NAME

Marpa::R2::HTML - High-level HTML Parser

SYNOPSIS

Delete all tables:

    use Marpa::R2::HTML qw(html);

    my $with_table = 'Text<table><tr><td>I am a cell</table> More Text';
    my $no_table   = html( \$with_table, { table => sub { return q{} } });

Delete everything but tables:

    my %handlers_to_keep_only_tables = (
        table  => sub { return Marpa::R2::HTML::original() },
        ':TOP' => sub { return \( join q{}, @{ Marpa::R2::HTML::values() } ) }
    );
    my $only_table = html( \$with_table, \%handlers_to_keep_only_tables );

The above works by turning the original text of the HTML into values and concatenating the values at the top of the parse. The same logic works even if a table is very defective:

    my $with_bad_table = 'Text<tr>I am a cell</table> More Text';
    my $only_bad_table =
        html( \$with_bad_table, \%handlers_to_keep_only_tables );

Delete all comments:

    my $with_comment = 'Text <!-- I am a comment --> I am not a comment';
    my $no_comment   = html( \$with_comment,
        { ':COMMENT' => sub { return q{} } });

By default, text is passed through unchanged, so that the user need only specify semantic actions for those components she wants changed. To change the title of a document:

    my $old_title = '<title>Old Title</title>A little html text';
    my $new_title = html(
        \$old_title,
        {   'title' => sub { return '<title>New Title</title>' }
        }
    );

Delete all elements with a class attribute of "delete_me":

    my $stuff_to_be_edited = '<p>A<p class="delete_me">B<p>C';
    my $edited_stuff       = html( \$stuff_to_be_edited,
        { '.delete_me' => sub { return q{} } });

Marpa::R2::HTML recognizes elements even if they have missing start and/or end tags. Marpa::R2::HTML can supply missing tags:

    sub supply_missing_tags {
        my $tagname = Marpa::R2::HTML::tagname();
        return if $empty_elements{$tagname};
        return ( Marpa::R2::HTML::start_tag() // "<$tagname>\n" )
            . Marpa::R2::HTML::contents() .
            ( Marpa::R2::HTML::end_tag() // "</$tagname>\n" );
    }
    my $html_with_just_a_title = '<title>I am a title and That is IT!';
    my $valid_html_with_all_tags =
        html( \$html_with_just_a_title, { q{*} => \&supply_missing_tags } );

Marpa::R2::HTML understands the hierarchical structure of an HTML document. Finding the maximum nesting depth in elements is straightforward:

    sub depth_below_me {
        return List::Util::max( 0, @{ Marpa::R2::HTML::values() } );
    }
    my %handlers_to_calculate_maximum_element_depth = (
        q{*}   => sub { return 1 + depth_below_me() },
        ':TOP' => sub { return depth_below_me() },
    );
    my $maximum_depth_with_just_a_title = html( \$html_with_just_a_title,
        \%handlers_to_calculate_maximum_element_depth );

Marpa::R2::HTML tracks actual elements, however tagged. The above code returns the same depth for $valid_html_with_all_tags as for $html_with_just_a_title.

DESCRIPTION

Marpa::R2::HTML does "high-level" parsing of HTML. It allows handlers to be specified for elements, terminals and other components in the hierarchical structure of an HTML document. Marpa::R2::HTML is an extremely liberal HTML parser. Marpa::R2::HTML does not reject any documents, no mater how poorly they fit the HTML standards.

THE Marpa::R2::HTML::html STATIC METHOD

The interface to Marpa::R2::HTML is through the Marpa::R2::HTML::html static method. It is the only Marpa::R2::HTML method not part of the API for the semantic actions.

html takes one or more arguments. The first argument is required, and must be a reference to a string to be parsed as HTML. The second and subsequent arguments (all optional) are hash references with handler descriptions. (See the synopsis for several examples of calls using the html method.)

CSS-style Handler Options

Handler descriptions in Marpa::R2::HTML are key-value pairs in a hash. In each pair, the key is a CSS-style handler specifier, and the value is a closure, which is called the action for the handler.

Specifiers are "CSS-style" -- their syntax imitates some of the basic cases of CSS specifiers. No attempt is planned to implement the full CSS specifier syntax.

