Thesis
Enhancing extremist data classification through textual analysis
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2023
- Thesis identifier
- T16775
- Person Identifier (Local)
- 201686484
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- The high volume of extremist materials on the Internet has created the need for intelligence gathering via the Web and real-time monitoring of potential websites for evidence of extremist activities. However, the manual classification for such contents is practically difficult and time-consuming. In response to this challenge, the work reported here developed several classification frameworks. Each framework provides a basis of text representation before being fed into machine learning algorithm. The basis of text representation are Sentiment-rule, Posit-textual analysis with word-level features, and an extension of Posit analysis, known as Extended-Posit, which adopts character-level as well as word-level data. Identifying some gaps in the aforementioned techniques created avenues for further improvements, most especially in handling larger datasets with better classification accuracy. Consequently, a novel basis of text representation known as the Composite-based method was developed. This is a computational framework that explores the combination of both sentiment and syntactic features of textual contents of a Web page. Subsequently, these techniques are applied on a dataset that had been subjected to a manual classification process, thereafter fed into machine learning algorithm. This is to generate a measure of how well each page can be classified into their appropriate classes. The classifiers considered are both Neural Network (RNN and MLP) and Machine Learning classifiers (such as J48, Random Forest and KNN). In addition, features selection and model optimisation were evaluated to know the cost when creating machine learning model. However, considering all the result obtained from each of the framework, the results indicated that composite features are preferable to solely syntactic or sentiment features which offer improved classification accuracy when used with machine learning algorithms. Furthermore, the extension of Posit analysis to include both word and character-level data out-performed word-level feature alone when applied on the assembled textual data. Moreover, Random Forest classifier outperformed other classifiers explored. Taking cost into account, feature selection improves classification accuracy and save time better than hyperparameter turning (model optimisation).
- Advisor / supervisor
- Weir, George
- Resource Type
- DOI
- Funder
Relazioni
Articoli
| Thumbnail | Titolo | Data caricata | Visibilità | Azioni |
|---|---|---|---|---|
|
|
PDF of thesis T16775 | 2023-11-15 | Pubblico | Scaricare |