(Somewhat) advanced NLP: text vectorization

### Load standardpackages
library(tidyverse) # Collection of all the good stuff like dplyr, ggplot2 ect.
library(magrittr) # For extra-piping operators (eg. %<>%)

library(tidytext)

This session

This session, we will

1. Review NLP workflows and data structures in R
2. Explore different type of DTM matrix type vector representations of text.
3. Add different types of dimensionality reduction techniques to the repertoir.
4. HAve a peak into word-embeddings
5. Add some goddies on top

Refresher:

Bag of words model

• In order for a computer to understand text we need to somehow find a useful representation.
• If you need to compare different texts e.g. articles, you will probably go for keywords. These keywords may come from a keyword-list with for example 200 different keywords
• In that case you could represent each document with a (sparse) vector with 1 for “keyword present” and 0 for “keyword absent”
• We can also get a bit more sophoistocated and count the number of times a word from our dictionary occurs.
• For a corpus of documents that would give us a document-term matrix.

example

Let’s try creating a bag of words model from our initial example.

text <- tibble(id = c(1:6),
               text = c('A text about cats.',
                        'A text about dogs.',
                        'And another text about a dog.',
                        'Why always writing about cats and dogs, always dogs?',
                        'There are too little text about cats but to many about dogs',
                        'Cats, cats, cats! I love cats soo much. Cats are way better than dogs'))

text_tidy <- text %>% 
  unnest_tokens(word, text, token = 'words') %>% 
  count(id, word)

The document-term matrix (DTM)

• The simplest form of vector representation of text is a ddocument-term matrix
• How to we get a document-term matrix now?
• We could do it by hand, with well-known dplyr syntax (Note: only works when you have one row per unique document-word pair)

text_tidy %>%
  pivot_wider(names_from = word, values_from = n, values_fill = 0)

• We could also use cast_dtm() to create a DTM in the format of the tm package.

text_dtm <- text_tidy %>%
  cast_dtm(id, word, n)

text_dtm 

<<DocumentTermMatrix (documents: 6, terms: 25)>>
Non-/sparse entries: 42/108
Sparsity           : 72%
Maximal term length: 7
Weighting          : term frequency (tf)

• We can simply convert ig to a tibble. Since there exists no direct transfer function, we have to first transform it to a matrix.
• Notice how we recover the rownames

text_dtm %>% as.matrix() %>% as_tibble(rownames = 'id') 

• Sidenote: We can also tidy the DTM again to a tidy token-dataframe.

text_dtm %>% tidy()

• We also can directly use a similar function to cast a sparse matrix (which we for sure then also could transform to a tibble again)

text_tidy %>% cast_sparse(row = id, column = word, value = n)

6 x 25 sparse Matrix of class "dgCMatrix"
                                                   
1 1 1 1 1 . . . . . . . . . . . . . . . . . . . . .
2 1 1 . 1 1 . . . . . . . . . . . . . . . . . . . .
3 1 1 . 1 . 1 1 1 . . . . . . . . . . . . . . . . .
4 . 1 1 . 2 1 . . 2 1 1 . . . . . . . . . . . . . .
5 . 2 1 1 1 . . . . . . 1 1 1 1 1 1 1 . . . . . . .
6 . . 5 . 1 . . . . . . 1 . . . . . . 1 1 1 1 1 1 1

• Finally, we could just apply a text recipe here

library(recipes)
library(textrecipes)

text %>%
  recipe(~.) %>% 
  step_tokenize(text, token = 'words') %>% # tokenize
  step_tf(text) %>% # TFIDF weighting
  prep() %>% juice()

TF-IDF - Term Frequency - Inverse Document Frequency

• A token is important for a document if appears very often
• A token becomes less important for comparison across a corpus if it appears all over the place in the corpus
• Cat in a corpus of websites talking about cats is not that important

\[w_{i,j} = tf_{i,j}*log(\frac{N}{df_i})\]

• \(w_{i,j}\) = the TF-IDF score for a term i in a document j
• \(tf_{i,j}\) = number of occurence of term i in document j
• \(N\) = number of documents in the corpus
• \(df_i\) = number of documents with term i

