moving_embedings.R

library(tidyverse)

read_lines("revizor.txt") |> 
  str_squish() |> 
  tibble(text = _) |> 
  filter(!str_detect(text, "^ДЕЙСТВИЕ"),
         !str_detect(text, "^Явление"),
         text != "") |> 
  mutate(id = 1:n()) ->
  revizor

window <- 20
indecies <- 1:(nrow(revizor)-window)

map_dfr(indecies, function(i){
  revizor |> 
    slice(i:(i+window-1)) |> 
    reframe(text = str_c(text, collapse = " ")) |> 
    mutate(doc_id = i)
}) ->
  result

library(word2vec)

result |> 
  mutate(text = tolower(text)) |> 
  pull(text) |> 
  word2vec(type = "cbow") ->
  #word2vec(type = "skip-gram") ->
  model

result |> 
  mutate(text = tolower(text)) |> 
  doc2vec(object = model, type = "embedding") ->
  embeddings

library(uwot)

embeddings |>
  umap(X = _, metric = "cosine", init = "pca") ->
  umap

as_tibble(umap) |> 
  rename(umap_v1 = V1, umap_v2 = V2) |> 
  mutate(id = 1:n()) |> 
  ggplot(aes(umap_v1, umap_v2, color = id))+
  geom_path()+
  geom_point()+
  theme_minimal()+
  labs(title = "moving embedings: Ревизор",
       caption = str_glue("window: {window}"))
ggsave("revizor.png", bg = "white", height = 7, width = 7)