Supported specifier syntaxes are as follows:

Tagname Specifiers
        table  => sub { return Marpa::R2::HTML::original() },

If a specifier contains no special characters it is taken as the name of an element. (A "special" character is anything except an alphanumeric, a hyphen or an underscore.) Consistent with HTML::Parser's default behavior, element names must be specified in lowercase.

Class Specifiers

A specifier which is a dot or period followed by a name will match any element whose class attribute is the same as the specified name. For example, the specifier ".delete_me" will match any element whose class attribute is "delete_me".

Tagname-Class Pair Specifiers

A specifier which contains a dot or period somewhere other than the first position (such as "span.label") is treated as a dotted tagname-class pair. Its action will be called for any component whose tagname and class attribute both match the specifiers.

The Tagname Wildcard Specifier

A specifier of just an asterisk ("*") matches all elements. Be careful to note that matching all elements is not the same as matching all components. The element wildcard specifier will not match any pseudoclasses.

Pseudoclass Specifiers
     ':COMMENT' => sub { return q{} }

A specifier which begins with a colon (":") matches a pseudoclass. Marpa::R2::HTML defines pseudoclasses to deal with terminals and other non-element components of the HTML hierarchy.

Conflicting Specifiers

At most one semantic action is called for each component. Where an element component matches several specifiers, the action is picked based on the most specific match.

1. Matches by tagname-class pair are the most specific.
2. Matches by class are the next most specific.
3. Matches by tagname are considered less specific than matches by class.
4. The wildcard match is the least specific.

Here's an example:

    my $html = <<'END_OF_HTML';
    <span class="high">High Span</span>
    <span class="low">Low Span</span>
    <div class="high">High Div</div>
    <div class="low">Low Div</div>
    <div class="oddball">Oddball Div</div>
    END_OF_HTML

    our @RESULTS = ();
    Marpa::R2::HTML::html(
        \$html,
        {   q{*} => sub {
                push @RESULTS, 'wildcard handler: ' . Marpa::R2::HTML::contents();
            },
            'div' => sub {
                push @RESULTS, '"div" handler: ' . Marpa::R2::HTML::contents();
            },
            '.high' => sub {
                push @RESULTS, '".high" handler: ' . Marpa::R2::HTML::contents();
            },
            'div.high' => sub {
                push @RESULTS,
                    '"div.high" handler: ' . Marpa::R2::HTML::contents();
            },
            '.oddball' => sub {
                push @RESULTS,
                    '".oddball" handler: ' . Marpa::R2::HTML::contents();
            },
            'body' => sub {undef},
            'head' => sub {undef},
            'html' => sub {undef},
            'p'    => sub {undef},
        }
    );

Here is what $result would contain after the above code was run:

    ".high" handler: High Span
    wildcard handler: Low Span
    "div.high" handler: High Div
    "div" handler: Low Div
    ".oddball" handler: Oddball Div

Details of the Specifier Syntax

For elements and class names only alphanumerics, hyphens and underscores are supported. Elements must be specified in lowercase, but they will match tagnames in the original document on a case-insensitive basis.

Forcing element names to be lowercase follows the default behavior of HTML::Parser, which coerces all tagnames to lowercase. This is consistent with the HTML standards. It is not consistent with the XML standards, and an option to configure this behavior may be added in the future.

Pseudoclass names special to Marpa::R2::HTML are case-sensitive, and must be all uppercase. Lowercase is reserved for CSS pseudoclasses. The CSS standard specifies that its pseudoclass names are case-indifferent. No CSS pseudoclasses are supported at this writing.

PSEUDOCLASSES

Marpa::R2::HTML uses HTML::Parser to do its low-level parsing. HTML::Parser "events" become the terminals for Marpa::R2::HTML.

Besides terminals and elements, three other HTML components are recognized: the SGML prolog (:PROLOG), the SGML trailer (:TRAILER), and the HTML document as a whole (:TOP).

:CDATA

The :CDATA pseudoclass specifies the action for CDATA terminals. Its action is called once for each non-whitespace raw text event that is not reclassed as cruft. (Raw text is text in which any markup and entities should be left as is.)

