僅會使用幾個欄位:

user_id: 用戶id status_id : 推文id created_at : 發文時間 text : 推文內容 source : 發文來源

了解資料的資料筆數以及時間分布

created_at已經是一個date類型的欄位，因此可以直接用min,max來看最遠或最近的日期 註:rtweet最多只能抓到距今10天的資料

nrow(df)
## [1] 4999
min(df$created_at)
## [1] "2021-04-05 14:22:40 UTC"
max(df$created_at)
## [1] "2021-04-12 16:22:07 UTC"

1.2串接CoreNLP API

(1). API呼叫的設定

server端 : + 需先在terminal開啟corenlp server + 在corenlp的路徑下開啟terminal輸入 java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 15000

# 產生coreNLP的api url，將本地端的網址轉成符合coreNLP服務的url
generate_API_url <- function(host, port="9000",
                    tokenize.whitespace="false", annotators=""){ #斷詞依據不是空格
    url <- sprintf('http://%s:%s/?properties={"tokenize.whitespace":"%s","annotators":"%s"}', host, port, tokenize.whitespace, annotators)
    url <- URLencode(url)
}
#指定服務的位置
host = "127.0.0.1"

generate_API_url(host)
# 呼叫coreNLP api
call_coreNLP <- function(server_host, text, host="localhost", language="eng",
                    tokenize.whitespace="true", ssplit.eolonly="true", annotators=c("tokenize","ssplit","pos","lemma","ner","parse","sentiment")){
  # 假設有兩個core-nlp server、一個負責英文（使用9000 port）、另一個則負責中文（使用9001 port）
  port <- ifelse(language=="eng", 9000, 9001);
  # 產生api網址
  url <- generate_API_url(server_host, port=port,
                    tokenize.whitespace=tokenize.whitespace, annotators=paste0(annotators, collapse = ','))
  
  result <- POST(url, body = text, encode = "json")
  doc <- httr::content(result, "parsed","application/json",encoding = "UTF-8")
  return (doc)
}
#文件使用coreNLP服務
coreNLP <- function(data,host){
  # 依序將每個文件丟進core-nlp進行處理，每份文件的回傳結果為json格式
  # 在R中使用objects來儲存處理結果
  result <- apply(data, 1 , function(x){
    object <- call_coreNLP(host, x['text'])
    list(doc=object, data=x)
  })
  
  return(result)
}

(2). 資料整理function

從回傳的object中整理斷詞出結果，輸出為 tidydata 格式

coreNLP_tokens_parser <- function(coreNLP_objects){
  
  result <- do.call(rbind, lapply(coreNLP_objects, function(obj){
    original_data <- obj$data
    doc <- obj$doc
    # for a sentences
    sentences <- doc$sentences
   
    sen <- sentences[[1]]
    
    tokens <- do.call(rbind, lapply(sen$tokens, function(x){
      result <- data.frame(word=x$word, lemma=x$lemma, pos=x$pos, ner=x$ner)
      result
    }))
    
    tokens <- original_data %>%
      t() %>% 
      data.frame() %>% 
      select(-text) %>% 
      slice(rep(1:n(), each = nrow(tokens))) %>% 
      bind_cols(tokens)
    
    tokens
  }))
  return(result)
}

從回傳的core-nlp object中整理出詞彙依存關係，輸出為 tidydata 格式

coreNLP_dependency_parser <- function(coreNLP_objects){
  result <- do.call(rbind, lapply(coreNLP_objects, function(obj){
    original_data <- obj$data
    doc <- obj$doc
    # for a sentences
    sentences <- doc$sentences
    sen <- sentences[[1]]
    dependencies <- do.call(rbind, lapply(sen$basicDependencies, function(x){
      result <- data.frame(dep=x$dep, governor=x$governor, governorGloss=x$governorGloss, dependent=x$dependent, dependentGloss=x$dependentGloss)
      result
    }))
  
    dependencies <- original_data %>%
      t() %>% 
      data.frame() %>% 
      select(-text) %>% 
      slice(rep(1:n(), each = nrow(dependencies))) %>% 
      bind_cols(dependencies)
    dependencies
  }))
  return(result)
}

