This report shows a clean, reader-friendly workflow for analyzing one sustained email correspondence. The focus is to surface tone, cadence, and thematic signals that are easy to interpret for marketing analysis.
data_root <- "data"
output_root <- "outputs"
min_emails <- 20
min_span_days <- 5
max_terms <- 15
plot_window_days <- 5
min_date <- as.POSIXct("1995-01-01", tz = "UTC")
brand_ink <- "#2f2f2f"
brand_teal <- "#1f4e5f"
brand_clay <- "#8c3b2a"
brand_sand <- "#e5d5b3"
brand_olive <- "#8a9a5b"library(dplyr)
library(readr)
library(stringr)
library(tidyr)
library(lubridate)
library(ggplot2)
library(tidytext)
library(igraph)
library(ggraph)
theme_set(theme_minimal(base_size = 12))
theme_update(panel.grid.minor = element_blank(), plot.title.position = "plot")extract_header_field <- function(header, field) {
pattern <- paste0("(?m)^", field, ":\\s*.*$")
value <- str_extract(header, pattern)
str_remove(value, paste0("(?m)^", field, ":\\s*"))
}
clean_body <- function(text) {
text %>%
str_replace_all("\\r", "") %>%
str_replace("(?s)-----Original Message-----.*$", " FORWARDED_MESSAGE ") %>%
str_replace("(?s)----- Forwarded by.*$", " FORWARDED_MESSAGE ") %>%
str_replace_all("\\S+@\\S+", "EMAIL_ADDRESS") %>%
str_replace_all("[\\n\\t]+", " ") %>%
str_squish()
}We parse the header fields, clean the body text, and keep only the metadata needed for correspondence analysis.
email_path <- file.path(data_root, "emails.csv")
if (!file.exists(email_path)) {
stop("Missing input file: ", email_path)
}
emails_raw <- read_csv(email_path, show_col_types = FALSE)
required_cols <- c("file", "message")
missing_cols <- setdiff(required_cols, names(emails_raw))
if (length(missing_cols) > 0) {
stop("Missing columns: ", paste(missing_cols, collapse = ", "))
}
emails_parsed <- emails_raw %>%
mutate(message = str_replace_all(message, "\\r", "")) %>%
mutate(
header_end = str_locate(message, "\\n\\n")[, 1],
header = if_else(!is.na(header_end), str_sub(message, 1, header_end - 1), ""),
body_raw = if_else(!is.na(header_end), str_sub(message, header_end + 2), message),
body = clean_body(body_raw),
date_raw = extract_header_field(header, "Date"),
from = str_to_lower(extract_header_field(header, "From")),
to = str_to_lower(extract_header_field(header, "To")),
subject = extract_header_field(header, "Subject"),
date_raw = str_replace(date_raw, "\\s*\\(.*\\)$", "")
) %>%
mutate(
date = parse_date_time(
date_raw,
orders = c("a, d b Y H:M:S z", "a, d b Y H:M z", "d b Y H:M:S z", "d b Y H:M z"),
tz = "UTC"
)
) %>%
select(file, date, from, to, subject, body) %>%
filter(!is.na(date), date >= min_date)We keep single-recipient emails, build bidirectional pairs, and select the top pair that meets the threshold.
emails_clean <- emails_parsed %>%
filter(!is.na(date), !is.na(from), !is.na(to)) %>%
mutate(to_count = str_count(to, ",|;") + 1) %>%
filter(to_count == 1) %>%
separate_rows(to, sep = "\\s*,\\s*|\\s*;\\s*") %>%
mutate(to = str_trim(to)) %>%
filter(to != "", from != "", from != to)
emails_pairs <- emails_clean %>%
mutate(
pair_id = if_else(
from < to,
paste(from, to, sep = " <-> "),
paste(to, from, sep = " <-> ")
)
)
pair_stats <- emails_pairs %>%
group_by(pair_id) %>%
summarise(
email_count = n(),
start_date = min(date),
end_date = max(date),
span_days = as.numeric(difftime(end_date, start_date, units = "days")),
.groups = "drop"
) %>%
filter(email_count >= min_emails, span_days > min_span_days) %>%
arrange(desc(email_count), desc(span_days))
if (nrow(pair_stats) == 0) {
stop("No correspondence pair meets the thresholds. Adjust min_emails or min_span_days.")