More precisely, a :CDATA terminal is created from any HTML::Parser text event that has the is_cdata flag on; that contains a non-whitespace character as defined in the HTML 4.01 specification (http://www.w3.org/TR/html4/struct/text.html#h-9.1); and that is not reclassed as cruft.

:COMMENT

The :COMMENT pseudoclass specifies the action for HTML comments. Its action is called once for every HTML::Parser comment event that is not reclassed as cruft.

:CRUFT

The :CRUFT pseudoclass specifies the action for cruft. Its action is called once for every HTML::Parser event that Marpa::R2::HTML reclasses as cruft.

Marpa::R2::HTML reclasses terminals as cruft when they do not fit the structure of an HTML document. One example of a terminal that Marpa::R2::HTML would reclass as cruft is a </head> end tag in the HTML body.

Reclassing terminals as cruft is only done as the last resort. When it can, HTML::Parser forgives violations of the HTML standards and accepts terminals as non-cruft.

Cruft is treated in much the same way as comments. It is preserved, untouched, in the original text view.

:DECL

The :DECL pseudoclass specifies the action for SGML declarations. Its action is called once for every HTML::Parser declaration event that is not reclassed as cruft.

:PCDATA

The :PCDATA pseudoclass specifies the action for PCDATA terminals. Its action is called once for each non-whitespace non-raw text event that is not reclassed as cruft.

More precisely, a :PCDATA terminal is created from any HTML::Parser text event that has the is_cdata flag off; that contains a non-whitespace character as defined in the HTML 4.01 specification (http://www.w3.org/TR/html4/struct/text.html#h-9.1); and that is not reclassed as cruft.

Markup and entities in :PCDATA text are expected to be interpreted eventually, but it can be counter-productive to do this during parsing. An application may, for example, be rewriting a document for display on the web. In that case it will often want to leave markup and entities for the client's browser to interpret.

Marpa::R2::HTML leaves interpretation of markup and entities entirely to the application. An application which chooses to do the interpretation itself may do it in the actions, or deal with it in post-processing. CPAN has excellent tools for this, some of which are part of HTML::Parser.

:PI

The :PI pseudoclass specifies the action for SGML processing instructions. Its action is called once for every HTML::Parser process event that is not reclassed as cruft.

:PROLOG

The :PROLOG pseudoclass specifies the action for SGML prolog. This is the part of the HTML document which precedes the HTML root element. Components valid in the prolog include SGML comments, processing instructions and whitespace.

:TOP

The action specified for the :TOP pseudoclass will be called once and only once in every parse, and will be the last action called in every parse. The :TOP component is the entire physical document, including the SGML prolog, the root element, and the SGML trailer. All the other HTML components in a document will be descendants of the :TOP component.

The :TOP action is unique, in that there is always an action for it, even if one is not specified. The html method returns the value returned by the :TOP action. The default :TOP action returns a reference to a string with the literal text value of all of its descendants.

:TRAILER

The :TRAILER pseudoclass specifies the action for SGML trailer. This is the part of the HTML document which follows the HTML root element. Components valid in the trailer include SGML comments, processing instructions, and whitespace. Cruft can also be found here, though for Marpa::R2::HTML that is a last resort.

:WHITESPACE

A Marpa::R2::HTML :WHITESPACE terminal is created for every HTML::Parser text event that is entirely whitespace as defined in the HTML 4.01 specification (http://www.w3.org/TR/html4/struct/text.html#h-9.1) and that is not reclassed as cruft. Whitespace is acceptable in places where non-whitespace is not, and the difference can be very significant structurally.

VIEWS

I hope the synopsis convinces the reader that the action semantics of Marpa::R2::HTML are natural. This naturalness is achieved at the price of some novelty. This section explains the ideas behind the semantic action API. Depending on taste, readers may want to skip this section and go straight to the API.

The components of an HTML document form a hierarchy, with the :TOP component on top, and the terminals on the bottom. The traditional syntax tree method requires semantic actions to know precisely what children every component will have. This processing model is not a good fit to HTML. Marpa::R2::HTML gives the writer of semantic actions "views" of each component that better fit situations where the number and type of children is unknown or vaguely defined.