從回傳的core-nlp object中整理出語句情緒，輸出為 tidydata 格式

coreNLP_sentiment_parser <- function(coreNLP_objects){
  result <- do.call(rbind, lapply(coreNLP_objects, function(obj){
    original_data <- obj$data
    doc <- obj$doc
    # for a sentences
    sentences <- doc$sentences
    sen <- sentences[[1]]
    
    sentiment <- original_data %>%
      t() %>% 
      data.frame() %>% 
      bind_cols(data.frame(sentiment=sen$sentiment, sentimentValue=sen$sentimentValue))
  
    sentiment
  }))
  return(result)
}

圖形化 Dependency tree

程式參考來源：https://stackoverflow.com/questions/35496560/how-to-convert-corenlp-generated-parse-tree-into-data-tree-r-package

# 圖形化顯示dependency結果
parse2tree <- function(ptext) {
  stopifnot(require(NLP) && require(igraph))
  
  # this step modifies coreNLP parse tree to mimic openNLP parse tree
  ptext <- gsub("[\r\n]", "", ptext)
  ptext <- gsub("ROOT", "TOP", ptext)


  ## Replace words with unique versions
  ms <- gregexpr("[^() ]+", ptext)                                      # just ignoring spaces and brackets?
  words <- regmatches(ptext, ms)[[1]]                                   # just words
  regmatches(ptext, ms) <- list(paste0(words, seq.int(length(words))))  # add id to words
  
  ## Going to construct an edgelist and pass that to igraph
  ## allocate here since we know the size (number of nodes - 1) and -1 more to exclude 'TOP'
  edgelist <- matrix('', nrow=length(words)-2, ncol=2)
  
  ## Function to fill in edgelist in place
  edgemaker <- (function() {
    i <- 0                                       # row counter
    g <- function(node) {                        # the recursive function
      if (inherits(node, "Tree")) {            # only recurse subtrees
        if ((val <- node$value) != 'TOP1') { # skip 'TOP' node (added '1' above)
          for (child in node$children) {
            childval <- if(inherits(child, "Tree")) child$value else child
            i <<- i+1
            edgelist[i,1:2] <<- c(val, childval)
          }
        }
        invisible(lapply(node$children, g))
      }
    }
  })()
  
  ## Create the edgelist from the parse tree
  edgemaker(Tree_parse(ptext))
  tree <- FromDataFrameNetwork(as.data.frame(edgelist))
  return (tree)
}

將句子丟入服務

取得coreNLP回傳的物件 先不要跑這段，會花大概12分鐘（如果你記憶體只有4G可能會當掉）

gc() #釋放不使用的記憶體
##            used  (Mb) gc trigger  (Mb) limit (Mb) max used  (Mb)
## Ncells  6098453 325.7   11495671 614.0         NA  7025696 375.3
## Vcells 17715543 135.2   23351897 178.2      16384 19098245 145.8
t0 = Sys.time()
obj = df[,c(2,5)]  %>% filter(text != "") %>% coreNLP(host) #丟入本地執行
#丟入coreNLP的物件 必須符合: 是一個data.frame 有一個text欄位

Sys.time() - t0 #執行時間
## Time difference of 13.75128 mins
#Time difference of 12 mins

save.image("coreNLP_all.RData")
#先將會用到的東西存下來，要用可直接載RData
#tokens =  coreNLP_tokens_parser(obj)
#dependencies = coreNLP_dependency_parser(obj)
#sentiment = coreNLP_sentiment_parser(obj)
#save.image("coreNLP_all.RData")

1.3 提取結果

(1). 斷詞、詞彙還原、詞性標註、NER

tokens =  coreNLP_tokens_parser(obj)
head(tokens,20)
##              status_id        word       lemma pos    ner
## 1  1378792114037870600       Simon       Simon NNP PERSON
## 2  1378792114037870600      Parkes      Parkes NNP PERSON
## 3  1378792114037870600      Update      Update NNP      O
## 4  1378792114037870600        With        with  IN      O
## 5  1378792114037870600       great       great  JJ      O
## 6  1378792114037870600      regret      regret  NN      O
## 7  1378792114037870600           I           I PRP      O
## 8  1378792114037870600        must        must  MD      O
## 9  1378792114037870600    announce    announce  VB      O
## 10 1378792114037870600        that        that  IN      O
## 11 1378792114037870600    children       child NNS      O
## 12 1378792114037870600        were          be VBD      O
## 13 1378792114037870600       being          be VBG      O
## 14 1378792114037870600 transported   transport VBN      O
## 15 1378792114037870600          in          in  IN      O
## 16 1378792114037870600       cargo       cargo  NN      O
## 17 1378792114037870600 containers. containers.  NN      O
## 18 1378792114037870600         The         the  DT      O
## 19 1378792114037870600   operation   operation  NN      O
## 20 1378792114037870600          in          in  IN      O

