Web Scraping into an E-book with BeautifulSoup and Pandoc

I recently learned how to use BeautifulSoup, a Python library for manipulating HTML and XML parse trees, and it’s been a fantastic addition to my virtual toolbelt. In the past when I’ve needed to process raw HTML, I’ve tried nasty hacks with Unix pipes, or routing the content through a web browser so that I could manipulate it via the DOM API. None of that worked very well, but now I finally have BeautifulSoup fill that gap. It’s got a selector interface and, except for rendering, it’s is basically as comfortable with HTML as JavaScript.

Today’s problem was that I wanted to read a recommended online book called Interviewing Leather, a story set “in a world where caped heroes fight dastardly villains on an everyday basis.” I say “online book” because the 39,403 word story is distributed as a series of 14 blog posts. I’d rather not read it on the website in a browser, instead preferring it in e-book form where it’s more comfortable. The last time I did this, I manually scraped the entire book into Markdown, spent a couple of weeks editing it for mistakes, and finally sent the Markdown to Pandoc to convert into an e-book.

For this book, I just want a quick-and-dirty scrape in order to shift formats. I’ve never read it and I may not even like it, so I definitely don’t want to spend much time on the conversation. Despite having fun with typing lately, I’d also prefer to keep all the formating — italics, etc. — without re-entering it all manually.

Fortunately Pandoc can consume HTML as input, so, in theory, I can feed it the original HTML and preserve all of the original markup. The challenge is that the HTML is spread across 14 pages surrounded by all the expected blog cruft. I need some way to extract the book content from each page, concatenate it together along with chapter headings, and send the result to Pandoc. Enter BeautifulSoup.

First, I need to construct the skeleton HTML document. Rather than code my own HTML, I’m going to build it with BeautifulSoup. I start by creating a completely empty document and adding a doctype to it.

from bs4 import BeautifulSoup, Doctype

doc = BeautifulSoup()

Next I create the html root element, then add the head and body elements. I also add a title element. The original content has fancy Unicode markup — left and right quotation marks, em dash, etc. — so it’s important to declare the page as UTF-8, since otherwise these characters are likely to be interpreted incorrectly. It always feels odd declaring the encoding within the content being encoded, but that’s just the way things are.

html = doc.new_tag('html', lang='en-US')
head = doc.new_tag('head')
meta = doc.new_tag('meta', charset='utf-8')
title = doc.new_tag('title')
title.string = 'Interviewing Leather'
body = doc.new_tag('body')

If I print(doc.prettify()) then I see the skeleton I want:

<!DOCTYPE html>
<html lang="en-US">
  <meta charset="utf-8"/>
   Interviewing Leather

Next, I assemble a list of the individual blog posts. When I was actually writing the script, I first downloaded them locally with my favorite download tool, curl, and ran the script against local copies. I didn’t want to hit the web server each time I tested. (Note: I’ve truncated these URLs to fit in this article.)

chapters = [

I visit a few of these pages in my browser to determine which part of the page I want to extract. I want to look closely enough to see what I’m doing, but not too closely as to not spoil myself! Right clicking the content in the browser and selecting “Inspect Element” (Firefox) or “Inspect” (Chrome) pops up a pane to structurally navigate the page. “View Page Source” would work, too, especially since this is static content, but I find the developer pane easier to read. Plus it hides most of the content, revealing only the structure.

The content is contained in a div with the class entry-content. I can use a selector to isolate this element and extract its child p elements. However, it’s not quite so simple. Each chapter starts with a bit of commentary that’s not part of the book, and I don’t want to include in my extract. It’s separated from the real content by an hr element. There’s also a footer below another hr element, likely put there by someone who wasn’t paying attention to the page structure. It’s not quite the shining example of semantic markup, but it’s regular enough I can manage.

  <main class="site-main">
    <div class="entry-body">
      <div class="entry-content">
        <p>A little intro.</p>
        <p>Some more intro.</p>
        <p>Actual book content.</p>
        <p>More content.</p>
        <p>Footer navigation junk.</p>

The next step is visiting each of these pages. I use enumerate since I want the chapter numbers when inserting h1 chapter elements. Pandoc will use these to build the table of contents.

for i, chapter in enumerate(chapters):
    # Construct h1 for the chapter
    header = doc.new_tag('h1')
    header.string = 'Chapter %d' % (i + 1,)

Next grab the page content using urllib and parse it with BeautifulSoup. I’m using a selector to locate the div with the book content.

    # Load chapter content
    with urllib.request.urlopen(chapter) as url:
        page = BeautifulSoup(url)
    content = page.select('.entry-content')[0]

Finally I iterate over the child elements of the div.entry-content element. I keep a running count of the hr element and only extract content when we’ve seen exactly one hr element.

    # Append content between hr elements
    hr_count = 0
    for child in content.children:
        if (child.name == 'hr'):
            hr_count += 1
        elif (child.name == 'p' and hr_count == 1):
            child.attrs = {}
            if (child.string == '#'):

If it’s a p element, I copy it into the output document, taking a moment to strip away any attributes present on the p tag, since, for some reason, some of these elements have old-fashioned alignment attributes in the original content.

The original content also uses the text “#” by itself in a p to separate sections rather than using the appropriate markup. Despite being semantically incorrect, I’m thankful for this since more hr elements would have complicated matters further. I convert these to the correct markup for the final document.

Finally I pretty print the result:


Alternatively I could pipe it through tidy.

$ python3 extract.py | tidy -indent -utf8 > output.html

A brief inspection with a browser indicates that everything seems to have come out correctly. I won’t know for sure, though, until I actually read through the whole book. Finally I have Pandoc perform the conversion.

$ pandoc -t epub3 -o output.epub output.html 

And that’s it! It’s ready to read offline in my e-book reader of choice. The crude version of my script took around 15–20 minutes to write and test, so I had an e-book conversion in under 30 minutes. That’s about as long as I was willing to spend to get it. Tidying the script up for this article took a lot longer.

I don’t have permission to share the resulting e-book, but I can share my script so that you can generate your own, at least as long as it’s hosted at the same place with the same structure.

Load Comments

null program

Chris Wellons