# TFIDF weights
text_tidy %<>%
  bind_tf_idf(term = word,
              document = id,
              n = n)

• We obviously could also cast a tf_idf weighted dtm…

text_tidy %>%
  select(id, word, tf_idf) %>%
  pivot_wider(names_from = word, values_from = tf_idf, values_fill = 0)

• btw: this is equivalent to just running a textrecipe like that:

text %>%
  recipe(~.) %>% 
  step_tokenize(text, token = 'words') %>% # tokenize
  step_tfidf(text) %>% # TFIDF weighting
  prep() %>% juice()

• Sidenote, when we use a POS engine such as spacyr for tokenization, we can also add recipes for lematization, filter for POS etc.

text %>%
  recipe(~.) %>% 
  step_tokenize(text, engine = "spacyr") %>%
  step_pos_filter(text, keep_tags = "NOUN") %>%
  step_lemma(text) %>%
  step_tf(text) %>%
  prep() %>%
  juice()

• A last reminder on the powerful pairwise_xx() functions from the widyr package
• For instance, pairwise similarities/distances

library(widyr)

text_tidy %>% pairwise_dist(id, word, tf_idf, method = "manhattan") %>%
  mutate(similarity = 1 - (distance / max(distance)) ) %>%
  select(-distance) %>%
  arrange(desc(similarity))

Dimensionality reduction techniques

rm(list=ls())

• Ok, lets get first some more interesting data. We will work with the CORDIS project descriptions of EU Horizon 2020 projects again.

text <- read_csv('https://github.com/SDS-AAU/SDS-master/raw/master/M2/data/cordis-h2020reports.gz')

colnames(text) <- colnames(text) %>% str_to_lower()
text %<>%
  select(-x1) %>%
  rename(id = projectid) %>%
  relocate(id) %>%
  filter(language == 'en') %>%
  drop_na(id)

• Lets create a tidy tokenlist

text_tidy <- text %>%
  rename(text = summary) %>%
  select(id, text) %>%
  unnest_tokens(word, text, token = "words")

• some preprocessing

# preprocessing
text_tidy %<>%
  filter(str_length(word) > 2 ) %>% # Remove words with less than  3 characters
  filter(!(word %in% c('project', 'research'))) %>%
  anti_join(stop_words, by = 'word') 

• We can also ad bigrams

text_tidy %<>%
  unnest_tokens(word, word, token = 'ngrams', n = 2, n_min = 1) %>%
  group_by(word) %>% filter(n() > 25) %>% ungroup() 

text_tidy %>%
  count(word, sort = TRUE)

• Lets finish this up and also add TF-IDF weights

text_tidy %<>%
  count(id, word) %>%
  bind_tf_idf(term = word,
              document = id,
              n = n) %>%
  select(-tf, -idf)

• Is there a big difference?

text_tidy %>%
  count(word, wt = tf_idf, sort = TRUE)

• And finally, lets get a DTM dataframe

text_dtm <- text_tidy %>%
  select(id, word, n) %>%
  pivot_wider(names_from = word, values_from = n, values_fill = 0)

• And, just in case, a TFIDF weighted version

• We could also prepare a recipe which doe pretty much the same…

recipe_base <- text %>%
  rename(text = summary) %>%
  select(id, text) %>%
  # BAse recipe starts
  recipe(~.) %>% 
  update_role(id, new_role = "id") %>% # Update role of ID
  step_tokenize(text, token = 'words') %>% # tokenize
  step_stopwords(text, keep = FALSE) %>% # remove stopwords
  step_untokenize(text) %>% # Here we now have to first untokenize
  step_tokenize(text, token = "ngrams", options = list(n = 1, n_min = 1)) %>% # and tokenize again
  step_tokenfilter(text, min_times = 25) 

• Sidenote

• Here, we can further preprocess to do whatever we would like, such as obtaining a dtm

recipe_base %>% 
  step_tf(text) %>% 
  prep() %>% 
  juice() %>% 
  head(100)

text_pca <- text_dtm %>% 
  column_to_rownames('id') %>% 
  prcomp(center = TRUE, scale. = TRUE, rank. = 10)

text_pca %>% glimpse()