coreNLP_tokens_parser欄位: status_id : 對應原本df裡的status_id，為一則tweets的唯一id word: 原始斷詞 lemma : 對斷詞做詞形還原 pos : part-of-speech,詞性 ner: 命名實體

(2). 命名實體標註(NER)

從NER查看特定類型的實體，辨識出哪幾種類型

unique(tokens$ner)
##  [1] "PERSON"            "O"                 "CITY"             
##  [4] "LOCATION"          "NUMBER"            "ORDINAL"          
##  [7] "ORGANIZATION"      "TITLE"             "MISC"             
## [10] "DATE"              "DURATION"          "MONEY"            
## [13] "PERCENT"           "COUNTRY"           "TIME"             
## [16] "NATIONALITY"       "SET"               "STATE_OR_PROVINCE"
## [19] "CAUSE_OF_DEATH"    "CRIMINAL_CHARGE"   "IDEOLOGY"         
## [22] "RELIGION"

#除去entity為Other，有多少種word有被標註entity

length(unique(tokens$word[tokens$ner != "O"]))
## [1] 1880

(3). 轉小寫

因為大小寫也會影響corenlp對NER的判斷，因此我們一開始給的推文內容是沒有處理大小寫的，但在跑完anotator後，為了正確計算詞頻，創建新欄位lower_word與lower_lemma，存放轉換小寫的word與lemma。轉成小寫的目的是要將不同大小寫的同一字詞（如Evergiven與evergiven）都換成小寫，再來計算詞頻

tokens$lower_word = tolower(tokens$word)
tokens$lower_lemma = tolower(tokens$lemma)

1.4 探索分析 - NER

####涉及到的國家(COUNTRY) 我們可以透過coreNLP中的NER解析出在Twitter上面談論台灣的台積電，所涉及到的國家(COUNTRY)，以初步了解這個議題的主要國家。

tokens %>%
  filter(ner == "COUNTRY") %>%  #篩選NER為COUNTRY
  group_by(lower_word) %>% #根據word分組
  summarize(count = n()) %>% #計算每組
  top_n(n = 13, count) %>%
  ungroup() %>% 
  mutate(word = reorder(lower_word, count)) %>%
  ggplot(aes(word, count)) + 
  geom_col()+
  ggtitle("Word Frequency (NER is COUNTRY)") +
  theme(text=element_text(size=14))+
  coord_flip()

  • 台積電讓Taiwan讓更多人知道這個國家
  • 台積電在USA.China貿易戰中有舉足輕重的地位

####涉及到的組織(ORGANIZATION) 我們可以透過coreNLP中的NER解析出在Twitter上面談論台灣的台積電，所涉及到的組織(ORGANIZATION)，以初步了解這個議題的主要公司/單位。

tokens %>%
  filter(ner == "ORGANIZATION") %>%  #篩選NER為ORGANIZATION
  group_by(lower_word) %>% #根據word分組
  summarize(count = n()) %>% #計算每組
  top_n(n = 10, count) %>%
  ungroup() %>% 
  mutate(word = reorder(lower_word, count)) %>%
  ggplot(aes(word, count)) + 
  geom_col()+
  ggtitle("Word Frequency (NER is ORGANIZATION)") +
  theme(text=element_text(size=14))+
  coord_flip()

  • 英代爾Intel、三星Samsung是台積電的競爭對象
  • 蘋果公司Apple是台積電的重要客戶

涉及到的人物(PERSON)

我們可以透過coreNLP中的NER解析出在Twitter上面談論台灣的台積電，所涉及到的人物(PERSON)，以初步了解這個議題的主要人物。

tokens %>%
  filter(ner == "PERSON") %>%  #篩選NER為PERSON
  group_by(lower_word) %>% #根據word分組
  summarize(count = n()) %>% #計算每組
  top_n(n = 20, count) %>%
  ungroup() %>% 
  mutate(word = reorder(lower_word, count)) %>%
  ggplot(aes(word, count)) + 
  geom_col()+
  ggtitle("Word Frequency (NER is PERSON)") +
  theme(text=element_text(size=14))+
  coord_flip()