Marpa::R2::HTML's semantics focus more widely -- on a component's descendants instead of just its direct children. (The terms ancestor and descendant are used in the standard way: If a component X is above Y in the hierarchy, X is an ancestor of Y; and Y is a descendant of the X.)

The Original View

The original view sees the text of a component as it was originally passed to the parser. The original view never changes. The original view is seen through the "Marpa::R2::HTML::original" API method.

The Terminals View

The terminals view sees the terminals corresponding to the original text of a component. The terminals view never changes. The terminals view is usually seen as part of other views.

At this writing the API does not contain a "pure" terminals view method. For a terminals view of the whole HTML document, HTML::Parser does the job with significantly lower overhead. For views and sections of views with no values defined, the descendants view (described below) is equivalent to the terminals view.

The Values View

When actions are called, they return a value. If that value is defined, it becomes visible to the values view of its ancestors. The values view of a component sees the visible values for its descendants.

The values view is an array, with the values ordered according to the lexical order of the components whose actions returned them. If no descendants have visible values, then the values view is a zero-length array.

The values view is hierarchical. When a component produces a visible value, it makes the values of its descendants disappear. That is, whenever the semantic action for a component X returns anything other than a Perl undef, it has two effects:

  • That return value becomes the visible value associated with component X.

  • All the values previously visible due to semantic actions for the descendants of component X disappear.

Values which disappear are gone forever. There is no mechanism to make them "reappear".

As a special case, if an action for a component returns a Perl undef, not only do the values of all its descendants disappear, the component for the action also will not appear in the values view. When its semantic action returns undef, a component permanently "drops out" of the values view taking all descendants with it. The original view is seen through the "Marpa::R2::HTML::values" API method.

The Literal View

The literal view can be thought of as a mix between the original view and the values view. It sees a text string, like the original view. But unlike the original view, the literal view includes the visible values.

Values appear in the literal view in stringized form. For sections of the original text without visible values, the literal view is the same as the original view. In all Marpa::R2::HTML's views, whether descendants are seen as text or values, they are seen in the original lexical order. The literal view is seen through the "Marpa::R2::HTML::literal" API method.

The Descendants View

Just as the literal view can be thought of as a mix between the original view and the values view, the descendants view can be thought of a mix between the terminals view and the values view.

The descendants view sees an array of elements with data for each of the component's descendants, in lexical order. Where a value is visible, the descendants view sees data for the component with the visible value. Where no value is visible, the descendants view sees data for the terminals. This means that when no values are visible, the descendants view is the same as the terminals view.

The descendants view is implemented via the "Marpa::R2::HTML::descendants" method. It is the most fine-grained and detailed way to look at the descendants of a component. The descendants view can do anything that the other views can do, but the other views should be preferred when they fit the application. Other views are typically more intuitive and efficient.

Views versus Syntax Trees

Views are a generalization of the traditional method for processing semantics: syntax trees. The values view is the view that most closely resembles a syntax tree. But there are important differences.

In its purest form, the syntax tree model required the semantic actions to define exactly how many and what kind of immediate children each node had. Each node in a syntax tree worked with its immediate children. Children in a syntax tree appeared as values.

The values view, on the other hand, sees all its descendants, not just its immediate children, but only if they make themselves visible. Because of this, the values view lends itself to being mixed with other views. The values view allows pieces of the tree to decide when they will come into sight and when they will fall out of view.

Views and Efficiency

In most applications, views are more efficient than syntax trees. In terms of Marpa::R2::HTML views, traditional syntax tree processing corresponds most closely to the values view when every component in the parse has a visible value. For Marpa::R2::HTML this is close to the worst case.

Marpa::R2::HTML optimizes for unvalued components. Unvalued components are represented as terminal spans. Adjacent descendant spans are automatically merged. This means the size and time required do not increase as processing rises up the component hierarchy.

Terminals views are calculated on a just-in-time basis when they are requested through the action API. The terminals view is produced quickly from the merged terminal span.

Original views are also calculated on a just-in-time basis as requested. Each terminal tracks the text it represents as a character span. The original text can be quickly reconstructed as the text in the source document from the first character location of its component's first terminal to the last character location of the component's last terminal.

