There are two general mechanisms provided by this library that are used to build web scrapers: Selectors and Scrapers.
Selectors describe a location within an HTML DOM tree. The simplest selector,
that can be written is a simple string value. For example, the selector
"div" matches every single div node in a DOM. Selectors can be combined
using tag combinators. The
// operator to define nested relationships within a
DOM tree. For example, the selector
"div" // "a" matches all anchor tags
nested arbitrarily deep within a div tag.
In addition to describing the nested relationships between tags, selectors can
also include predicates on the attributes of a tag. The
@: operator creates a
selector that matches a tag based on the name and various conditions on the
tag's attributes. An attribute predicate is just a function that takes an
attribute and returns a boolean indicating if the attribute matches a criteria.
There are several attribute operators that can be used to generate common
@= operator creates a predicate that matches the name and
value of an attribute exactly. For example, the selector
"div" @: ["id" @=
"article"] matches div tags where the id attribute is equal to
Scrapers are values that are parameterized over a selector and produce a value
from an HTML DOM tree. The
Scraper type takes two type parameters. The first
is the string like type that is used to store the text values within a DOM tree.
Any string like type supported by
Text.StringLike is valid. The second type
is the type of value that the scraper produces.
There are several scraper primitives that take selectors and extract content from the DOM. Each primitive defined by this library comes in two variants: singular and plural. The singular variants extract the first instance matching the given selector, while the plural variants match every instance.
Complete examples can be found in the examples folder in the scalpel git repository.
The following is an example that demonstrates most of the features provided by
this library. Supposed you have the following hypothetical HTML located at
"http://example.com/article.html" and you would like to extract a list of all
of the comments.
<html> <body> <div class='comments'> <div class='comment container'> <span class='comment author'>Sally</span> <div class='comment text'>Woo hoo!</div> </div> <div class='comment container'> <span class='comment author'>Bill</span> <img class='comment image' src='http://example.com/cat.gif' /> </div> <div class='comment container'> <span class='comment author'>Susan</span> <div class='comment text'>WTF!?!</div> </div> </div> </body> </html>
The following snippet defines a function,
allComments, that will download
the web page, and extract all of the comments into a list:
type Author = String data Comment = TextComment Author String | ImageComment Author URL deriving (Show, Eq) allComments :: IO (Maybe [Comment]) allComments = scrapeURL "http://example.com/article.html" comments where comments :: Scraper String [Comment] comments = chroots ("div" @: [hasClass "container"]) comment comment :: Scraper String Comment comment = textComment <|> imageComment textComment :: Scraper String Comment textComment = do author <- text $ "span" @: [hasClass "author"] commentText <- text $ "div" @: [hasClass "text"] return $ TextComment author commentText imageComment :: Scraper String Comment imageComment = do author <- text $ "span" @: [hasClass "author"] imageURL <- attr "src" $ "img" @: [hasClass "image"] return $ ImageComment author imageURL
Tips & Tricks
The primitives provided by scalpel are intentionally minimalistic with the assumption being that users will be able to build up complex functionality by combining them with functions that work on existing type classes (Monad, Applicative, Alternative, etc.).
This section gives examples of common tricks for building up more complex behavior from the simple primitives provided by this library.
AttributeName are all
IsString instances, and
thus it is convenient to use scalpel with
OverloadedStrings enabled. If not
OverloadedStrings, all tag names must be wrapped with
Scalpel has 3 different wildcard values each corresponding to a distinct use case.
anySelectoris used to match all tags:
textOfAllTags = texts anySelector
AnyTagis used when matching all tags with some attribute constraint. For example, to match all tags with the attribute
textOfTagsWithClassButton = texts $ AnyTag @: [hasClass "button"]
AnyAttributeis used when matching tags with some arbitrary attribute equal to a particular value. For example, to match all tags with some attribute equal to
textOfTagsWithAnAttributeWhoseValueIsButton = texts $ AnyTag @: [AnyAttribute @= "button"]
It is possible to run into scenarios where the name and attributes of a tag are not sufficient to isolate interesting tags and properties of child tags need to be considered.
In these cases the
guard function of the
Alternative type class can be
anySelector to implement predicates of arbitrary
Building off the above example, consider a use case where we would like find the html contents of a comment that mentions the word "cat".
The strategy will be the following:
Isolate the comment div using
Then within the context of that div the textual contents can be retrieved with
text anySelector. This works because the first tag within the current context is the div tag selected by chroot, and the
anySelectorselector will match the first tag within the current context.
Then the predicate that
"cat"appear in the text of the comment will be enforced using
guard. If the predicate fails, scalpel will backtrack and continue the search for divs until one is found that matches the predicate.
