Analyzing Film Censorship in India: CBFC Watch

---

Note: This notebook was updated on Tuesday 16th September, 3:00 PM to correct a data preprocessing error, which erroneously filtered out some data from being including in the summaries. The correction is viewable here. This has been corrected.

---

1 Introduction

This notebook contains the data analysis for the charts and numbers used on our site, CBFC.WATCH. To ensure full transparency with our readers, we are publishing our methodology and the R code used to generate the statistics for various parts of the site, as well as to provide a starting point for others who may be curious about how to use the data.

We explore these questions:

1. Do some regional CBFC offices require more modifications than others?
2. How have the primary reasons for censorship (e.g., violence, profanity, sexual content) trended over time?
3. How much footage is typically removed from a film?
4. Which specific censorship categories are becoming more frequent, and which are declining?
5. What specific words and themes are most associated with each major censorship category?
6. Which specific films have had the most footage removed by each regional office?

1.1 Data Loading and Preparation

Importing the dataset, correcting data types, and standardizing categorical information like language and regional office names.

data <- read_csv("https://github.com/diagram-chasing/censor-board-cuts/raw/refs/heads/master/data/data.csv", 
                 col_types = cols(.default = "c")) %>%
  mutate(
    cert_date = as.Date(cert_date),
    total_modified_time_secs = as.numeric(total_modified_time_secs),
    deleted_secs = as.numeric(deleted_secs),
    replaced_secs = as.numeric(replaced_secs),
    inserted_secs = as.numeric(inserted_secs)
  ) %>%
  # The 'certifier' column contains a string like "Examining Committee, Mumbai".
  # We extract the regional office name by splitting the string by the comma
  # and taking the last element.
  mutate(
    office = str_split(certifier, ",") %>%
      map_chr(last) %>%
      str_trim()
  ) %>%
  separate_rows(ai_content_types, sep = "\\|") %>%
  mutate(ai_content_types = str_trim(ai_content_types))%>%
  # Filter out any rows where the content type is empty after separation.
 # filter(ai_content_types != "") %>%
  # Standardize language names to correct for typos and variations in the raw data
  mutate(
    language = case_when(
      language == "Oriya" ~ "Odia",
      language == "Gujrati" ~ "Gujarati",
      language == "Chhatisgarhi" ~ "Chhattisgarhi",
      language == "Hariyanvi" ~ "Haryanvi",
      language == "Hindi Dub" ~ "Hindi Dubbed",
      TRUE ~ language
    )
  ) %>%
  filter(!is.na(language))

1.2 Exploratory Data Analysis

1.2.1 Languages in the Dataset

Counting how many films for each language in the dataset. We’ll show the top ones.

films_by_language_data <- data %>%
  distinct(id, language) %>%
  count(language, sort = TRUE) %>%
  top_n(10, n) %>%
  mutate(language = fct_reorder(language, n))

write_json(films_by_language_data, "films_by_language.json", pretty = TRUE, auto_unbox = TRUE)

films_by_language_data %>%
  ggplot(aes(x = n, y = language)) +
  geom_col(fill = tertiary_color, alpha = 0.9) +
  geom_text(aes(label = comma(n)), hjust = -0.15, size = 3.5, color = "gray20") +
  scale_x_continuous(
    labels = comma,
    expand = expansion(mult = c(0, 0.12))
  ) +
  labs(
    title = "Top 15 Languages by Number of Films Censored",
    subtitle = "Hindi, Telugu, and Tamil are the languages with the most films",
    x = "Number of Unique Films",
    y = NULL
  ) +
  theme_cbfc()

1.2.2 Reasons for Modification

The most common reasons for modifications, based on the AI-classified content types.

mods_by_content_data <- data %>%
  filter(!is.na(ai_content_types)) %>%
  count(ai_content_types, sort = TRUE) %>%
  top_n(15, n) %>%
  mutate(
    pretty_name = str_replace_all(ai_content_types, "_", " ") %>% str_to_title(),
    pretty_name = fct_reorder(pretty_name, n)
  )

write_json(mods_by_content_data, "modifications_by_content.json", pretty = TRUE, auto_unbox = TRUE)

mods_by_content_data %>%
  ggplot(aes(x = n, y = pretty_name)) +
  geom_col(fill = secondary_color, alpha = 0.9) +
  geom_text(aes(label = comma(n)), hjust = -0.15, size = 3, color = "gray20") +
  scale_x_continuous(
    labels = comma,
    expand = expansion(mult = c(0, 0.1))
  ) +
  labs(
    title = "Most Common Reasons for Film Modifications",
    x = "Number of Modifications",
    y = "Content Category"
  ) +
  theme_cbfc()

1.2.3 Duration of Edits by Action

We’ve classified each modification log into verbs depending on what action was taken. For example, adding a smoking disclaimer might be an ‘Insertion’ but removing an entire scene is ‘Deletion’. There are also replacements, audio modifications, visual modifications (such as blurs), and so on. We can look at the general distribution for each of these edits.