When a handler does not need to return a value, the most efficient thing to do is to return undef. This reverts that component and all its descendants to the efficient unvalued representation.

THE SEMANTIC ACTION API

Marpa::R2::HTML's semantic action API is implemented mainly through context-aware static methods. No arguments are passed to the user's semantics action callbacks. Instead the semantic actions get whatever data they need by calling these static methods.

API Static Methods

Marpa::R2::HTML::attributes

Returns a hash ref to the attributes of the start tag. This hash ref is exactly the hash ref returned for the attr arg specification of HTML::Parser. The attributes API method returns an empty hash if there were no attributes, if there was no start tag for this element, or if the current component is not an element.

Marpa::R2::HTML::contents

For an element, returns the literal view of the contents. The contents of an element are its entire text except for its start tag and its end tag. For an non-element component, returns undef.

Marpa::R2::HTML::descendants

This static method implements the descendants view. It takes one argument, the "dataspec". The dataspec is a string specifying the data to be returned for each descendant. The descendants method returns a reference to an array with one element per descendant, in lexical order. Each element in the array is a reference to an array whose elements are the per-descendant data requested in the string.

The descendant data specification string has a syntax similar to that of the argspec strings of HTML::Parser. Details of that syntax are given below

Marpa::R2::HTML::values

The Marpa::R2::HTML::values method implements the values view. It returns a reference to an array of the descendant values visible from this component, in lexical order. No elements of this array will be undefined. The array will be zero length if no descendant has a visible value.

Marpa::R2::HTML::end_tag

For an element with an explicit end tag, returns the original text of the end tag. For non-element components, returns undef. For elements with no end tag, returns undef.

Marpa::R2::HTML::literal

The Marpa::R2::HTML::literal method implements the literal view. Returns a string containing the literal view of the component -- its text as modified by any the visible values of its descendants.

Marpa::R2::HTML::literal_ref

Returns a reference to a string containing the literal view of the component. This can be useful for very long strings.

Marpa::R2::HTML::offset

Returns the start offset of the component. This is a zero-based location in the source document. Some components are zero-length, containing none of the tokens in the physical input. The Marpa::R2::HTML::offset method return undef for these.

Marpa::R2::HTML::original

The Marpa::R2::HTML::original method implements the original view. Returns a string containing the original view of the component -- its text unchanged from the source document.

Marpa::R2::HTML::start_tag

For an element with an explicit start tag, returns the original text of the start tag. For non-element components, returns undef. For elements with no explicit start tag, returns undef.

Marpa::R2::HTML::tagname

For an element component, returns its tagname. There is a tagname even if there are no explicit tags. Tagname is determined based on structure. For non-element components, returns undef.

Marpa::R2::HTML::title

Returns the value of the title attribute. For a non-element component, returns undef. If there was no explicit start tag, returns undef. If there was no title attribute, returns undef.

Dataspecs

    Marpa::R2::HTML::descendants('token_type,literal,element')

The data specification string, or dataspec, is a comma separated list of descendant data specifiers. The Marpa::R2::HTML::descendants method takes a dataspec as its argument. The Marpa::R2::HTML::descendants method returns a reference to an array of references to arrays of per-descendant data. The contents of the per-descendant data arrays and their order is as specified by the dataspec. These are the valid descendant data specifiers:

element

For an element descendant, returns the tagname. A valid tagname is returned even if there were no explicit tags. For non-element descendants, returns undef.

literal

Returns a string containing the literal view of the descendant.

original

Returns a string containing the original view of the descendant.

token_type

If the descendant is a terminal, returns the token type. These token types are the event types from HTML::Parser: "T" for text, "S" for a start tag, "E" for an end tag, "PI" for a processing instruction, "D" for an SGML declaration, and "C" for a comment. For components with visible values, returns undef.

value

For element descendants with a value, returns that value. In all other cases, returns undef.

The Instance Hash

Each Marpa::R2::HTML instance makes available a per-instance variable as a scratchpad for the application: $Marpa::R2::HTML::INSTANCE. Each call to Marpa::R2::HTML::html creates a $Marpa::R2::HTML::INSTANCE variable which is reserved for that application using the local keyword. Marpa::R2::HTML::html initializes it to an empty hash, but after that does not touch it. When programming via side effects is more natural than passing data up the parse tree (and it often is), $Marpa::R2::HTML::INSTANCE can be used to store the data.

