The Ontopia Automatic Classifier

A Developer's Guide

Published by: Ontopia
Date:2013-08-01
Version:5.3.0

Table of contents

Abstract

This Developer's Guide is an introduction to the automatic classification in Ontopia, and shows how to use it to classify content into a topic map.

1. Introduction

The automatic classification module can be used to process content to either classify it automatically, or give suggestions for classification to a user who can then fine-tune the suggestions into the final classification.

Given a file the module does the following:

The module returns a set of terms, each one of which has a score between 0 and 1 expressing how confident the module is that the content is about this term. The module does stemming of terms found in the content, based on the detected language, so that different variants of a single term, like "widget" and "widgets" are treated as a single term. The API provides mechanisms to access all variants of a term.

It's also possible to provide a topic map to the classifier, and if so the classifier will use information in this topic map to provide better classification results. For example, terms which occur in the topic map will be ranked higher than terms which do not. The information is also used to help the classifier detect compound terms (like "Topic Maps" or "New York").

As an example, below is the output of running the classification on this developer's guide:

1: classifier 1.0, 10
2: Ontopia 0.9398362722937127, 2
3: automatic classification 0.6733045538011962, 2
4: topic map 0.32963578857473336, 7
5: Developer Guide 0.31782661003640583, 2
6: widget 0.31465856836714945, 2
7: fine-tune 0.2391013657364179, 1
8: single term 0.2283465338494093, 2
9: API 0.20243323677418937, 1
10: OOXML 0.19983477366712343, 2
11: XML 0.15688292149123131, 1
12: HTML 0.15535359913968713, 1
13: PDF 0.1549748891676282, 1
14: Powerpoint 0.15236326757637844, 1
15: ppt 0.15162942789649314, 1
16: WordProcessingML 0.15053865122008855, 1
17: docx 0.14981801021049407, 1
18: PresentationML 0.14874667734265895, 1
19: pptx 0.14803876861331663, 1
20: module 0.13035138155842613, 5
21: optionally 0.12971709388058605, 1
22: Microsoft 0.12555221252942206, 2
23: java 0.11704152864641565, 1
24: process content 0.11542729078259414, 2
25: URI 0.1055602118063965, 1
26: command-line 0.10543579259361881, 1
27: classification 0.10537143393837285, 8
28: TermDatabase 0.104641226752405, 1
29: textual 0.10399562569001171, 2
30: suggestions 0.10272639505645716, 3
Total: 186 terms.

The module is still fairly immature, and so it's hard to predict how well the module will perform on any given collection of content. In general, however, the longer the text and the more focused on particular concepts it is, the better the module will perform. It also tends to be confused by mixed-language content, and by embedded code and markup. These last two problems can be handled, but not through the PUBLIC API. For help on this sort of issue, please contact the mailing list.

The following formats are supported:

At the moment, the languages supported are Norwegian and English only, but it is possible to plug in support for more languages. If you are interested in doing this, please contact the mailing list.

2. Running the classifier

The simplest way to run the classifier is to use Chew, the command-line tool. This takes a content file and, optionally, a topic map as input, and provides a textual dump of the classification results (as shown above). The main benefit of this is that it makes it easy to try out the classifier and see how it works on your content.

To try Chew run the command java net.ontopia.topicmaps.classify.Chew and follow the instructions you get.

To really make use of the classifier you have to write your own code to process content and update the topic map accordingly. To do that, use the net.ontopia.topicmaps.classify.SimpleClassifier class. This will let you provide a reference to a file or URI, and will return a TermDatabase with the classification results. From this you can update the topic map as desired, or even provide suggestions to users.

For more details, please consult the javadoc.

We are well aware that this limited API probably provides less functionality than users may need, but are awaiting further user feedback before we provide a more detailed API. Feedback to the mailing list is very much welcome.