Herodotus' Histories

Florian Klohn, 21 May 2022 (first version 13 May 2022)

#Preliminaries:
knitr::opts_chunk$set(message=FALSE, warning=FALSE, eval = FALSE) #set eval = TRUE when run first

rm(list=ls())

library(tidyverse)
library(tidytext)
library(gutenbergr)
library(ggwordcloud)
library(topicmodels) 
library(ldatuning)
library(ggsci)

theme_set(theme_light())

dir.create("Histories", showWarnings = FALSE)

Introduction

Herodotus (484 – 425 BC), the first known author to do a systematic investigation of historical events, is famous for his important work on the Greco-Persian wars: Histories, originally meaning ‘critical inquiry’, is therefore foundational to the Western historian tradition. The book accompanied the well-known journalist and author Ryszard Kapuściński on his journeys across the globe, and it is on my reading list for some time now.

However, maybe due to the book’s old age or other priorities, I have not managed to read it so far. How can we use data science methods to have a first impression of the book’s content instead? Fortunately, text mining tools are available for natural language processing that can help us do the job. Interesting sources concerning text mining in R are eg. LADAL and Automated Content Analysis with R. Data scientists interested in history may find Programming Historian quite illuminating. But now let us play a little bit with the text using tidytext.

Word Count

One first way to look at the ‘Histories’ is by using the number of words. What are the most prevalent words in the book, after excluding stop words (ie. irrelevant words)?

#Download data:
#gutenberg_metadata
#gutenberg_works(str_detect(title, "Herodotus"))
#hist_in <- gutenberg_download(c(2707, 2456))
#saveRDS(hist_in, "Histories/hist_gut.rds")
hist_in <- readRDS("Histories/hist_gut.rds")

#preparation: add chapter-information to data:
hist_df <- hist_in %>%
  arrange(desc(gutenberg_id)) %>%
  mutate(chapter=case_when(str_sub(text, 1, 5)=="BOOK " ~ str_sub(text, 1, 9), 
                           str_sub(text, 1, 8)=="NOTES TO" ~ str_sub(text, 1, 8))) %>%
  fill(chapter) %>%
  filter(str_sub(chapter, 1, 4)=='BOOK') %>% #removing footnotes
  mutate(line=row_number(), chapter=if_else(str_sub(chapter, 9,9)=='T', str_sub(chapter, 1,7), chapter),
         chapter=as_factor(chapter))

#tidy + clean data: exclude stop words + numbers
hist_tidy <- hist_df %>%
  unnest_tokens(word, text) %>% # -> tidy data
  anti_join(stop_words) %>% # exclude stop words
  filter(is.na(as.numeric(word))) %>% # exclude numbers
  filter(!word %in% c("thou", "thee", "thy", "ye")) #drop very frequent old words not detected by stop_words

#create wordcloud (as shown above):
set.seed(42)
hist_wordcloud <- hist_tidy %>%
  group_by(word) %>%
  summarize(num=n(), .groups="drop") %>%
  arrange(desc(num)) %>%
  slice(1:100) %>%
  mutate(angle = 90 * sample(c(0, 1), n(), replace = TRUE, prob = c(80, 20))) %>%
  ggplot(aes(label = word, size=num, angle=angle)) +
  geom_text_wordcloud(shape="square") + #, ylim=c(0,1), xlim=c(0.25, 0.85)
  theme_minimal() +
  scale_size_area(max_size = 15)
ggsave(hist_wordcloud, file="Histories/p0_hist_wordcloud.png") # import plot via markdown


#plot word count as bar char:
p1_hist_tidy <- hist_tidy %>%
  count(word, sort = TRUE) %>%
  filter(n > 300) %>%
  mutate(word = reorder(word, n)) %>%
  ggplot(aes(n, word)) +
  geom_col( fill='darkgreen') +
  labs(y = 'word', x='number occurrences', title='Figure 1: Most frequent words in Histories')
ggsave(p1_hist_tidy, file="Histories/p1_hist_tidy.png")  # import via markdown

We see that time, land, son, and king are on the top of the list followed by the main cultural players of the story – the Persians and the Greeks, represented by Hellenes and Athenians.

A more sophisticated approach for revealing representative document words is to focus on important words that are most distinctive to each of the book’s 9 chapters. We do this by weighting the term frequency (tf) with the inverse document frequency (idf).

hist_tf_idf <- hist_df %>%
  unnest_tokens(word, text) %>%
  group_by(chapter, word) %>%
  summarize(n=n()) %>% 
  bind_tf_idf(word, chapter, n) %>% #calculate inverse document frequency
  arrange(desc(tf_idf))

p2_hist_tf_idf <- hist_tf_idf %>%
  group_by(chapter) %>%
  slice_max(tf_idf, n = 10) %>%
  ungroup() %>%
  mutate(tf_idf=tf_idf*100, word=reorder_within(word, tf_idf, chapter)) %>% 
  ggplot(aes(x=tf_idf, y=word, fill = chapter)) + 
  geom_col(show.legend = FALSE) +
  scale_y_reordered() + #override ggplot default
  facet_wrap(~chapter, ncol = 3, scales = "free") +
  scale_fill_simpsons() +
  labs(x = "term frequency (in %) * inverse document frequency", y = 'words', title='Figure 2: Most distinctive words per chapter') 

ggsave(p2_hist_tf_idf, file="Histories/p2_hist_tf_idf.png") 

These most distinctive words per chapter show a strong emphasis on geographical and biographical details, as expected in historical literature. For example, the second book deals to large extent with Egypt, which can easily be seen by the top keywords.