duration_summary <- data %>%
  filter(total_modified_time_secs > 0, !is.na(ai_action)) %>%
  group_by(ai_action) %>%
  summarise(
    min = min(total_modified_time_secs, na.rm = TRUE),
    q1 = quantile(total_modified_time_secs, 0.25, na.rm = TRUE),
    median = median(total_modified_time_secs, na.rm = TRUE),
    q3 = quantile(total_modified_time_secs, 0.75, na.rm = TRUE),
    max = max(total_modified_time_secs, na.rm = TRUE),
    count = n(),
    .groups = 'drop'
  ) %>%
  mutate(pretty_name = str_replace_all(ai_action, "_", " ") %>% str_to_title()) %>%
  arrange(median)


# Export sample of raw data for BoxX component
duration_boxplot_data <- data %>%
   filter(
    total_modified_time_secs > 0, 
    total_modified_time_secs < 1400, # Found that anything above this is mostly incorrectly entered data
    !is.na(ai_action),
    ai_action != 'content_overlay',
    movie_name != 'ARISHADVARGA' # this movie clearly has a data processing error where the times are added cumulatively https://archive.org/details/cbfc-ecinepramaan-100020292100000261
  ) %>%
  mutate(pretty_name = str_replace_all(ai_action, "_", " ") %>% str_to_title()) %>%
  group_by(ai_action, pretty_name) %>%
  # Sample up to 500 points. If a group has < 500 points, it will take all of them.
  slice_sample(n = 500) %>%
  ungroup() %>%
  select(
    action = ai_action,
    category = pretty_name, 
    duration = total_modified_time_secs
  )


write_json(duration_summary, "duration_by_action_summary.json", pretty = TRUE, auto_unbox = TRUE)
write_json(duration_boxplot_data, "duration_by_action_boxplot.json", pretty = TRUE, auto_unbox = TRUE)

  duration_boxplot_data %>% 
  ggplot(aes(
    x = fct_reorder(category, duration, .fun = median),
    y = duration,
    fill = category
  )) +
  geom_boxplot(show.legend = FALSE, alpha = 0.8, fill = wes_colors[3]) +
  scale_y_log10(
    breaks = c(1, 10, 60, 300, 1800),
    labels = c("1 sec", "10 secs", "1 min", "5 mins", "30 mins")
  ) +
  coord_flip() +
  labs(
    title = "How Long are Different Types of Edits?",
    subtitle = "Distribution of modification durations for each action type on a logarithmic scale.",
    x = NULL,
    y = "Duration of Modification"
  ) +
  theme_cbfc()

1.3 Modification Reasons by Movie Rating

The dataset includes tags for what reason a particular modification was made; violence, profanity and so on. While I have a pretty good idea of what this will give, we can see what is the breakdown of each reason by the rating for that movie (rating is the U, UA, A classification for who can watch the movie).

rating_breakdown <- data %>%
  filter(rating %in% c("U", "UA", "A"), !is.na(ai_content_types)) %>%
  mutate(content_category = case_when(
    ai_content_types %in% c("sexual_explicit", "sexual_suggestive") ~ "Sexual Content",
    ai_content_types == "profanity" ~ "Profanity",
    ai_content_types == "violence" ~ "Violence",
    ai_content_types == "substance" ~ "Substance Use",
    ai_content_types == "religious" ~ "Religious Content",
    ai_content_types == "political" ~ "Political Content",
    TRUE ~ "Other"
  )) %>%
  count(rating, content_category, name = "modification_count", sort = TRUE) %>%
  group_by(rating) %>%
  mutate(percentage = modification_count / sum(modification_count)) %>%
  ungroup()

common_types <- rating_breakdown %>%
  filter(content_category != "Other") %>%
  group_by(content_category) %>%
  summarise(total = sum(modification_count), .groups = 'drop') %>%
  pull(content_category)

rating_breakdown %>%
  filter(content_category %in% common_types) %>%
  ggplot(aes(
    x = percentage,
    y = fct_reorder(content_category, percentage, .desc = FALSE),
    fill = rating
  )) +
  geom_col(color = "white", linewidth = 0.5) +
  facet_wrap(~rating, scales = "free_y", ncol = 3) +
  scale_x_continuous(labels = percent_format(accuracy = 1)) +
  scale_fill_manual(
    values = c("U" = wes_colors[1], "UA" = wes_colors[2], "A" = wes_colors[3])
  ) +
  labs(
    title = "Censorship reasons by film rating",
    subtitle = "Proportional breakdown of modification reasons within each film rating category.",
    x = "% of all modifications for this rating",
    y = NULL
  ) +
  theme_cbfc() +
  theme(
    strip.text = element_text(size = 14, face = "bold", color = "gray20", margin = margin(t = 5, b = 10)),
    strip.background = element_blank(),
    panel.grid.major.x = element_line(color = "gray90", linetype = "dotted"),
    panel.spacing.x = unit(2, "lines"),
    legend.position = "none"
  )

write_json(rating_breakdown, "rating_breakdown.json", pretty = TRUE, auto_unbox = TRUE)