  • biden:現任美國總統
  • john推測為美國半導體產業協會執行長John Neuffer

1.5 探索分析 - Dependency

語句依存關係結果
dependencies = coreNLP_dependency_parser(obj)
head(dependencies,20)
##              status_id        dep governor governorGloss dependent
## 1  1378792114037870600       ROOT        0          ROOT        26
## 2  1378792114037870600   compound        3        Update         1
## 3  1378792114037870600   compound        3        Update         2
## 4  1378792114037870600   obl:tmod       26       success         3
## 5  1378792114037870600       case        6        regret         4
## 6  1378792114037870600       amod        6        regret         5
## 7  1378792114037870600        obl        9      announce         6
## 8  1378792114037870600      nsubj        9      announce         7
## 9  1378792114037870600        aux        9      announce         8
## 10 1378792114037870600  acl:relcl        3        Update         9
## 11 1378792114037870600       mark       14   transported        10
## 12 1378792114037870600 nsubj:pass       14   transported        11
## 13 1378792114037870600        aux       14   transported        12
## 14 1378792114037870600   aux:pass       14   transported        13
## 15 1378792114037870600      ccomp        9      announce        14
## 16 1378792114037870600       case       17   containers.        15
## 17 1378792114037870600   compound       17   containers.        16
## 18 1378792114037870600        obl       14   transported        17
## 19 1378792114037870600        det       19     operation        18
## 20 1378792114037870600      nsubj       26       success        19
##    dependentGloss
## 1         success
## 2           Simon
## 3          Parkes
## 4          Update
## 5            With
## 6           great
## 7          regret
## 8               I
## 9            must
## 10       announce
## 11           that
## 12       children
## 13           were
## 14          being
## 15    transported
## 16             in
## 17          cargo
## 18    containers.
## 19            The
## 20      operation
視覺化 Dependency tree
parse_tree <- obj[[113]]$doc[[1]][[1]]$parse
tree <- parse2tree(parse_tree)
SetNodeStyle(tree, style = "filled,rounded", shape = "box")
plot(tree)

1.6 探索分析 - Sentiment

語句情緒值

情緒分數從最低分0~最高分4
+ 0,1 : very negative,negative
+ 2 : neutral
+ 3,4 : very positive,postive

