collective.classification 0.1b2

Introduction

collective.classification aims to provide a set of tools for automatic document classification. Currently it makes use of the Natural Language Toolkit and features a trainable document classifier based on Part Of Speech (POS) tagging, heavily influenced by topia.termextract. This product is mostly intended to be used for experimentation and development. Currently english and dutch are supported.

What is this all about?

It’s mostly about having fun! The package is in a very early experimental stage and awaits eagerly contributions. You will get a good understanding of what works or not by looking at the tests. You might also be able to do some useful things with it:

1) Term extraction can be performed to provide quick insight on what a document is about. 2) On a large site with a lot of content and tags (or subjects in the plone lingo) it might be difficult to assign tags to new content. In this case, a trained classifier could provide useful suggestions to an editor responsible for tagging content. 3) Similar documents can be found based on term similarity. 4) Clustering can help you organize unclassified content into groups.

How it works?

At the moment there exist the following type of utilities:

• POS taggers, utilities for classifying words in a document as Parts Of Speech. Two are provided at the moment, a Penn TreeBank tagger and a trigram tagger. Both can be trained with some other language than english which is what we do here.

• Term extractors, utilities responsible for extracting the important terms from some document. The extractor we use here, assumes that in a document only nouns matter and uses a POS tagger to find those mostly used in a document. For details please look at the code and the tests.

• Content classifiers, utilities that can tag content in predefined categories. Here, a naive Bayes classifier is used. Basically, the classifier looks at already tagged content, performs term extraction and trains itself using the terms and tags as an input. Then, for new content, the classifier will provide suggestions for tags according to the extracted terms of the content.

• Utilities that find similar content based on the extracted terms.

• Clusterers, utilities that without prior knowledge of content classification can group content into groups according to feature similarity. At the moment NLTK’s k-means clusterer is used.

Installation & Setup

Before running buildout, make sure you have yaml and its python bindings installed (use macports on osx, or your package installer on linux). If nltk exists for your OS you might as well install that, otherwise it will be fetched when you run buildout.

To get started you will simply need to add the package to your “eggs” section and run buildout, restart your Plone instance and install the “collective.classification” package using the quick-installer or via the “Add-on Products” section in “Site Setup”.

WARNING: Upon first time installation linguistic data will be fetched from NLTK’s repository and stored locally on your filesystem. It’s not big (about 400kb) but you need the plone user to have access to its “home”. Running the tests will also fetch more data from nltk bringing the total to about 225Mb, so not for the faint at disk space.

How to use it?

• For a parsed document you can call the term view to display the identified terms (just append @@terms to the url of the content to call the view).

• In order to use the classifier and get suggested tags for some content, you can call @@suggest-categories on the content. This comes down to appending @@suggest-categories to the url in your browser. A form will come up with suggestions, choose the ones that seem appropriate and apply. You will need to have the right to edit the document in order to call the view.

• You can find similar content for some content based on its terms by calling the @@similar-items view.

• For clustering you can just call the @@clusterize view from anywhere. The result is not deterministic but hopefully helpful;). You need manager rights for this so as to not allow your users to DOS your site!

Integration test

Here, we’ll test the classifier using a sample of the Brown corpus. The Brown corpus has a list of POS-tagged english articles which are also conveniently categorized. The test consists of training the classifier using 20 documents from each of the categories ‘news’,’editorial’ and ‘hobbies’. Then we’ll ask the classifier to classify 5 more documents from each category and see what happens.

We can now start adding documents, starting with the first 20 documents in the Brown corpus categorized as ‘news’.