Topic Modeling

How can we access the content of the book in a more meaningful way by using abstract topics? We do so by applying topic modeling, a statistical framework for unsupervised classification that discovers topics occurring in a collection of documents (here: chapters). We use Latent Dirichlet Allocation (LDA), a typical approach for topic modeling, in order to automatically find both associations between words and the topics, and associations between topics and the book’s nine chapters. Although the number of topics is a priori unknown and serves as an input for the algorithm, there are different approaches to automatically determine the optimal number of topics. So let us first figure out the relevant topic number at first.

#prepare data fro LDA
hist_dtm <- hist_tidy  %>%
  count(chapter, word) %>%
  cast_dtm(term=word, document=chapter, value=n)


ldatuning.metriken <- FindTopicsNumber(hist_dtm, topics = seq(from = 2, to = 15, by = 1), metrics = c("Griffiths2004", "CaoJuan2009", "Arun2010", "Deveaud2014"), method = "Gibbs", control = list(seed = 42), mc.cores = 2)

#What is the optimal number of topics to consider? Well, there are different ways to approach this problem, by maximizing dissimilarity between topics.

#Plot metrics:
#FindTopicsNumber_plot(ldatuning.metriken)

#use ggplot:
p3_top_num <- ldatuning.metriken %>% 
  pivot_longer(cols=c(Griffiths2004, CaoJuan2009, Arun2010, Deveaud2014), names_to='metrics') %>%
  group_by(metrics) %>%
  mutate(min_val=min(value), max_val=max(value), value_norm=(value-min_val)/(max_val-min_val)) %>%
  mutate(opt=case_when(metrics %in% c("CaoJuan2009", "Arun2010") ~ 'minimize', TRUE ~ 'maximize')) %>%
  ggplot(aes(x=topics, y=value_norm)) +
  geom_line(aes(color=metrics), size=2) + 
  scale_x_continuous(breaks = c(2:15)) +
  scale_color_simpsons() +
  facet_wrap(~ opt, ncol=1) +
  labs(y=NULL, x='number of topics', title="Figure 3: Metrics to choose topic number")
ggsave(p3_top_num, file="Histories/p3_top_num.png") 

Given the criteria above we choose 8 as the number of topics in Histories and use them as the input for our final LDA. The result will allow us to extract the most distinctive words for each of the generated topics.

#LDA with 8 topics:
hist_lda <- LDA(hist_dtm, k = 8, control = list(seed = 42), method="Gibbs") 

#word topic association (beta):
topic_word <- tidy(hist_lda, matrix = "beta")

#visualize topics with top-10 word-probabilities per chapter
p4_topic_word <- topic_word %>%
  group_by(topic) %>%
  slice_max(beta, n = 10) %>%
  ungroup() %>%
  mutate(term=reorder_within(term, beta, topic)) %>% 
  ggplot(aes(x=beta, y=term,   fill = topic)) + 
  geom_col(show.legend = FALSE) +
  scale_y_reordered() + 
  scale_fill_material("deep-orange") + #cyan
  facet_wrap(~topic, ncol = 2, scales = "free") +
  labs(x = "beta (term frequency)", y = "term", title='Figure 4: Most important words per topic')

ggsave(p4_topic_word, file="Histories/p4_topic_word.png") 

One approach to understand the 8 topics is to look at their most important words. We can see that topic 7 contains words such as ‘son’, ‘time’, and ‘king’ – words that are also most prevalent overall as shown above. How are the 8 topics distributed over the nine chapters?

#Association between Documents & Themen: (gamma)
doc_topic <- tidy(hist_lda, matrix = "gamma") %>% 
  mutate(document=as_factor(document))

#visualize the assiciation between topics and chapter:
p5_doc_topic <- doc_topic %>%
  ggplot(aes(y=as.factor(topic), x=document,  fill=gamma)) + 
  scale_fill_material("blue-grey") +
  geom_tile(colour="white") +
  theme_minimal(base_size = 8) +
  labs(title="Figure 5: Relationship between topics and chapters", fill="gamma", y="topic", x="chapter")  

ggsave(p5_doc_topic, file="Histories/p5_doc_topic.png") 

A visual inspection suggests all chapters are a mixture of topics 4 and 7 + some individual component. Whereas topic 7 captures ideas of ancestry, topic 4 also contains words that are associated with religion. We also see that chapters 2-3 and 8-9 seem to have some overlap in their content.