1.4 Average Cut Time by Regional Office

Different regional offices of the CBFC may apply censorship standards differently. Here, we calculate the average total time modified per film for each major office. We filter out extreme outliers (films with more than 30 minutes of cuts) and only include offices that have certified a sufficient number of films (at least 50).

film_summary <- data %>%
  filter(!is.na(office) & !str_detect(office, "\\.mp4$")) %>%
  group_by(id, movie_name, office) %>%
  summarise(
    total_secs = sum(total_modified_time_secs, na.rm = TRUE),
    .groups = 'drop'
  ) %>%
  filter(total_secs > 0, total_secs < 1800)

office_stats <- film_summary %>%
  group_by(office) %>%
  filter(n() >= 50) %>%
  summarise(
    film_count = n(),
    mean_total_secs = mean(total_secs, na.rm = TRUE),
    se = sd(total_secs, na.rm = TRUE) / sqrt(n()),
    ci_lower_secs = pmax(0, mean_total_secs - 1.96 * se),
    ci_upper_secs = mean_total_secs + 1.96 * se,
    .groups = 'drop'
  ) %>%
  arrange(desc(mean_total_secs))

write_json(office_stats, "cuts_by_office.json", pretty = TRUE, auto_unbox = TRUE)

summary_table <- office_stats %>%
  mutate(
    `Average Time` = format_seconds(mean_total_secs),
    `95% CI Lower` = format_seconds(ci_lower_secs),
    `95% CI Upper` = format_seconds(ci_upper_secs)
  ) %>%
  select(
    Office = office,
    `Films Analyzed` = film_count,
    `Average Time`,
    `95% CI Lower`,
    `95% CI Upper`
  )

summary_table %>%
  gt() %>%
  tab_header(
    title = "Average Modification Time per Film by CBFC Regional Office",
    subtitle = "Analysis of offices with at least 50 certified films"
  ) %>%
  tab_style(
    style = list(
      cell_fill(color = wes_colors[1], alpha = 0.3),
      cell_text(weight = "bold")
    ),
    locations = cells_column_labels()
  ) %>%
  tab_style(
    style = cell_text(align = "center"),
    locations = cells_body(columns = c(`Films Analyzed`, `Average Time`, `95% CI Lower`, `95% CI Upper`))
  ) %>%
  cols_align(
    align = "left",
    columns = Office
  ) %>%
  tab_options(
    table.font.names = "Atkinson Hyperlegible",
    heading.title.font.size = 16,
    heading.subtitle.font.size = 12,
    column_labels.font.size = 12,
    table.font.size = 11
  )

Average Modification Time per Film by CBFC Regional Office
Analysis of offices with at least 50 certified films
Office	Films Analyzed	Average Time	95% CI Lower	95% CI Upper
Cuttack	118	9m 10s	7m 38s	10m 43s
Delhi	442	4m 06s	3m 33s	4m 38s
Mumbai	6573	3m 51s	3m 42s	3m 59s
Thiruvananthpuram	688	3m 16s	2m 55s	3m 36s
Chennai	1657	3m 05s	2m 52s	3m 19s
Kolkata	209	2m 32s	1m 59s	3m 04s
Bangalore	982	2m 13s	1m 59s	2m 27s
Hyderabad	1055	2m 01s	1m 49s	2m 13s

1.5 Trends in Censorship Content Types

Have the reasons for film censorship changed over time? This analysis tracks the changes in different modification types (Violence, Profanity, Sexual Content, etc.) on a monthly basis.

For each category, we calculate its deviation from its own historical average (the “baseline”). This method clearly shows periods when a certain type of content was censored more or less frequently than usual. We also apply a 3-month moving average to smooth out short-term noise and reveal the underlying trend.

interpolate_at_zero <- function(df) {
  sign_change <- which(diff(sign(df$deviation_pct_smooth)) != 0)
  if (length(sign_change) == 0) return(df)
  
  interpolated <- map_df(sign_change, ~{
    slice <- df[.x:(.x + 1), ]
    zero_cross_date <- approx(slice$deviation_pct_smooth, as.numeric(slice$year_month), xout = 0)$y
    if (is.na(zero_cross_date)) return(tibble())
    tibble(
      year_month = as.Date(zero_cross_date, origin = "1970-01-01"),
      deviation_pct_smooth = 0,
      above_baseline = slice$above_baseline[2],
      content_type = slice$content_type[1],
      baseline_rate = slice$baseline_rate[1]
    )
  })
  