Ordinarily, $Marpa::R2::HTML::INSTANCE is destroyed, with the rest of the parse instance, when Marpa::R2::HTML::html returns. But it can be useful for the :TOP semantic action to return $Marpa::R2::HTML::INSTANCE as the value of the parse.

Undefined Actions versus Actions Which Return undef

It is worth emphasizing that the effect of not defining a semantic action for a component is different from the effect of defining a semantic action which returns a Perl undef. The difference lies in what happens to any visible values of the descendants of that component.

Where no action is defined for a component, it leaves all that component's views as they were before. That is, all values which were visible remain visible and no new values become visible. When an action is defined for a component, but that action returns undef, no new values become visible, and all descendant values which were visible disappear.

Root Element versus :TOP Pseudoclass

It is important to understand the very special function of the :TOP component, and to avoid confusing it with the HTML root element. The most important distinctions are that

  • The semantic action for :TOP pseudoclass is always the last action to be called in a parse.

  • The :TOP component is always the entire HTML document. This can be true of the root element, but it is not true in all cases.

  • The value that the action for the :TOP component returns becomes the value that the Marpa::R2::HTML::html method returns.

The root element is the HTML element whose tagname is "html", though its start and end tags are optional and can be omitted even in strictly valid HTML. Tags or no tags, every HTML document has a root element. (The :TOP component is not an element, so it does not have a tagname and never has tags.)

The root element is always a descendant of the :TOP component. The SGML prolog and SGML trailer are always descendants of the :TOP component. The SGML prolog and SGML trailer are never descendants of the root element.

If an action for the root element is specified, it will also be called once and only once in every parse. An action for the root element can be specified in same way as actions for other elements, using its tagname of "html". An element wildcard action also becomes the action for the root element, if no more specific handler declaration takes precedence.

A :TOP action will be called once and only once in every parse. The :TOP action is unique in that there is a default action. No other component has a default action.

Tags versus Structure

Where tags conflict with structure, HTML::Parser follows structure. "Following structure" means that, for example, if semantic actions for the html, head, and body elements exist, they will be called once and only once during every parse.

Consider this short and very defective HTML document:

    <title>Short</title><p>Text</head><head>

HTML::Parser starts the HTML document's body when it encounters the <p> start tag. That means that, even if they were in the right order, the two head tags cannot be fit into any reasonable parse structure.

If an action is specified for the head element, it will be called for the actual header, and the original view of the head element component will be the text "<title>Short</title>". The action for the head element will not be called again. The two stray tags, </head> and <head>, will be treated as descendants of the body element, and reclassed as "cruft" terminals.

Explicit and Implicit Elements

If a semantic action is specified for a tagname, it is called whenever an element is found with that tagname, even if there are no explicit tags for that element. The HTML standards allow both start and end tags to be missing for html, head, body and tbody elements. Marpa::R2::HTML is more liberal, and will recognize virtual tags for table, tr, and td elements as required to repair a defective table.

Marpa::R2::HTML is more even liberal about recognizing virtual end tags than it is about start tags. Virtual start tags are recognized only for the specific elements listed above. For any non-empty HTML element, there is some circumstance under which Marpa::R2::HTML will recognize a virtual end tag. At end of file, as one example, Marpa::R2::HTML will do its best to produce a balanced HTML structure by creating a virtual end tag for every element in the stack of currently active elements.

EXPORTS

Marpa::R2::HTML exports nothing by default. Optionally, Marpa::R2::HTML::html may be exported.

Copyright and License

  Copyright 2012 Jeffrey Kegler
  This file is part of Marpa::R2.  Marpa::R2 is free software: you can
  redistribute it and/or modify it under the terms of the GNU Lesser
  General Public License as published by the Free Software Foundation,
  either version 3 of the License, or (at your option) any later version.

  Marpa::R2 is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  Lesser General Public License for more details.

  You should have received a copy of the GNU Lesser
  General Public License along with Marpa::R2.  If not, see
  http://www.gnu.org/licenses/.