List of 5
 $ sdev    : num [1:499] 3.54 3.05 2.86 2.8 2.66 ...
 $ rotation: num [1:599, 1:10] 0.01561 0.00169 0.07306 -0.03103 0.01753 ...
  ..- attr(*, "dimnames")=List of 2
  .. ..$ : chr [1:599] "aim" "allowing" "based" "blood" ...
  .. ..$ : chr [1:10] "PC1" "PC2" "PC3" "PC4" ...
 $ center  : Named num [1:599] 0.2265 0.0541 0.6733 0.0701 0.1543 ...
  ..- attr(*, "names")= chr [1:599] "aim" "allowing" "based" "blood" ...
 $ scale   : Named num [1:599] 0.537 0.235 1.049 0.445 0.856 ...
  ..- attr(*, "names")= chr [1:599] "aim" "allowing" "based" "blood" ...
 $ x       : num [1:499, 1:10] -3.24 -0.94 -1.62 -1.25 -1.4 ...
  ..- attr(*, "dimnames")=List of 2
  .. ..$ : chr [1:499] "115844" "633197" "633249" "633261" ...
  .. ..$ : chr [1:10] "PC1" "PC2" "PC3" "PC4" ...
 - attr(*, "class")= chr "prcomp"

text_pca[['x']] %>%
  head()

             PC1       PC2        PC3        PC4        PC5        PC6        PC7        PC8         PC9       PC10
115844 -3.240084 -1.224083  1.0929219 -0.2966866  0.7300978 -0.1272715  0.1145893 -2.3920794 -0.01109556  0.1170174
633197 -0.939826  4.805016  1.9730777  2.0355810 -0.9698387  0.5595339  1.9143958 -2.0685189 -1.93477227 -1.4869024
633249 -1.620804  4.488997  0.3095817  2.5046982 -1.6895767  1.9497913 -0.5535784 -0.6742253  0.01409018 -1.1999591
633261 -1.249307  4.793337  1.3972610  3.9865897 -0.6584607  0.7980526  1.1628729 -2.7636819 -3.20517847 -0.6888338
633382 -1.399995  5.265843  0.1248408  4.1248219 -0.9627588  1.3542756  1.4368296 -0.6137203 -1.93784759 -0.8948685
633571  1.445999  2.273599 -0.7621658  1.2438623 -2.1143075  0.4455806  0.8693148  0.8680666  2.56183743  2.5522762

text_pca %>% tidy()

• Again, alternatively with a recipe…

recipe_pca <- recipe_base %>% # tokenize
  step_tfidf(text, prefix = '') %>% # TFIDF weighting
  step_pca(all_predictors(), num_comp = 10) %>% # PCA
  prep() 

recipe_pca %>% juice()

• Some plotting

recipe_pca %>% juice() %>%
  ggplot(aes(x = PC01, y = PC02)) +
  geom_point() 

• we can also use the tidy results of the recipe to do some more analytics

recipe_pca %>%
  tidy(7) %>%
  filter(component %in% paste0("PC", 1:4)) %>%
  group_by(component) %>%
    arrange(desc(value)) %>%
    slice(c(1:2, (n()-2):n())) %>%
  ungroup() %>%
  mutate(component = fct_inorder(component)) %>%
  ggplot(aes(value, terms, fill = terms)) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~component, nrow = 1) +
  labs(y = NULL)

• Note: Also check further for further dimensionlity reduction steps:
  • tep_kpca():
  • step_ica()
  • step_isomap()
  • step_nnmf()

Topic Models: Latent-Dirichlet-Allocation (LDA)

• While we already did it somewhat ‘on-the-fly’, here a more formal introduction to LDA
• In contrast to dimnesionality reduction techiques mostly aiming at preprocessing data or easing visualization, LDA more aims at EDA and interpretation
• It is a generative approach to identify topics (clusters) within the word-usage in documents.
  • Topics are represented as a probability distribution over the words in the vocabulary. Hhigh probability words can be used to charactrize the topic.
  • Documents are represented as a mixture of topics.

alt text

library(topicmodels)

text_dtm <- text_tidy %>%
  cast_dtm(document = id, term = word, value = n)

text_lda <- text_dtm %>% 
  LDA(k = 6, method = "Gibbs",
      control = list(seed = 1337))

• \(\beta\) is an output of the LDA model, indicating the propability that a word occurs in a certain topic.
• Therefore, loking at the top probability words of a topic often gives us a good intuition regarding its properties.

