Skip to content

Latest commit

 

History

History
210 lines (151 loc) · 7.18 KB

page-classifiers.rst

File metadata and controls

210 lines (151 loc) · 7.18 KB
Raw

Target Page Classifiers

ACHE uses target page classifiers to distinguish between relevant and irrelevant pages. Page classifiers are flexible and can be as simple as a simple regular expression, or a sophisticated machine-learning based classification model.

Configuring Page Classifiers

To configure a page classifier, you will need to create a new directory containing a file named pageclassifier.yml specifying the type of classifier that should be used and its parameters. ACHE contains several page classifier implementations available. The following subsections describe how to configure them:

title_regex

Classifies a page as relevant if the HTML tag title matches a given pattern defined by a provided regular expression. You can provide this regular expression using the pageclassifier.yml file. Pages that match this expression are considered relevant. For example:

type: title_regex
parameters:
  regular_expression: ".*sometext.*"

url_regex

Classifies a page as relevant if the URL of the page matches any of the regular expression patterns provided. You can provide a list of regular expressions using the pageclassifier.yml file as follows:

type: url_regex
parameters:
  regular_expressions: [
    "https?://www\\.somedomain\\.com/forum/.*"
    ".*/thread/.*",
    ".*/archive/index.php/t.*",
  ]

body_regex

Classifies a page as relevant if the HTML content of the page matches any of the regular expression patterns provided. You can provide a list of regular expressions using the pageclassifier.yml file as follows:

type: body_regex
parameters:
  regular_expressions:
  - pattern1
  - pattern2

regex

Classifies a page as relevant by matching the lists of regular expressions provided against multiple fields: url, title, and content. You can provide a list of regular expressions for each of these fields, and also the type of boolean operation to combine the results:

  • AND (default): All regular expressions must match
  • OR: At least one regular expression must match

Besides the combination method for each regular expression within a list, you cab also specify how the final result for each field should be combined. The file pageclassifier.yml should be organized as follows:

type: regex
parameters:
    boolean_operator: AND|OR
    url:
      boolean_operator: AND|OR
      regexes:
        - pattern1-for-url
        - pattern2-for-url
    title:
      boolean_operator: AND|OR
      regexes:
        - pattern1-for-title
        - pattern2-for-title
    content:
      boolean_operator: AND|OR
      regexes:
        - pattern1-for-content

For example, in order to be classified as relevant using the following configuration, a page would have to:

  • match regexes .*category=1.* OR .*post\.php.* in the URL
  • AND
  • it would have to match .*bar.* OR .*foo.* in the title.
type: regex
parameters:
    boolean_operator: "AND"
    url:
      boolean_operator: "OR"
      regexes:
        - .*category=1.*
        - .*post\.php.*
    title:
      boolean_operator: "OR"
      regexes:
        - .*bar.*
        - .*foo.*

smile (a.k.a "weka" before version 0.11.0)

Warning

This classifier was previously known as weka before version 0.11.0, and has been re-implemented using SMILE library which uses a more permissive open-source license (Apache 2.0). If you have models built using a previous ACHE version, you will need to re-build your model before upgrading ACHE to a version equal or greater than 0.11.0.

Classifies pages using a machine-learning based text classifier (SVM, Random Forest) trained using ACHE's buildModel command. A smile classifier consists of a features_file, a model_file, and a stopwords_file file (which contains the stop-words used during the training process):

type: smile
parameters:
  features_file: pageclassifier.features
  model_file: pageclassifier.model
  stopwords_file: stoplist.txt

All these files can be built automatically by training a model using the command ache buildModel, as detailed in the next sub-section. You can also run ache help buildModel to see more options available.

Alternatively, you can use the Domain Discovery Tool (DDT) to gather training data and build automatically these files. DDT is a interactive web-based application that helps the user with the process of training a page classifier for ACHE.

Building a model for the smile page classifier

To create the necessary configuration files, you will need to gather positive (relevant) and negative (irrelevant) examples of web pages to train the page classifier. You should store the HTML content of each web page in a plain text file. These files should be placed in two directories, named positive` and ``negative, which reside in another empty directory. See an example at config/sample_training_data.

Here is how you build a model from these examples using ACHE's command line:

ache buildModel -t <training data path> -o <output path for model> -c <stopwords file path>

where,

  • <training data path> is the path to the directory containing positive and negative examples.
  • <output path> is the new directory that you want to save the generated model that consists of two files: pageclassifier.model and pageclassifier.features.
  • <stopwords file path> is a file with list of words that the classifier should ignore. You can see an example at config/sample_config/stoplist.txt.

Example of building a page classifier using our test data:

ache buildModel -c config/sample_config/stoplist.txt -o model_output -t config/sample_training_data

Testing Page Classifiers

Once you have configured your classifier, you can verify whether it is working properly to classify a specific web page by running the following command:

ache run TargetClassifierTester --input-file {html-file} --model {model-config-directory}

where,

  • {html-file} is the path to a file containing the page's HTML content and
  • {model-config-directory} is a path to the configuration directory containing your page classifier configuration.