So far, we have used tools to figure out what the Histories are about. According to Kapuściński (2007), the main theme of Histories can be described by the following 3 rules, which governed antiquity, in comparison with the modern world:

1. The law of vengeance applies.
2. Human happiness is not permanent.
3. No one can escape fate, even not a god.

From today’s perspective these rules, especially 1 and 3, seem very fatalistic, but they governed people’s life with cycles of crime and punishment over a long period of time. If you know how to derive such an interpretation with text mining methods, please let me know.

One last aspect worth discussing is the fact that Herodotus had a Hellenian background. Was his description of the Persian-Greek conflict culturally biased and how can data analytics support detect this?

Sentiment Analysis

In order to tackle this question we use a sentiment word lexicon in which words are classified as positive or negative. Are the terms ‘Persian’, ‘Hellenic’, and ‘Athenian’ to the same extent associated with positive/negative words in Histories?

As a first step, we generate bi-grams (or word pairs) and extract neighboring words of the terms Persia, Hellen, and Athen, reflecting different cultures. Then these neighboring words are matched with the sentiment lexicon and the number of sentiments is summed up for each of the 3 cultures.

#remove stop words before calculating bi-grams:
bg_herod <- hist_df %>%
  unnest_tokens(word, text) %>% 
  anti_join(stop_words) %>% # exclude stop words stop_words
  filter(is.na(as.numeric(word))) %>% # exclude numbers
  filter(!word %in% c("thou", "thee")) %>% #drop very frequent old words not detected by stop_words
  group_by(chapter, line) %>% #undo unnest words
  summarize(text = str_c(word, collapse = " ")) %>%
  ungroup() %>%
  unnest_tokens(bigram, text, token = "ngrams", n = 2)  %>% #nest n-grams
  count(bigram, sort = TRUE) 

#Which words are associated with the persian, greek or the athenian culture?
#dominate positive or negative sentiments?

#function to create bi-grams containing specific values:
cult_fct <- function(x) {

y <- bg_herod %>%
  filter(str_detect(bigram, x)) %>% #bi-grams containing "persian"
  separate(bigram, c("word1", "word2"), sep = " ") %>%
  mutate(culture=x, word1 = if_else(!str_detect(word1, x), word1, NA_character_ ),
         word2 = if_else(!str_detect(word2, x), word2, NA_character_ )) %>%
  unite(word1, word2, col=bigram, na.rm = TRUE) %>%
  filter(bigram!='')

return(y)

}


comp <- c('persia', 'hellen', 'athen')

#run function  
comp_cult <- comp %>%
  map(cult_fct) %>%
  bind_rows() 
 

#sentiment analysis:
sent_bing <- get_sentiments("bing") 

p6_comp_cult <- comp_cult %>%
  mutate(word=bigram) %>%
  inner_join(sent_bing, by="word") %>% 
  group_by(culture, sentiment) %>%
  summarize(n=sum(n)) %>%
  ggplot(aes(x=n, y=culture ,  fill = sentiment)) + 
  geom_bar(stat="identity", position = "dodge") +
  labs(x = "number sentiments", y = "culture", title='Figure 6: Total number of sentiments by culture')

ggsave(p6_comp_cult, file="Histories/p6_comp_cult.png") 

The results show that negative sentiments dominate all 3 cultures, compared to positive ones. However, it is easy to see that in relation to the total number of sentiments in each culture, the negative sentiments are actually more pronounced for the terms Hellenian and Athenian. So our results do not suggest that Herodotus’ Histories are culturally biased in favor of the Greek culture.

Wrap Up

We have seen how to use text mining tools together with visualization in order to get a first impression of The Histories. In addition to analyzing words, we have seen how to generate abstract topics from text data. Also, sentiment analysis was applied to see whether Herodotus’ description was positively biased toward the Greek culture.

Is the book worth reading after all? After mining Histories the book will certainly stay on my reading list.

References

Gentzkow, Matthew, Bryan Kelly, and Matt Taddy. 2019. “Text as Data.” Journal of Economic Literature 57 (3): 535–74.

Grün, Bettina, and Kurt Hornik. 2011. “topicmodels: An R Package for Fitting Topic Models.” Journal of Statistical Software 40 (13): 1–30. https://doi.org/10.18637/jss.v040.i13.

Kapuściński, Ryszard. 2007. Travels with Herodotus. Random House.

Le Pennec, Erwan, and Kamil Slowikowski. 2019. Ggwordcloud: A Word Cloud Geom for ’Ggplot2’. https://CRAN.R-project.org/package=ggwordcloud.

Nikita, Murzintcev. 2020. Ldatuning: Tuning of the Latent Dirichlet Allocation Models Parameters. https://CRAN.R-project.org/package=ldatuning.

Robinson, David. 2021. Gutenbergr: Download and Process Public Domain Works from Project Gutenberg. https://CRAN.R-project.org/package=gutenbergr.

Silge, Julia, and David Robinson. 2016. “Tidytext: Text Mining and Analysis Using Tidy Data Principles in r.” JOSS 1 (3). https://doi.org/10.21105/joss.00037.

Xiao, Nan. 2018. Ggsci: Scientific Journal and Sci-Fi Themed Color Palettes for ’Ggplot2’. https://CRAN.R-project.org/package=ggsci.