# LDA output is defined for tidy(), so we can easily extract it
lda_beta <- text_lda %>% 
  tidy(matrix = "beta") 

lda_beta %>%
  # slice
  group_by(topic) %>%
  arrange(topic, desc(beta)) %>%
  slice(1:10) %>%
  ungroup() %>%
  # visualize
  mutate(term = reorder_within(term, beta, topic)) %>%
  group_by(topic, term) %>%    
  arrange(desc(beta)) %>%  
  ungroup() %>%
  ggplot(aes(term, beta, fill = as.factor(topic))) +
  geom_col(show.legend = FALSE) +
  coord_flip() +
  scale_x_reordered() +
  labs(title = "Top 10 terms in each LDA topic",
       x = NULL, y = expression(beta)) +
  facet_wrap(~ topic, ncol = 3, scales = "free")

• Documents are represented as a mix of topics. This association of a document to a topic is captured by \(\gamma\)

lda_gamma <- text_lda %>% 
  tidy(matrix = "gamma")

lda_gamma %>%
  group_by(topic) %>%
    arrange(desc(gamma)) %>% 
    slice(1:10) %>%
  ungroup() %>%
  left_join(text %>% select(id, projectacronym) %>% mutate(id = id %>% as.character()), by = c('document' = 'id'))

• Note that an LDA can also be performed via a recipe:

recipe_lda <- recipe_base %>% # tokenize
  step_lda(text, num_topics = 6) %>% # LDA
  prep() 

recipe_lda %>% juice() %>% 
  head(100)

• As a bonus, a great way to interactively visualize LDA’s.
• It’s a bit cumbersome in R, though…

library(LDAvis)

# A bit of a lenghty function....
topicmodels_json_ldavis <- function(fitted, doc_dtm, method = "PCA"){
  require(topicmodels); require(dplyr); require(LDAvis)
  
  # Find required quantities
  phi <- posterior(text_lda)$terms %>% as.matrix() # Topic-term distribution
  theta <- posterior(fitted)$topics %>% as.matrix() # Document-topic matrix
  
  text_tidy <- doc_dtm %>% tidy()
  vocab <- colnames(phi)
  doc_length <- tibble(document = rownames(theta)) %>% left_join(text_tidy %>% count(document, wt = count), by = 'document')
  tf <- tibble(term = vocab) %>% left_join(text_tidy %>% count(term, wt = count), by = "term") 
  
  if(method == "PCA"){mds <- jsPCA}
  if(method == "TSNE"){library(tsne); mds <- function(x){tsne(svd(x)$u)} }
  
  # Convert to json
  json_lda <- LDAvis::createJSON(phi = phi, theta = theta, vocab = vocab, doc.length = doc_length %>% pull(n), term.frequency = tf %>% pull(n),
                                 reorder.topics = FALSE, mds.method = mds,plot.opts = list(xlab = "Dim.1", ylab = "Dim.2")) 
  return(json_lda)
}

library(LDAvis)
json_lda <- topicmodels_json_ldavis(fitted = text_lda, 
                                    doc_dtm = text_dtm, 
                                    method = "TSNE")

json_lda %>% serVis() # For direct output
# json_lda %>% serVis(out.dir = 'LDAviz') # For saving the html

Didnt really figure out how to embedd the resulting plot, but the outcome can be seen here

Embeddings (Bonus)

• One last thing we did not venture in yet, are embeddings

• I will not go into details here, just see it as a peak of what’s to come in further sessions.

• The idee of word embedding is (in a nutshell) that

• There are packages on how to train own embeddings such as text2vec, but we will for now not bother with that.

