ETM - R package for Topic Modelling in Embedding Spaces

This repository contains an R package called topicmodels.etm which is an implementation of ETM

ETM is a generative topic model combining traditional topic models (LDA) with word embeddings (word2vec)
- It models each word with a categorical distribution whose natural parameter is the inner product between a word embedding and an embedding of its assigned topic
- The model is fitted using an amortized variational inference algorithm on top of libtorch (https://torch.mlverse.org)
The techniques are explained in detail in the paper: “Topic Modelling in Embedding Spaces” by Adji B. Dieng, Francisco J. R. Ruiz and David M. Blei, available at https://arxiv.org/pdf/1907.04907.pdf

Installation

For installing the package from CRAN:

pkgs <- c("torch", "topicmodels.etm", "word2vec", "doc2vec", "udpipe", "uwot")
install.packages(pkgs)
library(torch)
library(topicmodels.etm)

For installing the development version of this package: you can perform the following installations in R:

install.packages("torch")
install.packages("word2vec")
install.packages("doc2vec")
install.packages("udpipe")
install.packages("remotes")
library(torch)
remotes::install_github('bnosac/ETM', INSTALL_opts = '--no-multiarch')

For allowing to plot the models:

install.packages("textplot")
install.packages("ggrepel")
install.packages("ggalt")

Example

Build a topic model on questions answered in Belgian parliament in 2020 in Dutch.

a. Get data

Example text of +/- 6000 questions asked in the Belgian parliament (available in R package doc2vec).
Standardise the text a bit

library(torch)
library(topicmodels.etm)
library(doc2vec)
library(word2vec)
data(be_parliament_2020, package = "doc2vec")
x      <- data.frame(doc_id           = be_parliament_2020$doc_id, 
                     text             = be_parliament_2020$text_nl, 
                     stringsAsFactors = FALSE)
x$text <- txt_clean_word2vec(x$text)

b. Build a word2vec model to get word embeddings and inspect it a bit

w2v        <- word2vec(x = x$text, dim = 25, type = "skip-gram", iter = 10, min_count = 5, threads = 2)
embeddings <- as.matrix(w2v)
predict(w2v, newdata = c("migranten", "belastingen"), type = "nearest", top_n = 4)
$migranten
      term1               term2 similarity rank
1 migranten              lesbos  0.9434163    1
2 migranten               chios  0.9334459    2
3 migranten vluchtelingenkampen  0.9269973    3
4 migranten                kamp  0.9175452    4

$belastingen
        term1                term2 similarity rank
1 belastingen            belasting  0.9458982    1
2 belastingen          ontvangsten  0.9091899    2
3 belastingen              geheven  0.9071115    3
4 belastingen            ontduiken  0.9029559    4

c. Build the embedding topic model

Create a bag of words document term matrix (using the udpipe package but other R packages provide similar functionalities)
Keep only the top 50% terms with the highest TFIDF
Make sure document/term/matrix and the embedding matrix have the same vocabulary

library(udpipe)
dtm   <- strsplit.data.frame(x, group = "doc_id", term = "text", split = " ")
dtm   <- document_term_frequencies(dtm)
dtm   <- document_term_matrix(dtm)
dtm   <- dtm_remove_tfidf(dtm, prob = 0.50)

vocab        <- intersect(rownames(embeddings), colnames(dtm))
embeddings   <- dtm_conform(embeddings, rows = vocab)
dtm          <- dtm_conform(dtm,     columns = vocab)
dim(dtm)
dim(embeddings)

Learn 20 topics with a 100-dimensional hyperparameter for the variational inference

set.seed(1234)
torch_manual_seed(4321)
model     <- ETM(k = 20, dim = 100, embeddings = embeddings)
optimizer <- optim_adam(params = model$parameters, lr = 0.005, weight_decay = 0.0000012)
loss      <- model$fit(data = dtm, optimizer = optimizer, epoch = 20, batch_size = 1000)
plot(model, type = "loss")

d. Inspect the model

terminology  <- predict(model, type = "terms", top_n = 5)
terminology
[[1]]
              term       beta
3891 zelfstandigen 0.05245856
2543      opdeling 0.02827548
5469  werkloosheid 0.02366866
3611          ocmw 0.01772762
4957  zelfstandige 0.01139760