  bind_rows(df, interpolated) %>% arrange(year_month)
}

chart_data <- data %>%
  mutate(
    year_month = floor_date(cert_date, "month"),
    content_type = case_when(
      ai_content_types %in% c("sexual_explicit", "sexual_suggestive") ~ "Sexual Content",
      ai_content_types == "profanity" ~ "Profanity",
      ai_content_types == "violence" ~ "Violence",
      ai_content_types == "substance" ~ "Substance Use",
      ai_content_types == "religious" ~ "Religious Content",
      ai_content_types == "political" ~ "Political Content",
      TRUE ~ NA_character_
    )
  ) %>%
  filter(!is.na(content_type)) %>%
  group_by(year_month, content_type) %>%
  summarise(films_with_mods = n_distinct(certificate_id), .groups = "drop") %>%
  left_join(
    data %>%
      mutate(year_month = floor_date(cert_date, "month")) %>%
      group_by(year_month) %>%
      summarise(total_films = n_distinct(certificate_id)),
    by = "year_month"
  ) %>%
  mutate(rate = films_with_mods / total_films) %>%
  filter(total_films >= 10) %>%
  group_by(content_type) %>%
  arrange(year_month) %>%
  mutate(
    baseline_rate = mean(rate),
    deviation_pct = (rate - baseline_rate) / baseline_rate * 100,
    deviation_pct_smooth = zoo::rollmean(deviation_pct, k = 3, fill = NA, align = "center"),
    above_baseline = ifelse(is.na(deviation_pct_smooth), NA, deviation_pct_smooth > 0)
  ) %>%
  ungroup() %>%
  filter(!is.na(deviation_pct_smooth)) %>%
  mutate(content_type = fct_reorder(content_type, baseline_rate, .desc = TRUE)) %>%
  group_by(content_type) %>%
  group_modify(~ interpolate_at_zero(.x)) %>%
  ungroup()

ggplot(chart_data, aes(x = year_month, y = deviation_pct_smooth, color = above_baseline, group = 1)) +
  geom_hline(yintercept = 0, color = "grey40", linetype = "dashed") +
  geom_line(linewidth = 1.2) +
  facet_wrap(~content_type, ncol = 3) +
  scale_color_manual(
    values = c("TRUE" = wes_colors[5], "FALSE" = wes_colors[1]),
    labels = c("Below Baseline", "Above Baseline"),
    name = NULL
  ) +
  scale_x_date(
    date_labels = "'%y",
    date_breaks = "2 years",
    expand = expansion(mult = c(0.02, 0.02))
  ) +
  scale_y_continuous(
    labels = function(x) paste0(ifelse(x >= 0, "+", ""), round(x, 0), "%")
  ) +
  labs(
    title = "Censorship Patterns vs. Historical Average",
    caption = "Note: Dashed line is the baseline average for that category. Red indicates months with above-average censorship rates; blue is below-average.\nA 3-month moving average has been applied to smooth trends. Source: Central Board of Film Certification.",
    x = NULL,
    y = "Deviation from Baseline"
  ) +
  theme_cbfc() +
  theme(
    plot.caption = element_text(size = 9, color = "grey50", hjust = 0, margin = margin(t = 16)),
    panel.grid.major = element_line(color = "grey90", linewidth = 0.4),
    legend.position = "bottom"
  )

write_json(chart_data, "content_trends.json", pretty = TRUE, auto_unbox = TRUE)

1.6 Distribution of Modification Durations

This histogram shows the distribution of total modification times for films that had at least one cut. Approximately half of the films in the dataset have ‘zero seconds’ of edits (meaning modifications might have been made but they were not logged with a duration of edit); this chart focuses only on those that have duration values.

film_data <- data %>%
  group_by(id) %>%
  summarise(
    total_modified_time_secs = sum(total_modified_time_secs, na.rm = TRUE),
    .groups = 'drop'
  )

total_films <- nrow(film_data)
percent_zero <- mean(film_data$total_modified_time_secs == 0, na.rm = TRUE)
film_data_positive <- film_data %>% filter(total_modified_time_secs > 0)
median_val_positive <- median(film_data_positive$total_modified_time_secs)

p_dist <- ggplot(film_data_positive, aes(x = total_modified_time_secs)) +
  geom_histogram(
    aes(y = after_stat(count) / total_films * 100),
    fill = wes_colors[2],
    bins = 40,
    boundary = 0
  ) +
  geom_vline(
    xintercept = median_val_positive,
    linetype = "dashed",
    color = "gray20",
    linewidth = 0.8
  ) +
  # Use a log10 scale on the x-axis because the data is highly skewed.
  scale_x_log10(
    breaks = c(1, 10, 60, 300, 1800),
    labels = c("1s", "10s", "1m", "5m", "30m")
  ) +
  scale_y_continuous(
    expand = expansion(mult = c(0, 0.05)),
    labels = percent_format(scale = 1)
  ) +
  labs(
    title = "Distribution of Modification Times",
    subtitle = str_glue("Comparing total modification time for films with at least one cut.
                        This chart excludes the {percent(percent_zero, accuracy=1)} of films with zero modifications."),
    x = "Total Modification Time (Logarithmic Scale)",
    y = "Percent of All Films"
  ) +
  theme_cbfc() +
  theme(
    panel.grid.major.x = element_blank()
  )

print(p_dist)

# Export the data used for the histogram to a JSON file.
# We use ggplot_build() to extract the computed data from the plot object.
hist_data_for_export <- ggplot_build(p_dist)$data[[1]] %>%
  select(x_min = xmin, x_max = xmax, count, y_percent_total = y)

export_data_dist <- list(
  histogram_bins = hist_data_for_export,
  statistics = list(
    median_positive_secs = median_val_positive,
    total_films = total_films,
    percent_zero_mods = percent_zero
  ),
  axis_config = list(
    breaks = c(1, 10, 60, 300, 1800),
    labels = c("1s", "10s", "1m", "5m", "30m")
  )
)

write_json(export_data_dist, "histogram_data.json", pretty = TRUE, auto_unbox = TRUE)

1.7 Identifying Trending Censorship Categories

To determine which censorship categories are increasing or decreasing over time, we use a logistic regression model.

