LDA in practice

Introduction to Natural Language Processing in R

Kasey Jones

Research Data Scientist

Finalizing LDA results

select the number of topics
- perplexity/other metrics
- a solution that works for your situation

Perplexity

measure of how well a probability model fits new data
lower is better
used to compare models
- In LDA parameter tuning
- Selecting number of topics

sample_size <- floor(0.90 * nrow(doc_term_matrix))
set.seed(1111)
train_ind <- sample(nrow(doc_term_matrix), size = sample_size)
train <- matrix[train_ind, ]
test <- matrix[-train_ind, ]

¹ https://en.wikipedia.org/wiki/Perplexity

Perplexity in R

library(topicmodels)
values = c()
for(i in c(2:35)){
  lda_model <- LDA(train, k = i, method = "Gibbs",
                   control = list(iter = 25, seed = 1111))
  values <- c(values, perplexity(lda_model, newdata = test))  
}

plot(c(2:35), values, main="Perplexity for Topics", 
     xlab="Number of Topics", ylab="Perplexity")

Perplexity again!

Perplexity is a measure of how good our models extends to new data.

Practical selection

How many topics can the situation handle
- 20 might be difficult to cover
How are you displaying the results
- Graphics with 5 topics are easier than graphics with 100 topics
Rules of thumb:
- Use a small number of topics where each topic is represented by several documents
- Large topic counts can be used only if time allows exploring and dissecting each topic

Using results

Review or have reviewers find "themes" for each topic
- provide reviewer with a list of top words in the topic
- provide reviewer with a list of the top documents for that topic

Review output

betas <- tidy(lda_model, matrix = "beta")
betas %>%
  filter(topic == 1) %>%
  arrange(desc(beta)) %>%
  select(term)

# A tibble: 2,000 x 1
   term    
   <chr>   
 1 athletic   
 2 quick 
 3 strong    
 4 tough    
...

gammas <- tidy(lda_model, matrix = "gamma")
gammas %>%
  filter(topic == 1) %>%
  arrange(desc(gamma)) %>%
  select(document)

# A tibble: 1,000 x 1
   document
   <chr>   
 1 232     
 2 292     
 3 921    
 4 643    
 5 468

Summarize output

gammas <- tidy(lda_model, matrix = "gamma")
gammas %>%
  group_by(document) %>%
  arrange(desc(gamma)) %>%
  slice(1) %>%
  group_by(topic) %>% 
  tally(topic, sort=TRUE)

  topic     n
1     1  1326
2     5  1215
3     4   804
...

Summarize output again

gammas %>%
  group_by(document) %>%
  arrange(desc(gamma)) %>%
  slice(1) %>%
  group_by(topic) %>% 
  summarize(avg=mean(gamma)) %>%
  arrange(desc(avg))

  topic   avg
1     1 0.696
2     5 0.530
3     4 0.482
...

LDA practice.

Introduction to Natural Language Processing in R