[[2]]
              term        beta
3891 zelfstandigen 0.032309771
5469  werkloosheid 0.021119611
4957  zelfstandige 0.010217560
3611          ocmw 0.009712025
2543      opdeling 0.008961252

[[3]]
              term       beta
2537 gedetineerden 0.02914266
3827 nationaliteit 0.02540042
3079    gevangenis 0.02136421
5311 gevangenissen 0.01215335
3515  asielzoekers 0.01204639

[[4]]
             term       beta
3435          btw 0.02814350
5536    kostprijs 0.02012880
3508          pod 0.01218093
2762          vzw 0.01088356
2996 vennootschap 0.01015108

[[5]]
               term        beta
3372        verbaal 0.011172118
3264    politiezone 0.008422602
3546 arrondissement 0.007855867
3052      inbreuken 0.007204257
2543       opdeling 0.007149355

[[6]]
                  term       beta
3296        instelling 0.04442037
3540 wetenschappelijke 0.03434755
2652             china 0.02702594
3043    volksrepubliek 0.01844959
3893          hongkong 0.01792639

[[7]]
               term        beta
2133       databank 0.003111386
3079     gevangenis 0.002650804
3255            dvz 0.002098217
3614         centra 0.001884672
2142 geneesmiddelen 0.001791468

[[8]]
         term       beta
2547 defensie 0.03706463
3785  kabinet 0.01323747
4054  griekse 0.01317877
3750   turkse 0.01238277
3076    leger 0.00964661

[[9]]
            term        beta
3649        nmbs 0.005472604
3704      beslag 0.004442090
2457   nucleaire 0.003911803
2461 mondmaskers 0.003712016
3533   materiaal 0.003513884

[[10]]
               term        beta
4586   politiezones 0.017413139
2248     voertuigen 0.012508971
3649           nmbs 0.008157282
2769 politieagenten 0.007591151
3863        beelden 0.006747020

[[11]]
              term        beta
3827 nationaliteit 0.009992087
4912        duitse 0.008966853
3484       turkije 0.008940011
2652         china 0.008723009
4008  overeenkomst 0.007879931

[[12]]
           term        beta
3651 opsplitsen 0.008752496
4247   kinderen 0.006497230
2606  sciensano 0.006430181
3170      tests 0.006420473
3587  studenten 0.006165542

[[13]]
               term        beta
3052      inbreuken 0.007657704
2447          drugs 0.006734609
2195      meldingen 0.005259825
3372        verbaal 0.005117311
3625 cyberaanvallen 0.004269334

[[14]]
         term       beta
2234 gebouwen 0.06128503
3531 digitale 0.03030998
3895    bpost 0.02974019
4105    regie 0.02608073
3224 infrabel 0.01758554

[[15]]
         term       beta
3649     nmbs 0.08117295
3826  station 0.03944306
3911    trein 0.03548101
4965  treinen 0.02843846
3117 stations 0.02732874

[[16]]
                term       beta
3649            nmbs 0.06778506
3240 personeelsleden 0.03363639
2972        telewerk 0.01857295
4965         treinen 0.01807373
3785         kabinet 0.01702784

[[17]]
                 term        beta
2371              app 0.009092372
3265          stoffen 0.006641808
2461      mondmaskers 0.006462210
3025 persoonsgegevens 0.005374488
2319         websites 0.005372964

[[18]]
         term       beta
5296 aangifte 0.01940070
3435      btw 0.01360575
2762      vzw 0.01307520
2756 facturen 0.01233578
2658 rekenhof 0.01196285

[[19]]
               term        beta
3631      beperking 0.017481016
3069       handicap 0.010403863
3905 tewerkstelling 0.009714387
3785        kabinet 0.006984415
2600      ombudsman 0.006074827