A positive slope from the model indicates an increasing trend, while a negative slope indicates a decreasing trend. We then visualize the top three fastest-increasing and fastest-decreasing categories since 2018.

This is based on Julia Silge’s excellent article demonstrating a similar type of analysis.

categories_by_quarter <- data %>%
  filter(!is.na(cert_date), !is.na(ai_content_types), cert_date >= as.Date("2018-01-01")) %>%
  mutate(quarter = floor_date(cert_date, unit = "3 months")) %>%
  count(quarter, ai_content_types, name = "category_count") %>%
  group_by(quarter) %>%
  mutate(quarter_total_cuts = sum(category_count)) %>%
  ungroup() %>%
  group_by(ai_content_types) %>%
  filter(sum(category_count) > 100) %>%
  ungroup()

category_models <- categories_by_quarter %>%
  nest(data = -ai_content_types) %>%
  # (total - category) as a function of time (quarter).
  mutate(
    model = map(data, ~ glm(
      cbind(category_count, quarter_total_cuts - category_count) ~ quarter,
      data = .,
      family = "binomial"
    ))
  )

slopes <- category_models %>%
  mutate(tidied = map(model, tidy)) %>%
  unnest(tidied) %>%
  filter(term == "quarter") %>%
  arrange(desc(estimate))

trends_to_plot <- bind_rows(
  slopes %>% top_n(3, estimate),
  slopes %>% top_n(-3, estimate)
)

ggplot(
  categories_by_quarter %>% inner_join(trends_to_plot, by = "ai_content_types"),
  aes(x = quarter, y = category_count / quarter_total_cuts, 
      color = fct_reorder2(ai_content_types, quarter, category_count))
) +
  geom_line(alpha = 0.8, linewidth = 1.2) +
  geom_smooth(method = "lm", se = FALSE, linetype = "dashed", linewidth = 0.5) +
  facet_wrap(~ ifelse(estimate > 0, "Increasing Trends", "Decreasing Trends"), scales = "free_y") +
  scale_y_continuous(labels = percent_format()) +
  scale_color_manual(values = wes_colors) +
  labs(
    title = "Which Censorship Categories Have Trended Up or Down?",
    subtitle = "Proportion of all modifications per quarter since 2018, with a linear trendline.",
    x = "Date of Certification",
    y = "Percent of All Cuts in Quarter",
    color = "Content Type"
  ) +
  theme_cbfc() +
  theme(legend.position = "bottom")

# Export trending categories data
# Create clean version of trends_to_plot without model objects
trends_clean <- trends_to_plot %>% 
  select(ai_content_types, estimate, std.error, statistic, p.value)

trending_export <- list(
  slopes = slopes %>% select(ai_content_types, estimate, std.error, statistic, p.value),
  quarterly_data = categories_by_quarter %>%
    inner_join(trends_clean, by = "ai_content_types") %>%
    select(quarter, ai_content_types, category_count, quarter_total_cuts, 
           estimate, std.error, statistic, p.value)
)
write_json(trending_export, "trending_categories.json", pretty = TRUE, auto_unbox = TRUE)

1.8 Keyword Co-occurrence Network

This analysis explores which terms tend to appear together within the same film’s modification records.

We calculate the pairwise correlation between the 50 most common keywords and visualize the results as a network graph.

This contains a lot of NSFW stuff, but hiding it would…probably defeat the point.

reference_words <- data %>%
  filter(!is.na(ai_reference)) %>%
  mutate(word = str_split(tolower(ai_reference), "\\|")) %>%
  unnest(word) %>%
  filter(!is.na(word), !str_detect(word, "violence|scene|visual|dialogue")) %>%
  select(id, word)

word_counts <- reference_words %>%
  count(word, sort = TRUE)

word_pairs <- reference_words %>%
  filter(word %in% (word_counts %>% top_n(50, n) %>% pull(word))) %>%
  pairwise_cor(item = word, feature = id, sort = TRUE)

word_pairs %>%
  filter(correlation > 0.05) %>%
  graph_from_data_frame() %>%
  ggraph(layout = "fr") +
  geom_edge_link(aes(edge_alpha = correlation), show.legend = FALSE) +
  geom_node_point(color = wes_colors[3], size = 5) +
  geom_node_text(aes(label = name), repel = TRUE) +
  theme_void() +
  labs(
    title = "Which Censorship Terms Appear Together?",
    subtitle = "Co-occurrence network of the 50 most common keywords in CBFC modification records."
  )

# Export co-occurrence network data
network_export <- list(
  word_counts = word_counts %>% top_n(50, n),
  word_pairs = word_pairs %>% filter(correlation > 0.05)
)
write_json(network_export, "keyword_network.json", pretty = TRUE, auto_unbox = TRUE)

1.9 Most Important Keywords by Category (TF-IDF)

While the network graph shows which words co-occur, it doesn’t tell us which words are most characteristic of a specific censorship category. For that, we use a metric called Term Frequency-Inverse Document Frequency (TF-IDF).