• The only thing we will do for now is to load pretrained embeddings (GloVe, cf. Pennington et al, 2014)

library(textdata)

glove6b <- embedding_glove6b(dimensions = 100)
glove6b %>% head(1000)

• La voila, a large pretrained embedding model for around 400k of the most common words.
• We for now loaded the smallest of these embedding models, there exist way bigger ones.
• Lets join it with our tidy tokenlist

word_embeddings <- text_tidy %>%
  inner_join(glove6b, by = c('word' = 'token'))

word_embeddings %>% head()

• We could now create average document embeddings by taking the mean over all dimensions
• We could also (even better) weight that by then word’s tfidf score.

doc_embeddings <- word_embeddings %>%
  group_by(id) %>%
  summarise(across(starts_with("d"), ~mean(.x / tf_idf, na.rm = TRUE)))

• These embddings could now be used for instance for some clustering or SML exercise
• I guess you can already see how to use these embeddings in an SML model.

library(uwot) # for UMAP

embeddings_umap <- doc_embeddings  %>% 
  column_to_rownames("id") %>%
  umap(n_neighbors = 15, 
       metric = "cosine", 
       min_dist = 0.01, 
       scale = TRUE,
       verbose = TRUE, 
       n_threads = 8) 

0%   10   20   30   40   50   60   70   80   90   100%
[----|----|----|----|----|----|----|----|----|----|
**************************************************|

embeddings_umap %<>% as.data.frame()

embeddings_umap  %>% 
  ggplot(aes(x = V1, y = V2)) + 
  geom_point(shape = 21, alpha = 0.5) 

• Ok, we see a rather clear seperation of documents.
• Just for fun, lets add a density based clustering (very good for spatial clustering) on top (even though we already see the results)

library(dbscan)

• Do the hirarchical density based clustering

embeddings_hdbscan <- embeddings_umap %>% as.matrix() %>% hdbscan(minPts = 15)

• Plot it

embeddings_umap %>% 
  bind_cols(cluster = embeddings_hdbscan$cluster %>% as.factor(), 
            prob = embeddings_hdbscan$membership_prob) %>%
  ggplot(aes(x = V1, y = V2, col = cluster)) + 
  geom_point(aes(alpha = prob), shape = 21) 

• Note: We can also assigne the embeddings via a recipe
• Unfortunately, we can not do a TFIDF weighting here ‘out-of-the-box’, but have to work with average embeddings instead.

recipe_embedding <- recipe_base %>% # tokenize
  step_word_embeddings(text, embeddings = glove6b, aggregation = 'mean')

recipe_embedding %>% prep() %>% juice() %>% 
  head(100)

library(embed)

recipe_umap <- recipe_embedding %>%
  step_umap(starts_with('w_embed'), n_neighbors = 15) 

recipe_umap %>% prep() %>% juice() %>% 
  head(100)

—>

• So, that’s all I have for now

Summary

• There are many ways to convert text data into a vector representation.
• These range from simple and weighted bags-of-words, to topic models, over different types of dimensionality reduction to finally word and document embeddings.
• All of them are useful, depending on the purpose.

Endnotes

Packages & Ecosystem

• textrecipes: Text preprocessing recipes
• embed: Extra embedding recipes
• topicmodels: LDA topicmodelling in R
• LDAvis: A bit clunky but awesome interactive LDA visualizations
• text2vec: Package vor vector space modelling (aka embeddings & other vectorizations) of textdata
• textdata: Useful datasets for text, such as GloVe embeddings, sentiment lexica etc.
• uwot: UMAP for R

References

CHapters:

• Julia Silge and David Robinson (2020). Text Mining with R: A Tidy Approach, O’Reilly. Online available here
  • Chapter 6: xxx
• Emil Hvidfeldt and Julia Silge (2020). Supervised Machine Learning for Text Analysis in R, online available here
  • Chapter 5: Word Embeddings

Articles: * Blei, David M., Andrew Y. Ng, and Michael I. Jordan. “Latent dirichlet allocation.” Journal of machine Learning research 3, no. Jan (2003): 993-1022. * Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Conference on Empirical Methods on Natural Language Processing (EMNLP), pages 1532–1543, 2014

Further sources

• Julia Silge’s Blog: Full of great examples of predictive modeling, NLP, and the combination fo both, using tidy ecosystems
• Emil Hvitfeldt’s Blog: Likewise, full of great examples of applied tidy ML & NLP in

Session Info

sessionInfo()