[[20]]
          term       beta
3228    geweld 0.05881281
4178   vrouwen 0.05113553
4247  kinderen 0.04818219
2814  jongeren 0.01803746
2195 meldingen 0.01548613

e. Predict alongside the model

newdata <- head(dtm, n = 5)
scores  <- predict(model, newdata, type = "topics")
scores

f. Save / Load model

torch_save(model, "example_etm.ckpt")
model <- torch_load("example_etm.ckpt")

g. Optionally - visualise the model in 2D

Example plot shown above was created using the following code

This uses R package textplot >= 0.2.0 which was updated on CRAN on 2021-08-18
The summary function maps the learned embeddings of the words and topic centers in 2D using UMAP and textplot_embedding_2d plots the selected topics of interest in 2D

library(textplot)
library(uwot)
library(ggrepel)
library(ggalt)
manifolded <- summary(model, type = "umap", n_components = 2, metric = "cosine", n_neighbors = 15, 
                      fast_sgd = FALSE, n_threads = 2, verbose = TRUE)
space      <- subset(manifolded$embed_2d, type %in% "centers")
textplot_embedding_2d(space)
space      <- subset(manifolded$embed_2d, cluster %in% c(12, 14, 9, 7) & rank <= 7)
textplot_embedding_2d(space, title = "ETM topics", subtitle = "embedded in 2D using UMAP", 
                      encircle = FALSE, points = TRUE)

z. Or you can brew up your own code to plot things

Put embeddings of words and topic centers in 2D using UMAP

library(uwot)
centers    <- as.matrix(model, type = "embedding", which = "topics")
embeddings <- as.matrix(model, type = "embedding", which = "words")
manifold   <- umap(embeddings, 
                   n_components = 2, metric = "cosine", n_neighbors = 15, fast_sgd = TRUE, 
                   n_threads = 2, ret_model = TRUE, verbose = TRUE)
centers    <- umap_transform(X = centers, model = manifold)
words      <- manifold$embedding

Plot words in 2D, color by topic and add topic centers in 2D
- This uses R package textplot >= 0.2.0 (https://github.com/bnosac/textplot) which was put on CRAN on 2021-08-18

library(data.table)
terminology  <- predict(model, type = "terms", top_n = 7)
terminology  <- rbindlist(terminology, idcol = "cluster")
df           <- list(words   = merge(x = terminology, 
                                     y = data.frame(x = words[, 1], y = words[, 2], term = rownames(embeddings)), 
                                     by = "term"), 
                     centers = data.frame(x = centers[, 1], y = centers[, 2], 
                                          term = paste("Topic-", seq_len(nrow(centers)), sep = ""), 
                                          cluster = seq_len(nrow(centers))))
df           <- rbindlist(df, use.names = TRUE, fill = TRUE, idcol = "type")
df           <- df[, weight := ifelse(is.na(beta), 0.8, beta / max(beta, na.rm = TRUE)), by = list(cluster)]

library(textplot)
library(ggrepel)
library(ggalt)
x <- subset(df, type %in% c("words", "centers") & cluster %in% c(1, 3, 4, 8))
textplot_embedding_2d(x, title = "ETM topics", subtitle = "embedded in 2D using UMAP", encircle = FALSE, points = FALSE)
textplot_embedding_2d(x, title = "ETM topics", subtitle = "embedded in 2D using UMAP", encircle = TRUE, points = TRUE)

Or if you like writing down the full ggplot2 code

library(ggplot2)
library(ggrepel)
x$topic <- factor(x$cluster)
plt     <- ggplot(x, 
    aes(x = x, y = y, label = term, color = topic, cex = weight, pch = factor(type, levels = c("centers", "words")))) + 
    geom_text_repel(show.legend = FALSE) + 
    theme_void() + 
    labs(title = "ETM topics", subtitle = "embedded in 2D using UMAP")
plt + geom_point(show.legend = FALSE)

## encircle if topics are non-overlapping can provide nice visualisations
library(ggalt)
plt + geom_encircle(aes(group = topic, fill = topic), alpha = 0.4, show.legend = FALSE) + geom_point(show.legend = FALSE)

More examples are provided in the help of the ETM function see ?ETM Don’t forget to set seeds to have reproducible behaviour

Support in text mining

Need support in text mining? Contact BNOSAC: http://www.bnosac.be