TF-IDF identifies words that are common within one category (e.g., “blood” in the “violence” category) but are relatively rare in all other categories.

reference_tokens <- data %>%
  filter(!is.na(ai_reference), !is.na(ai_content_types)) %>%
  select(ai_content_types, ai_reference) %>%
  mutate(word = str_split(tolower(ai_reference), "\\|")) %>%
  unnest(word) %>%
  filter(word != "")

reference_tfidf <- reference_tokens %>%
  count(ai_content_types, word, sort = TRUE) %>%
  bind_tf_idf(term = word, document = ai_content_types, n = n) %>%
  arrange(desc(tf_idf))

top_terms_table <- reference_tfidf %>%
  group_by(ai_content_types) %>%
  slice_max(order_by = tf_idf, n = 5) %>%
  ungroup() %>%
  select(Category = ai_content_types, Term = word, `TF-IDF` = tf_idf) %>%
  filter(Category %in% c("violence", "profanity", "sexual_suggestive", "substance", "political", "religious"))

top_terms_table %>%
  gt() %>%
  tab_header(
    title = "Most Distinctive Keywords by Censorship Category",
    subtitle = "Using Term Frequency-Inverse Document Frequency (TF-IDF) analysis"
  ) %>%
  fmt_number(
    columns = `TF-IDF`,
    decimals = 3
  ) %>%
  tab_style(
    style = list(
      cell_fill(color = wes_colors[2], alpha = 0.3),
      cell_text(weight = "bold")
    ),
    locations = cells_column_labels()
  ) %>%
  tab_style(
    style = cell_text(transform = "capitalize"),
    locations = cells_body(columns = Category)
  ) %>%
  tab_style(
    style = cell_text(style = "italic"),
    locations = cells_body(columns = Term)
  ) %>%
  cols_align(
    align = "center",
    columns = `TF-IDF`
  ) %>%
  tab_options(
    table.font.names = "Atkinson Hyperlegible",
    heading.title.font.size = 16,
    heading.subtitle.font.size = 12,
    column_labels.font.size = 12,
    table.font.size = 11,
    row_group.font.weight = "bold"
  )

Most Distinctive Keywords by Censorship Category
Using Term Frequency-Inverse Document Frequency (TF-IDF) analysis
Category	Term	TF-IDF
political	national flag	0.020
political	modi	0.015
political	indian flag	0.009
political	pakistan	0.007
political	political leaders	0.007
profanity	munda	0.005
profanity	bhenchod	0.004
profanity	iththa	0.003
profanity	bhadkov	0.003
profanity	maal	0.003
religious	superstition	0.044
religious	black magic	0.015
religious	religion	0.014
religious	superstitions	0.012
religious	religious sentiments	0.011
sexual_suggestive	kissing scene	0.022
sexual_suggestive	cleavage	0.019
sexual_suggestive	sexual_suggestive	0.018
sexual_suggestive	kissing	0.017
sexual_suggestive	love making scene	0.017
substance	tobacco	0.028
substance	liquor label	0.017
substance	akshay kumar	0.016
substance	liquor labels	0.015
substance	rahul dravid	0.014
violence	blood	0.032
violence	bloodshed	0.009
violence	killing scene	0.007
violence	dead bodies	0.007
violence	dead body	0.006

write_json(reference_tfidf %>% slice_head(n = 300), "tfidf_analysis.json", pretty = TRUE, auto_unbox = TRUE)

1.10 Top Censored Films by Office

Identifies the top films that had the most time removed by each major CBFC regional office.

film_level_summary <- data %>%
  filter(!is.na(office) & !str_detect(office, "\\.mp4$")) %>%
  group_by(id, movie_name, office, language, cert_date, cert_no) %>%
  summarise(
    total_cuts = n(),
    total_time_removed_secs = sum(total_modified_time_secs, na.rm = TRUE),
    .groups = 'drop'
  ) %>%
  filter(total_time_removed_secs > 0, total_time_removed_secs < 1800) %>%
  mutate(
    year = map2_dbl(cert_date, cert_no, ~ extract_year(.x, .y)),
    cleaned_name = map_chr(movie_name, clean_name),
    slug = map2_chr(cleaned_name, year, make_slug)
  )

top_censored_by_office <- film_level_summary %>%
  group_by(office) %>%
  arrange(desc(total_time_removed_secs)) %>%
  slice_head(n = 10) %>%
  ungroup() %>%
  mutate(
    duration_formatted = format_seconds(total_time_removed_secs)
  ) %>%
  arrange(office, desc(total_time_removed_secs)) %>%
  select(Office = office, Language = language, `Film Name` = cleaned_name, 
         `Time Removed` = duration_formatted, `Total Cuts` = total_cuts, Slug = slug)

