Abstract
Text classification is the categorization of text into a predefined set of categories. Text classification is becoming increasingly important given the large volume of text stored electronically e.g. email, digital libraries and the World Wide Web (WWW). These documents represent a massive amount of information that can be accessed easily. To gain benefit from using this information requires organisation. One way of organising it automatically is to use text classification. A number of well known machine learning techniques have been used in text classification including Naïve Bayes, Support Vector Machines and Decision Trees, and the less commonly used are k-Nearest Neighbour, Neural Networks and Genetic Algorithms. 数据挖掘研究院
数据挖掘工具
One aspect of text classification is general message classification, the ability to correctly classify text messages containing text of different lengths. There are many applications that would benefit from this. An example of such applications are, personal emailing filtering, filtering email into different categories of business and personal email and spam email and email routing, e.g. routing email for a helpdesk, so that the email reaches the correct person.
数据挖掘工具
This thesis presents an investigation of applying a Case based Reasoning (CBR) approach to general text message classification. Case-based Reasoning was chosen as it was found to perform well for a particular type of message classification, spam filtering. CBR was found to have certain advantages over other machine learning techniques such as Naïve Bayes. It was able to handle the dynamic nature of spam better than other machine learning techniques and offered the ability for the training data to be easily updated continuously and to have new training data immediately available. 数据挖掘工具
The objective of this research is to extend previous work conducted on spam filtering to general message classification, which includes classifying short and long text messages into multiple categories. Short text message classification presents a particular challenge as the concept being learnt is weak. We investigated two types of similarity metrics used with CBR, feature based and featureless similarity metrics. We then compared CBR using both feature based and featureless similarity metrics with two well known machine learning techniques. Naïve Bayes (NB) and Support Vector machine (SVM). These two machine learning techniques serve as base line classifiers as they seem to be currently the classifier of choice in the text classification domain. The results of this search show that CBR using a featureless similarity metric achieves better performance than CBR using a feature base similarity metric. The results also show that when using CBR with a feature based similarity metric the classification task required different feature types and different feature representations, depending on the domain.
数据挖掘工具
We also investigated whether a case-base editing technique developed for spam case-bases improve the performance over unedited case-bases on different text domains. We found that the case-base editing technique used for spam filtering performs well for email based case-bases but not for other text domains of either short or long text messages.