}
selected_pair_id <- pair_stats$pair_id[2]
pair_members <- str_split(selected_pair_id, " <-> ", simplify = TRUE)
selected_pair <- emails_pairs %>%
filter(pair_id == selected_pair_id) %>%
arrange(date) %>%
mutate(email_id = row_number())
pair_summary <- tibble(
pair_id = selected_pair_id,
participant_a = pair_members[1],
participant_b = pair_members[2],
email_count = nrow(selected_pair),
start_date = min(selected_pair$date),
end_date = max(selected_pair$date),
span_days = as.numeric(difftime(max(selected_pair$date), min(selected_pair$date), units = "days"))
)
knitr::kable(pair_summary)| pair_id | participant_a | participant_b | email_count | start_date | end_date | span_days |
|---|---|---|---|---|---|---|
| all.worldwide@enron.com <-> enron.announcements@enron.com | all.worldwide@enron.com | enron.announcements@enron.com | 2206 | 1999-08-18 17:54:00 | 2001-07-08 17:31:00 | 689.984 |
The descriptive layer helps explain cadence and responsiveness before we interpret tone.
selected_pair <- selected_pair %>%
mutate(
day = as.Date(date),
direction = if_else(
from == pair_members[1],
paste(pair_members[1], "->", pair_members[2]),
paste(pair_members[2], "->", pair_members[1])
)
)
midpoint_date <- min(selected_pair$date) + (max(selected_pair$date) - min(selected_pair$date)) / 2
window_start <- midpoint_date - days(plot_window_days)
window_end <- midpoint_date + days(plot_window_days)
selected_pair_window <- selected_pair %>%
filter(date >= window_start, date <= window_end)
gap_hours <- selected_pair %>%
arrange(date) %>%
mutate(gap_hours = as.numeric(difftime(date, lag(date), units = "hours"))) %>%
filter(!is.na(gap_hours))
span_days <- as.numeric(difftime(max(selected_pair$date), min(selected_pair$date), units = "days"))
descriptive_summary <- tibble(
metric = c("Total emails", "Date span (days)", "Emails per day", "Median gap (hours)", "Mean gap (hours)"),
value = c(
nrow(selected_pair),
round(span_days, 1),
round(nrow(selected_pair) / max(span_days, 1), 2),
round(median(gap_hours$gap_hours), 2),
round(mean(gap_hours$gap_hours), 2)
)
)
knitr::kable(descriptive_summary)| metric | value |
|---|---|
| Total emails | 2206.00 |
| Date span (days) | 690.00 |
| Emails per day | 3.20 |
| Median gap (hours) | 0.00 |
| Mean gap (hours) | 7.51 |
daily_volume <- selected_pair %>%
count(day)
ggplot(daily_volume, aes(day, n)) +
geom_col(fill = brand_teal) +
labs(
title = "Email volume over time",
x = "Date",
y = "Emails per day"
)
We score each email using the AFINN lexicon and use that to track tone shifts over time.