sentiment = coreNLP_sentiment_parser(obj)
head(sentiment,20)
##              status_id
## 1  1378792114037870600
## 2  1376322876274532359
## 3  1378773153304956930
## 4  1378312805724602368
## 5  1378782399681589249
## 6  1378782341825372160
## 7  1378780519807193093
## 8  1376524772608196609
## 9  1376616766047186951
## 10 1376256378986201088
## 11 1378772889982410766
## 12 1378772308693688321
## 13 1378405661877403649
## 14 1375905467482931202
## 15 1376599080198103045
## 16 1376487077441855491
## 17 1378564346498928642
## 18 1376294537027584000
## 19 1376352746593456129
## 20 1376563871012507649
##                                                                                                                                                                                                     text
## 1  Simon Parkes Update With great regret I must announce that children were being transported in cargo containers. The operation in the Suez Canal was a success and warrants me to do a video update...
## 2                                                                                                                                                      Traffic jam at the Suez Canal. About 350 vessels.
## 3                                                                                                                                             To be clear The USS Dwight D. Eisenhower Aircraft Carrier.
## 4                                                                                                                                                                                                 Update
## 5                                                                                                                                                SEUZ CANAL BLOCKAGE  HOW SHIP FROM SEUZ CANAL IN DETAIL
## 6                                                                                                                                                SEUZ CANAL BLOCKAGE  HOW SHIP FROM SEUZ CANAL IN DETAIL
## 7                                 Who knows why the Asia Ruby III is still waiting to enter the? She was tugged out after thewas stranded and normally should have been one of the first vessels to pass
## 8                                                                                                                                                                                 Ever Given is moving !
## 9                                                                                                                        Good to see that Suez Canal Authorities give priority to vessels with livestock
## 10                                                                                                                                                  Yes ! The Dutch have arrived !! Salvation is near !!
## 11                                                                                                                    Can I name mine the Suez Canal because she's too shallow for the ? I hope it's not
## 12  Do you miss the fun ofblocking the Suez? The last time it happened was even weirder I really loved reading about the Great Bitter Lake Association via  this piece onis wonderful and full of photos
## 13                                                                                                                 The owner of the shipfiles a lawsuit against its operator for delinquency in theCanal
## 14                                                                                                                    TheAuthority reveals the case of.. and explains the story of the bulldozers photos
## 15                                                                                                              International newspapers after the float one of the largest rescue operations in history
## 16                                                                                                                                             How the shipgot stuck in thethrough satellite A new video
## 17                                                                                                                 The owner of the shipfiles a lawsuit against its operator for delinquency in theCanal
## 18                                                                                                                                       How the shipgot stuck in thethrough satellite A very new video 
## 19                                                                                                                                                  How the shipgot stuck in thethroughA very new video 
## 20                                                                                                                                                        The joy of the crew of theship as it exits the
##    sentiment sentimentValue
## 1   Positive              3
## 2    Neutral              2
## 3    Neutral              2
## 4    Neutral              2
## 5    Neutral              2
## 6    Neutral              2
## 7   Negative              1
## 8   Positive              3
## 9    Neutral              2
## 10  Positive              3
## 11  Negative              1
## 12  Positive              3
## 13   Neutral              2
## 14   Neutral              2
## 15   Neutral              2
## 16   Neutral              2
## 17   Neutral              2
## 18   Neutral              2
## 19  Positive              3
## 20   Neutral              2
資料集中的情緒種類
unique(sentiment$sentiment)
## [1] "Positive"     "Neutral"      "Negative"     "Verypositive" "Verynegative"
sentiment$sentimentValue = sentiment$sentimentValue %>% as.numeric
#了解情緒文章的分佈
sentiment$sentiment %>% table()
## .
##     Negative      Neutral     Positive Verynegative Verypositive 
##         1180         2875          869            2           12
平均情緒分數時間趨勢
df$date = as.Date(df$created_at)

sentiment %>% 
  merge(df[,c("status_id","source","date")]) %>%
  group_by(date) %>% 
  summarise(avg_sentiment = mean(sentimentValue,na.rm=T)) %>% 
  ggplot(aes(x=date,y=avg_sentiment)) + 
  geom_line()

不同用戶端情緒時間趨勢
sentiment %>% 
  merge(df[,c("status_id","source","date")]) %>%
  filter(source %in% c("Twitter Web Client","Twitter for iPhone","Twitter for Android")) %>% 
  group_by(date,source) %>% 
  summarise(avg_sentiment = mean(sentimentValue,na.rm=T)) %>% 
  ggplot(aes(x=date,y=avg_sentiment,color=source)) + 
  geom_line()
## `summarise()` has grouped output by 'date'. You can override using the `.groups` argument.

了解情緒分佈，以及在正面情緒及負面情緒下，所使用的文章詞彙為何?
#了解正面文章的詞彙使用
sentiment %>% 
  merge(tokens) %>% 
  anti_join(stop_words) %>% 
  filter(!lower_word %in% c('i','the')) %>% 
  filter(sentiment == "Verypositive" | sentiment =='Positive') %>%
  group_by(lower_lemma) %>% #根據lemma分組
  summarize(count = n()) %>% 
  filter(count >5 & count<400)%>%
  wordcloud2()
## Joining, by = "word"
#了解負面文章的詞彙使用
sentiment %>% 
  merge(tokens) %>% 
  anti_join(stop_words) %>% 
  filter(!lower_word %in% c('i','the')) %>% 
  filter(sentiment == "Verynegative" | sentiment =='Negative') %>%
  group_by(lower_lemma) %>% 
  summarize(count = n()) %>% 
  filter(count >10 &count<400)%>%
  wordcloud2()

“wordcloud”

2. Sentimentr 英文情緒分析

2.1 簡介sentimentr

library(sentimentr)

mytext <- c(
    'do you like it?  But I hate really bad dogs',
    'I am the best friend.',
    'Do you really like it?  I\'m not a fan'
)

mytext <- get_sentences(mytext) #物件，將character向量轉成list,list裡放著character向量(已斷句)
每個文本的情緒分數