display_table <- top_censored_by_office %>%
  select(-Slug)

display_table %>%
  gt() %>%
  tab_header(
    title = "Top 10 Most Censored Films by Time Removed",
    subtitle = "Films with the highest modification times for each regional office"
  ) %>%
  tab_style(
    style = list(
      cell_fill(color = wes_colors[3], alpha = 0.3),
      cell_text(weight = "bold")
    ),
    locations = cells_column_labels()
  ) %>%
  tab_style(
    style = cell_text(style = "italic"),
    locations = cells_body(columns = `Film Name`)
  ) %>%
  cols_align(
    align = "center",
    columns = c(Language, `Time Removed`, `Total Cuts`)
  ) %>%
  cols_align(
    align = "left",
    columns = c(Office, `Film Name`)
  ) %>%
  tab_options(
    table.font.names = "Atkinson Hyperlegible",
    heading.title.font.size = 16,
    heading.subtitle.font.size = 12,
    column_labels.font.size = 12,
    table.font.size = 11
  ) %>%
  tab_style(
    style = cell_borders(
      sides = c("top", "bottom"),
      color = "lightgray",
      weight = px(1)
    ),
    locations = cells_body()
  )

Top 10 Most Censored Films by Time Removed
Films with the highest modification times for each regional office
Office	Language	Film Name	Time Removed	Total Cuts
Bangalore	Kannada	NEGILA ODEYA	26m 22s	5
Bangalore	Tamil	ROCKY	24m 16s	6
Bangalore	Kannada	DAMAYANTHI	23m 21s	13
Bangalore	Tamil	GAJAKESSARI	23m 05s	8
Bangalore	Hindi	REHNA NAHI BIN TERE	23m 02s	1
Bangalore	Tamil	GOOGLY	22m 03s	6
Bangalore	Hindi	KISS	21m 36s	8
Bangalore	Kannada	M E S T R I (REVISED)	20m 52s	23
Bangalore	Tamil	RUSTUM	20m 19s	6
Bangalore	Kannada	TIGER GALLI	19m 45s	188
Chennai	Telugu	PULIDEBBA (REVISED)	28m 25s	18
Chennai	Tamil	PATCHIRAJA PARAVAI	26m 42s	4
Chennai	Tamil	'AALAPIRANTHAVAN' (REVISED)	25m 23s	27
Chennai	Tamil	NERAM NALLA NERAM (REVISED)	25m 01s	29
Chennai	Tamil	NAAN UNNA NINACHEN (REVISED)	24m 34s	28
Chennai	Tamil	THEEVIRAM	23m 43s	10
Chennai	Hindi	GHAJINI	23m 27s	7
Chennai	Hindi	AANGILA PADAM	22m 57s	6
Chennai	Telugu	BANDIPOTU SIMHAM (REVISED)	22m 47s	22
Chennai	Telugu	PAGA PATTINA SIMHAM (REVISED)	22m 28s	18
Cuttack	Bhojpuri	IDDARAMMAYILATHO	28m 53s	7
Cuttack	Odia	TOR MOR LOVE STORY	27m 57s	10
Cuttack	Awadhi	VETTAIKAARAN	26m 25s	6
Cuttack	Odia	MU LOVER NO - 1	26m 10s	11
Cuttack	Bhojpuri	V I P 2	25m 56s	6
Cuttack	Odia	PREMA TOTE DURU JUHARA	25m 54s	7
Cuttack	Odia	JHIATE TO PARI	25m 43s	6
Cuttack	Odia	DASAHARA	23m 23s	7
Cuttack	Bhojpuri	MISHRA FOR SALE	23m 12s	7
Cuttack	Bhojpuri	SOWKHYAM	23m 02s	6
Delhi	Hindustani	LAKSHMI	28m 11s	6
Delhi	Hindustani	ROWDY MUNNA	27m 39s	7
Delhi	Bhojpuri	KHALEJA (MAHESH KHALEJA)	27m 22s	7
Delhi	Odia	CHANDRAMUKHI	27m 00s	7
Delhi	Hindustani	D J	26m 50s	6
Delhi	Hindustani	SUPER	26m 35s	7
Delhi	Awadhi	POWER RETURNS ( RACE GURRAM )	24m 47s	5
Delhi	Hindustani	RACHHA	23m 30s	5
Delhi	Odia	DHARMADURAI	22m 31s	8
Delhi	Odia	TARAK	22m 28s	8
Guwahati	Hindi	CAPITAL	3m 00s	5
Guwahati	Manipuri	NGAMNABA LANFAMSE	2m 45s	8
Guwahati	Bodo	BEKAR ROMEO	1m 38s	6
Guwahati	Hindi Dubbed	BARUN RAI AND THE HOUSE ON THE CLIFF	1m 34s	10
Guwahati	Assamese	RAKSHAK - THE SAVIOUR	1m 08s	3