email_tokens <- selected_pair %>%
select(email_id, body) %>%
unnest_tokens(word, body) %>%
filter(!str_detect(word, "^\\d+$"))
sentiment_scores <- email_tokens %>%
anti_join(stop_words, by = "word") %>%
inner_join(get_sentiments("afinn"), by = "word") %>%
group_by(email_id) %>%
summarise(sentiment = sum(value), .groups = "drop")
selected_pair <- selected_pair %>%
left_join(sentiment_scores, by = "email_id") %>%
mutate(
sentiment = replace_na(sentiment, 0),
word_count = str_count(body, "\\S+"),
weekday = wday(date, label = TRUE)
)
selected_pair_window <- selected_pair %>%
filter(date >= as.POSIXct("2000-01-01 00:00:00", tz = "UTC"))ggplot(selected_pair_window, aes(date, sentiment, color = direction)) +
geom_hline(yintercept = 0, color = "grey75") +
geom_line(alpha = 0.4) +
geom_point(size = 2) +
scale_color_manual(values = c(brand_teal, brand_clay)) +
labs(
title = "Sentiment over time",
x = "Date",
y = "AFINN sentiment score",
color = "Direction"
)
ggplot(selected_pair, aes(weekday, sentiment, fill = direction)) +
geom_boxplot(alpha = 0.75, outlier.size = 1) +
scale_fill_manual(values = c(brand_teal, brand_clay)) +
labs(
title = "Sentiment by weekday",
x = "Weekday",
y = "AFINN sentiment score",
fill = "Direction"
)
These numbers are document-level sentiment sums for each email in selected_pair_window.
With AFINN, each matched word has an integer sentiment value:
Your code does this:
This view treats each email as a node on a time axis. Node size represents length (semantic load), and color reflects sentiment score.
graph_nodes <- selected_pair_window %>%
arrange(date) %>%
mutate(
node_id = row_number(),
name = as.character(node_id),
x = as.numeric(date),
y = 0,
date_label = format(as.Date(date), "%Y-%m-%d")
) %>%
select(name, everything())
graph_edges <- graph_nodes %>%
mutate(next_name = lead(name)) %>%
filter(!is.na(next_name)) %>%
transmute(from = name, to = next_name)
graph <- igraph::graph_from_data_frame(
graph_edges,
vertices = graph_nodes,
directed = TRUE
)
layout <- create_layout(
graph,
layout = "manual",
x = graph_nodes$x,
y = graph_nodes$y
)
ggraph(layout) +
geom_edge_link(
color = "grey70",
alpha = 0.6,
arrow = grid::arrow(length = grid::unit(2, "mm")),
end_cap = ggraph::circle(2, "mm")
) +
geom_node_point(aes(size = word_count, color = sentiment)) +
geom_node_text(
aes(label = date_label),
nudge_y = 0.08,
size = 2.8,
check_overlap = TRUE
) +
scale_color_gradient2(low = "#7f1d1d", mid = brand_sand, high = brand_teal) +
scale_size(range = c(2, 8)) +
scale_x_continuous(
breaks = function(x) pretty(x, n = 5),
labels = function(x) as.Date(x, origin = "1970-01-01")
) +
labs(
title = "Semantic timeline of the correspondence",
x = "Time",
y = NULL,
color = "Sentiment",
size = "Words"
) +
theme(
axis.text.y = element_blank(),
axis.ticks.y = element_blank()
)
This is a fast, marketing-friendly view of the terms that dominate the conversation.
top_terms <- email_tokens %>%
anti_join(stop_words, by = "word") %>%
count(word, sort = TRUE) %>%
slice_head(n = max_terms)
ggplot(top_terms, aes(reorder(word, n), n)) +
geom_col(fill = brand_olive) +
coord_flip() +
labs(
title = "Top terms in the selected correspondence",
x = NULL,
y = "Occurrences"
)
The following block is a placeholder for a Python + LLM pipeline that will detect questions raised within the correspondence.
# Load the selected correspondence emails exported from R
# Normalize text (strip signatures, quoted replies, and forwarded blocks)
# Use a classifier or prompt-based LLM to detect explicit and implicit questions
# Tag each email with question_type, urgency, and topic
# Aggregate question frequency over time and by sender
# Return a summary table and a trend chart back to the reportThis focused correspondence analysis shows how Liveo can translate raw email text into clear marketing signals: cadence, tone shifts, dominant themes or identifiy explicit and implicit questions. The workflow is compact, interpretable, and easy to extend with LLM-based question detection or richer semantic models when needed.