Skip to contents

Sequential LDA

Sentence-level topic modeling

Source: vignettes/pkgdown/sequential.Rmd
sequential.Rmd

Sequential LDA (Latent Dirichlet Allocation) can identify topics of sentences more accurately than the standard LDA, because it considers the topics of the preceding sentence to classify the current sentence. The topics of the preceding sentences not only make the transition of topics smoother but help to mitigate the data sparsity.

Preperation

We use the Sputnik corpus on Ukraine, but preparation in this example is slightly different from the introduction. Since Sequential LDA is created for classification of sentences, we apply corpus_reshape() to the corpus to segment the texts

library(seededlda)
library(quanteda)
library(ggmatplot)

corp <- readRDS("data_corpus_sputnik2022.rds") |>
    corpus_reshape(to = "sentences")
toks <- tokens(corp, remove_punct = TRUE, remove_symbols = TRUE, 
               remove_numbers = TRUE, remove_url = TRUE)
dfmt <- dfm(toks) |> 
    dfm_remove(stopwords("en")) |>
    dfm_remove("*@*") |>
    dfm_trim(max_docfreq = 0.1, docfreq_type = "prop")

As a result, the DFM has more documents (150925) with sentence number appended to the document names. The sentences make the sparsity of the DFM very high (99.98%).

print(dfmt)
#> Document-feature matrix of: 150,925 documents, 58,838 features (99.98% sparse) and 4 docvars.
#>                features
#> docs            monday president joe biden reiterated united states commitment
#>   s1092644731.1      1         1   1     1          1      1      1          1
#>   s1092644731.2      0         0   0     1          0      0      0          0
#>   s1092644731.3      0         2   0     1          0      0      0          0
#>   s1092644731.4      0         1   0     0          0      0      0          0
#>   s1092644731.5      0         0   0     0          0      0      0          0
#>   s1092644731.6      0         0   0     0          0      0      0          0
#>                features
#> docs            diplomacy tensions
#>   s1092644731.1         1        1
#>   s1092644731.2         1        0
#>   s1092644731.3         1        0
#>   s1092644731.4         1        0
#>   s1092644731.5         0        0
#>   s1092644731.6         0        0
#> [ reached max_ndoc ... 150,919 more documents, reached max_nfeat ... 58,828 more features ]

Sequential LDA

You can enable the sequential algorithm only by setting gamma = 0.5. If the value is smaller than this, the topics of the previous sentence affect less in classifying the current sentence.

lda_seq <- textmodel_lda(dfmt, k = 10, gamma = 0.5, 
                         batch_size = 0.01, auto_iter = TRUE,
                         verbose = TRUE)
#> Fitting LDA with 10 topics
#>  ...initializing
#>  ...using up to 16 threads for distributed computing
#>  ......allocating 1510 documents to each thread
#>  ...Gibbs sampling in up to 2000 iterations
#>  ......iteration 100 elapsed time: 14.36 seconds (delta: 0.27%)
#>  ......iteration 200 elapsed time: 27.44 seconds (delta: 0.03%)
#>  ......iteration 300 elapsed time: 40.55 seconds (delta: -0.04%)
#>  ...computing theta and phi
#>  ...complete
knitr::kable(terms(lda_seq))
topic1 topic2 topic3 topic4 topic5 topic6 topic7 topic8 topic9 topic10
gas world moscow biden president nato operation energy military ukrainian
energy people security discuss china finland military prices weapons sputnik
oil war foreign joins biden sweden ukrainian inflation ukrainian media
sanctions discuss nato president taiwan military donetsk uk nuclear people
european countries minister also military security special percent defense war
countries political also former washington countries forces oil forces one
eu joined countries sputnik war joined republics government kiev also
prices international military get also alliance donbass year also photo
europe sean united house nato hour lugansk crisis ministry military
also jacquie president new security turkey february price according kiev

Seeded Sequential LDA

You can also enable the sequential algorithm in Seeded LDA only by setting gamma > 0.

dict <- dictionary(file = "dictionary.yml")
print(dict)
#> Dictionary object with 5 key entries.
#> - [economy]:
#>   - market*, money, bank*, stock*, bond*, industry, company, shop*
#> - [politics]:
#>   - parliament*, congress*, white house, party leader*, party member*, voter*, lawmaker*, politician*
#> - [society]:
#>   - police, prison*, school*, hospital*
#> - [diplomacy]:
#>   - ambassador*, diplomat*, embassy, treaty
#> - [military]:
#>   - military, soldier*, terrorist*, air force, marine, navy, army
lda_seed <- textmodel_seededlda(dfmt, dict, residual = 2, gamma = 0.5, 
                                batch_size = 0.01, auto_iter = TRUE,
                                verbose = TRUE)
#> Fitting LDA with 7 topics
#>  ...initializing
#>  ...using up to 16 threads for distributed computing
#>  ......allocating 1510 documents to each thread
#>  ...Gibbs sampling in up to 2000 iterations
#>  ......iteration 100 elapsed time: 12.70 seconds (delta: 0.05%)
#>  ......iteration 200 elapsed time: 24.17 seconds (delta: -0.01%)
#>  ...computing theta and phi
#>  ...complete
knitr::kable(terms(lda_seed))
economy politics society diplomacy military other1 other2
energy biden ukrainian nato military china nato
gas discuss people security ukrainian countries sputnik
oil president police moscow operation also eu
prices joins one military forces world european
sanctions also war president special india finland
market congress also ukrainian donetsk president countries
countries sputnik minister foreign kiev new sweden
price political sputnik weapons republics economic also
inflation get photo states donbass international media
european joined johnson united lugansk foreign military

The plot show the topic probability of the first 100 sentences from an article in the corpus. Thanks to the sequential algorithm, adjacent sentences are classified into the same topics.

ggmatplot(lda_seed$theta[paste0("s1104078647.", 1:100),], 
          plot_type = "line", linetype = 1, 
          xlab = "Sentence", ylab = "Probability", legend_title = "Topic")

References