Guwahati	Assamese	BAD BOYS	0m 50s	2
Guwahati	Assamese	RONGATAPU 1982	0m 32s	3
Guwahati	Khasi	KYNJAH	0m 27s	7
Guwahati	Hindi	KOOKI	0m 24s	6
Guwahati	Assamese	MOI ETI NIXHASOR (KODUWA - THE NIGHTBIRD)	0m 18s	1
Hyderabad	Telugu	"NARAKASURA"	25m 00s	15
Hyderabad	Telugu	NANI	23m 51s	40
Hyderabad	Telugu	ALLUDA MAJAAKAA	23m 11s	38
Hyderabad	Malayalam	SIMHA MUGHAM	20m 46s	27
Hyderabad	Telugu	"I LOVE YOU IDIOT"	18m 54s	1
Hyderabad	Hindi	SAMAJAVARAGAMANA	18m 39s	6
Hyderabad	Telugu	AASHAADAM PELLIKODUKU ( A TO UA)	17m 03s	19
Hyderabad	Hindi	KONAPURAMLO JARIGINA KATHA	16m 48s	7
Hyderabad	Hindi	"SPARK LIFE"	16m 33s	16
Hyderabad	Telugu	1948 AKHANDA BHARATH	16m 25s	19
Kolkata	Bengali	BAGHER BACCHA	28m 24s	8
Kolkata	Bengali	DESHBHOKTO KHILADI	27m 57s	8
Kolkata	Bengali	PYAR KI JUNG	21m 39s	6
Kolkata	Bengali	VANDE BHARAT (SAVE INDIA)	16m 15s	4
Kolkata	Bengali	MAAHIYA (REVISED)	13m 23s	11
Kolkata	Bengali	MAHANAYAK UTTAM KUMAR ( THE METRO STATION ) ( REVISED)	11m 43s	9
Kolkata	Bengali	PORNOMOCHI (REVISED)	10m 24s	44
Kolkata	Bengali	LAAL RANGER DUNIYA (REVISED)	9m 36s	65
Kolkata	Bengali	KOLIJUGER SHIKHONDI	9m 23s	4
Kolkata	Bengali	CIRCLE (REVISED)	9m 19s	21
Mumbai	Hindustani	MASS	29m 52s	7
Mumbai	Bhojpuri	DAROGA BABUNI	29m 50s	35
Mumbai	Bhojpuri	JIYAB SHAAN SE	29m 46s	6
Mumbai	Hindustani	DON ( DON SEENU)	29m 44s	6
Mumbai	Urdu	PALAHATI BRAHMA NAIDU - SHEZADA	29m 43s	6
Mumbai	Hindustani	NAKSHATRAM	29m 43s	1
Mumbai	Bengali	SAPTAMA INDRIYA	29m 38s	6
Mumbai	Hindi	HAI	29m 37s	9
Mumbai	Hindustani	AATAADISTHA	29m 35s	7
Mumbai	Hindi	DARINGBAAZ KHILADI RETURNS (THOOTTAL POO MALARUM)	29m 24s	8
Thiruvananthpuram	Kannada	MARCO	29m 32s	2
Thiruvananthpuram	Malayalam	THANKAMANI	26m 05s	8
Thiruvananthpuram	Malayalam	ARDHARAATHRI (REVISED)	25m 10s	21
Thiruvananthpuram	Malayalam	ARTHARAATHRI PANTHRANDU MUTHAL AARU VARE	25m 07s	4
Thiruvananthpuram	Malayalam	PANI	24m 20s	5
Thiruvananthpuram	Malayalam	ATTENTION PLEASE	22m 28s	3
Thiruvananthpuram	Malayalam	VIRAL 2020	22m 24s	2
Thiruvananthpuram	Malayalam	UDAL	21m 39s	69
Thiruvananthpuram	Malayalam	KARIMPULI (RE-REVISED)	21m 32s	17
Thiruvananthpuram	Malayalam	ARIKU	20m 10s	4

# Export top censored films data with slug included
write_json(top_censored_by_office, "top_censored_films.json", pretty = TRUE, auto_unbox = TRUE)

1.11 Further Information

For more details, interactive charts, and the full explorer, please visit our project website: https://cbfc.watch

The analysis was conducted by Aman Bhargava and Vivek Matthew for Diagram Chasing.

1.11.1 How to Cite

Bhargava, A., Matthew, V., & Diagram Chasing. (2025). Analyzing Film Censorship in India: CBFC Watch. Retrieved from https://cbfc.watch.

BibTeX Entry

For use in LaTeX documents, you can use the following BibTeX entry:

@misc{Bhargava2025CBFC,
  author = {Bhargava, Aman and Matthew, Vivek and {Diagram Chasing}},
  title  = {Analyzing Film Censorship in India: CBFC Watch},
  year   = {2025},
  month  = {September},
  howpublished = {\url{https://cbfc.watch}},
  note   = {Analysis and visualizations by Diagram Chasing. Last accessed: \today}
}