>>> from nltk.corpus import brown
>>> for articleid in brown.fileids(categories='news')[:20]:
...     text = " ".join(brown.words(articleid))
...     id = self.folder.invokeFactory('Document',articleid,
...                                    title=articleid,
...                                    text=text,
...                                    subject='news')

Continuing with 20 documents categorized as ‘editorial’:

>>> for articleid in brown.fileids(categories='editorial')[:20]:
...     text = " ".join(brown.words(articleid))
...     id = self.folder.invokeFactory('Document',articleid,
...                                    title=articleid,
...                                    text=text,
...                                    subject='editorial')

And finally 20 documents categorized as ‘hobbies’:

>>> for articleid in brown.fileids(categories='hobbies')[:20]:
...     text = " ".join(brown.words(articleid))
...     id = self.folder.invokeFactory('Document',articleid,
...                                    title=articleid,
...                                    text=text,
...                                    subject='hobbies')

All these documents should have been parsed and indexed:

>>> catalog = self.folder.portal_catalog
>>> sr = catalog.searchResults(noun_terms='state')
>>> len(sr) > 5
True

Let’s see what terms we get for the first ‘editorial’ content:

>>> browser = self.getBrowser()
>>> browser.open(self.folder.absolute_url()+'/cb01/@@terms')
>>> browser.contents
'...state...year...budget...war...danger...nuclear war...united states...'

Nuclear war and United States? Scary stuff… Time to train the classifier:

>>> from zope.component import getUtility
>>> from collective.classification.interfaces import IContentClassifier
>>> classifier = getUtility(IContentClassifier)
>>> classifier.train()
>>> classifier.tags()
['editorial', 'hobbies', 'news']

For a start, the classifier should be pretty certain when asked about text already classified:

>>> browser.open(self.folder.absolute_url()+'/ca01/@@suggest-categories')
>>> browser.contents
'...news 100.0%...editorial 0.0%...hobbies 0.0%...'

So let’s see where this gets us, by asking the classifier to categorize 5 more documents for which we know the category. We will use the classifier’s functions directly this time instead of adding the documents to plone and calling the @@suggest-categories view. ‘News’ first:

>>> classificationResult = []
>>> for articleid in brown.fileids(categories='news')[20:25]:
...     text = " ".join(brown.words(articleid))
...     id = self.folder.invokeFactory('Document',articleid,
...                                    text=text)
...     uid = self.folder[id].UID()
...     classificationResult.append(classifier.classify(uid))
>>> classificationResult
['news', 'news', 'news', 'news', 'news']

Let’s see how we do with ‘editorials’

>>> classificationResult = []
>>> for articleid in brown.fileids(categories='editorial')[20:25]:
...     text = " ".join(brown.words(articleid))
...     id = self.folder.invokeFactory('Document',articleid,
...                                    text=text)
...     uid = self.folder[id].UID()
...     classificationResult.append(classifier.classify(uid))
>>> classificationResult
['editorial', 'editorial', 'editorial', 'editorial', 'editorial']

That’s excellent! What about ‘hobbies’?

>>> classificationResult = []
>>> for articleid in brown.fileids(categories='hobbies')[20:25]:
...     text = " ".join(brown.words(articleid))
...     id = self.folder.invokeFactory('Document',articleid,
...                                    text=text)
...     uid = self.folder[id].UID()
...     classificationResult.append(classifier.classify(uid))
>>> classificationResult
['hobbies', 'hobbies', 'editorial', 'hobbies', 'hobbies']

Not so bad! Overall: we got 14/15 right…

Let’s now pick again the first editorial item and see which documents are similar to it based on the terms we extracted:

>>> browser.open(self.folder['cb01'].absolute_url()+'/@@similar-items')

The most similar item (a Jaccard index of ~0.2) is cb15:

>>> browser.contents
'...cb15...0.212121212121...'

Let’s see what their common terms are:

>>> cb01terms = catalog.searchResults(
... UID=self.folder['cb01'].UID())[0].noun_terms[:20]
>>> cb15terms = catalog.searchResults(
... UID=self.folder['cb15'].UID())[0].noun_terms[:20]
>>> set(cb01terms).intersection(set(cb15terms))
set(['development', 'state', 'planning', 'year', 'area'])

which is fine, since both documents talk about development and budget planning…

What about stats? We can call the @@stats view to find out…

>>> self.setRoles('Manager')
>>> browser.open(self.folder.absolute_url()+'/@@classification-stats')
>>> browser.contents
'...state...True...editorial:hobbies...5.0...'

which basically tells us, that if the word ‘state’ is present the classifier gives 5 to 1 for the content to be in the ‘editorial’ category rather than the ‘hobbies’ category Changelog =========