R LDA Topic Modeling: Result topics contains very similar words

Question

All:

I am beginner in R topic modeling, it all started three weeks ago. So my problem is I can successfully processed my data into corpus, Document term matrix and LDA function. I have tweets as my input and about 460,000 tweets. But I am not happy with the result, the words across all topic are very similar.

packages <- c('tm','topicmodels','SnowballC','RWeka','rJava')
if (length(setdiff(packages, rownames(installed.packages()))) > 0) {
install.packages(setdiff(packages, rownames(installed.packages())))  
}

options( java.parameters = "-Xmx4g" )
library(tm)
library(topicmodels)
library(SnowballC)
library(RWeka)

print("Please select the input file");
flush.console();
ifilename <- file.choose();
raw_data=scan(ifilename,'string',sep="\n",skip=1);

tweet_data=raw_data;
rm(raw_data);
tweet_data = gsub("(RT|via)((?:\\b\\W*@\\w+)+)","",tweet_data)
tweet_data = gsub("http[^[:blank:]]+", "", tweet_data)
tweet_data = gsub("@\\w+", "", tweet_data)
tweet_data = gsub("[ \t]{2,}", "", tweet_data)
tweet_data = gsub("^\\s+|\\s+$", "", tweet_data)
tweet_data = gsub('\\d+', '', tweet_data)
tweet_data = gsub("[[:punct:]]", " ", tweet_data)

max_size=5000;
data_size=length(tweet_data);
itinerary=ceiling(data_size[1]/max_size);
if (itinerary==1){pre_data=tweet_data}else {pre_data=tweet_data[1:max_size]}

corp <- Corpus(VectorSource(pre_data));
corp<-tm_map(corp,tolower);
corp<-tm_map(corp,removePunctuation);
extend_stop_word=c('description:','null','text:','description','url','text','aca',
                   'obama','romney','ryan','mitt','conservative','liberal');
corp<-tm_map(corp,removeNumbers);
gc();
IteratedLovinsStemmer(corp, control = NULL)
gc();
corp<-tm_map(corp,removeWords,c(stopwords('english'),extend_stop_word));
gc();
corp <- tm_map(corp, PlainTextDocument)
gc();
dtm.control = list(tolower= F,removePunctuation=F,removeNumbers= F,
                   stemming= F, minWordLength = 3,weighting= weightTf,stopwords=F)

dtm = DocumentTermMatrix(corp, control=dtm.control)
gc();
#dtm = removeSparseTerms(dtm,0.99)
dtm = dtm[rowSums(as.matrix(dtm))>0,]
gc();

best.model <- lapply(seq(2,50, by=2), function(k){LDA(dtm[1:10,], k)})
gc();
best.model.logLik <- as.data.frame(as.matrix(lapply(best.model, logLik)))
best.model.logLik.df <- data.frame(topics=c(seq(2,50, by=2)), LL=as.numeric(as.matrix(best.model.logLik)))
k=best.model.logLik.df[which.max(best.model.logLik.df$LL),1];
cat("Best topic number is k=",k);
flush.console();
gc();
lda.model = LDA(dtm, k,method='VEM')
gc();
write.csv(terms(lda.model,50), file = "terms.csv");
lda_topics=topics(lda.model,1);

The following is the results I get:

> terms(lda.model,10)
      Topic 1     Topic 2    Topic 3    Topic 4    Topic 5   
 [1,] "taxes"     "medicare" "tax"      "tax"      "jobs"    
 [2,] "pay"       "will"     "returns"  "returns"  "plan"    
 [3,] "welfare"   "tax"      "gop"      "taxes"    "gop"     
 [4,] "will"      "care"     "taxes"    "will"     "military"
 [5,] "returns"   "can"      "abortion" "paul"     "will"    
 [6,] "plan"      "laden"    "can"      "medicare" "tax"     
 [7,] "economy"   "vote"     "tcot"     "class"    "paul"    
 [8,] "budget"    "economy"  "muslim"   "budget"   "campaign"
 [9,] "president" "taxes"    "campaign" "says"     "says"    
[10,] "reid"      "just"     "economy"  "cuts"     "can"     
      Topic 6     Topic 7     Topic 8    Topic 9    
 [1,] "medicare"  "tax"       "medicare" "tax"      
 [2,] "taxes"     "medicare"  "tax"      "president"
 [3,] "plan"      "taxes"     "jobs"     "jobs"     
 [4,] "tcot"      "tcot"      "tcot"     "taxes"    
 [5,] "budget"    "president" "foreign"  "medicare" 
 [6,] "returns"   "jobs"      "plan"     "tcot"     
 [7,] "welfare"   "budget"    "will"     "paul"     
 [8,] "can"       "energy"    "economy"  "health"   
 [9,] "says"      "military"  "bush"     "people"   
[10,] "obamacare" "want"      "now"      "gop"      
      Topic 10    Topic 11   Topic 12  
 [1,] "tax"       "gop"      "gop"     
 [2,] "medicare"  "tcot"     "plan"    
 [3,] "tcot"      "military" "tax"     
 [4,] "president" "jobs"     "taxes"   
 [5,] "gop"       "energy"   "welfare" 
 [6,] "plan"      "will"     "tcot"    
 [7,] "jobs"      "ohio"     "military"
 [8,] "will"      "abortion" "campaign"
 [9,] "cuts"      "paul"     "class"   
[10,] "paul"      "budget"   "just" 

As you can see the words "tax" "medicare" are across all topic. I noticed that while I playing with the dtm = removeSparseTerms(dtm,0.99) the results may changes a little. And the following is my sample input data

> tweet_data[1:10]
 [1] " While  Romney friends get richer  MT  Romney Ryan Economic Plans Would Increase Unemployment Deepen Recession"                 
 [2] "Wayne Allyn Root claims proof of Obama s foreign citizenship  During a radio show interview Resist"                             
 [3] " President Obama  Chief Investor  Leave Energy Upgrades to the Businesses  Reading President Obama誷 latest Execu   "           
 [4] " Brotherhood  starts crucifixions Opponents of Egypt s Muslim president executed  naked on trees   Obama s    tcot"             
 [5] "  Say you stand with President Obama裻he candidate in this election who trusts women to make their own health decisions     "   
 [6] " Romney  Ryan Descend Into Medicare Gibberish "                                                                                 
 [7] "Maddow  Romney demanded opponents tax returns and lied about residency in    The Raw Story"                                     
 [8] "Is it not grand  How can Jews reconcile Obama   Carter s treatment of Jews Israel  How ca    "                                  
 [9] "   The Tax Returns are Hurting Romney  Badly  "                                                                                 
[10] "  Replacing Gen Dempsey is cruicial to US security  Dempsey  disappointed  by anti Obama campaign by ex military members  h    "

Please Help!!Thanks!

Answer 1

Reduce the number of topics in your case. This would enhance the clustering capability of your topic model. Now you are overlapping existing models with another. Since topic index varies over iterations, it is difficult to follow through on the results/ compare too.