Return the desired HTML content of the comment div.
catComment :: Scraper String String catComment = -- 1. First narrow the current context to the div containing the comment's -- textual content. chroot ("div" @: [hasClass "comment", hasClass "text"]) $ do -- 2. anySelector can be used to access the root tag of the current context. contents <- text anySelector -- 3. Skip comment divs that do not contain "cat". guard ("cat" `isInfixOf` contents) -- 4. Generate the desired value. html anySelector
For the full source of this example, see complex-predicates in the examples directory.
The pluralized versions of the primitive scrapers (
allow the user to extract content from all of the tags matching a given
selector. For more complex scraping tasks it will at times be desirable to be
able to extract multiple values from the same tag.
Like the previous example, the trick here is to use a combination of the
chroots function and the
Consider an extension to the original example where image comments may contain some alt text and the desire is to return a tuple of the alt text and the URLs of the images.
The strategy will be the following:
to isolate each img tag using
Then within the context of each img tag, use the
anySelectorselector to extract the alt and src attributes from the current tag.
Create and return a tuple of the extracted attributes.
altTextAndImages :: Scraper String [(String, URL)] altTextAndImages = -- 1. First narrow the current context to each img tag. chroots "img" $ do -- 2. Use anySelector to access all the relevant content from the the currently -- selected img tag. altText <- attr "alt" anySelector srcUrl <- attr "src" anySelector -- 3. Combine the retrieved content into the desired final result. return (altText, srcUrl)
For the full source of this example, see generalized-repetition in the examples directory.
scalpel package relies on curl to provide networking support. For small
projects and one off scraping tasks this is likely sufficient. However when
using scalpel in existing projects or on platforms without curl this dependency
can be a hindrance.
For these scenarios users can instead depend on scalpel-core which does not provide networking support and does not depend on curl.
My Scraping Target Doesn't Return The Markup I Expected
Some websites return different markup depending on the user agent sent along with the request. In some cases, this even means returning no markup at all in an effort to prevent scraping.
To work around this, you can add your own user agent string with a curl option.
#!/usr/local/bin/stack -- stack runghc --resolver lts-6.24 --install-ghc --package scalpel-0.4.0 import Network.Curl import Text.HTML.Scalpel main = do html <- scrapeURLWithOpts opts url $ htmls anySelector maybe printError printHtml html where url = "https://www.google.com" opts = [ CurlUserAgent "some user agent string" ] printError = putStrLn "Failed" printHtml = mapM_ putStrLn
A list of user agent strings can be found here.
Building on Windows
Building scalpel on Windows can be a challenge because of the dependency on curl. In order to successfully build scalpel you must download curl and add the following to your stack.yaml file.
extra-lib-dirs: ["C:/Program Files/cURL/dlls"] extra-include-dirs: ["C:/Program Files/cURL/dlls"]
If you do not require network support, you can instead depend on scalpel-core which does not does not depend on curl.
Fix bug (#59, #54) in DFS traversal order.
scalpelinto two packages:
scalpel-core. The latter does not provide networking support and does not depend on curl.
- Add the
chroottricks (#23 and #25) to README.md and added examples.
- Fix backtracking that occurs when using
- Fix bug where the same tag may appear in the result set multiple times.
- Performance optimizations when using the (//) operator.
- Make Scraper an instance of MonadFail. Practically this means that failed
pattern matches in
<-expressions within a do block will evaluate to mzero instead of throwing an error and bringing down the entire script.
- Pluralized scrapers will now return the empty list instead mzero when there are no matches.
- Add the
positionscraper which provides the index of the current sub-tree within the context of a
- Added the
- Added the
matchfunction which allows for the creation of arbitrary attribute predicates.
- Fixed build breakage with GHC 8.0.1.
Make tag and attribute matching case-insensitive.
- Added benchmarks and many optimizations.
selectmethod is removed from the public API.
- Many methods now have a constraint that the string type parametrizing TagSoup's tag type now must be order-able.
scrapeUrlWithConfigthat will hopefully put an end to multiplying
- The default behaviour of the
scrapeUrl*methods is to attempt to infer the character encoding from the
Cleanup stale instance references in documentation of TagName and AttributeName.
Made Scraper an instance of MonadPlus.
Fixed examples in documentation and added an examples folder for ready to compile examples. Added travis tests to ensures that examples remain compilable.
Removed the StringLike parameter from the Selector, Selectable, AttributePredicate, AttributeName, and TagName types. Instead they are now agnostic to the underlying string type, and are only constructable with Strings and the Any type.
Tighten dependencies and drop download-curl all together.
Add the html and html scraper primitives for extracting raw HTML.
- Make scrapeURL follow redirects by default.
- Expose a new function scrapeURLWithOpts that takes a list of curl options.
- Fix bug (#2) where image tags that do not have a trailing "/" are not selectable.
Tighten dependencies on download-curl.
- First version!