情緒分數為-1~1之間，<0屬於負面，>0屬於正面，0屬於中性

sentiment_by(mytext) #document level
##    element_id word_count       sd ave_sentiment
## 1:          1         10 1.497465    -0.8088680
## 2:          2          5       NA     0.5813777
## 3:          3          9 0.284605     0.2196345
每個句子的情緒分數
sentiment(mytext) #sentence level
##    element_id sentence_id word_count  sentiment
## 1:          1           1          4  0.2500000
## 2:          1           2          6 -1.8677359
## 3:          2           1          5  0.5813777
## 4:          3           1          5  0.4024922
## 5:          3           2          4  0.0000000
  • 回傳4個欄位的dataframe:
    • element_id – 第幾個文本
    • sentence_id – 該文本中的第幾個句子
    • word_count – 句子字數
    • sentiment – 句子的情緒分數

2.2 使用twitter資料實踐在sentimentr

計算tweet中屬於正面的字
set.seed(10)
mytext <- get_sentences(tweets$text) #將text轉成list of characters型態
x <- sample(tweets$text, 1000, replace = FALSE) #隨機取1000筆，取後不放回
sentiment_words <- extract_sentiment_terms(x) #抓取其中帶有情緒的字
sentiment_counts <- attributes(sentiment_words)$counts #計算出現次數
sentiment_counts[polarity > 0,]   #正面的字
##            words polarity n
##   1:      please      1.0 9
##   2:      wonder      1.0 6
##   3:         top      1.0 5
##   4:   heartfelt      1.0 3
##   5:    welcomes      1.0 3
##  ---                       
## 528: cooperation      0.1 1
## 529:     markets      0.1 1
## 530:     develop      0.1 1
## 531:    thriller      0.1 1
## 532:       peaks      0.1 1
計算tweet中屬於負面的字
sentiment_counts[polarity < 0,] %>% arrange(desc(n)) %>% top_n(10) #出現次數最多的負面字
## Selecting by n
##        words polarity  n
##  1: shortage    -0.75 38
##  2:      war    -0.50 35
##  3:    fight    -0.50 33
##  4:  drought    -0.50 31
##  5:    force    -0.60 21
##  6:  foreign    -0.25 19
##  7: problems    -0.50 19
##  8:   demand    -0.50 17
##  9: abducted    -1.00 17
## 10:     hell    -0.80 16
highlight每個句子，判斷屬於正/負面
set.seed(12)
df%>%
    filter(status_id %in% sample(unique(status_id), 30)) %>% #隨機30筆貼文
    mutate(review = get_sentences(text)) %$% 
    sentiment_by(review, status_id) %>%
    highlight()
## Saved in /var/folders/kc/5zttnvj52nvdl3ckb52dvjnh0000gn/T//RtmpRjL5ho/polarity.html
## Opening /var/folders/kc/5zttnvj52nvdl3ckb52dvjnh0000gn/T//RtmpRjL5ho/polarity.html ...

2.3 用日期來了解情緒波動

code 參考 https://github.com/trinker/sentimentr

tweets$date = format(tweets$created_at,'%Y%m%d')

(out  = tweets  %>%  with(
    sentiment_by( #document level
        get_sentences(text), 
        list( date)
    )
))
plot(out)

2.4 用日期來了解不同用戶端情的緒波動

(out  = tweets %>% filter(source %in% c("Twitter Web Client","Twitter for iPhone","Twitter for Android")) %>%  with(
    sentiment_by(
        get_sentences(text), 
        list(source, date)
    )
))
plot(out)

轉換Emoji代碼為語意文字
replace_emoji("\U0001f4aa")
## [1] " flexed biceps "

總結

練習心得

  1. 練習如何申請Twitter API申請
  2. 抓取資料時的條件設定，如AND或OR會影響資料的多寡
  3. Windows與Mac系統不同，故CoreNLP伺服器設定方式也有差異
  4. 從中發現CoreNLP伺服器可以使用本地或文字平台的服務，但分析結果也會有差異

coreNLP

  1. 找出議題核心人物，組織，國家
  2. 用句法學的分析找出句子相依關係
  3. 分別找出正、負面文章的常用字

sentimentr

  1. 找到tweets中正負面的詞，並且計算每個文本中屬於正負面的句子有哪些
  2. 根據日期知道情緒的波動、不同用戶端的波動