Improving Retrieval Accuracy in Main Content Extraction from HTML Web Documents
Hadi Mohammadzadeh · Qucosa (Saxon State and University Library Dresden) · 2013
The rapid growth of text based information on the World Wide Web and various applications making use of this data motivates the need for efficient and effective methods to identify and separate the “main content” from the additional content items, such as navigation menus, advertisements, design elements or legal disclaimers. Firstly, in this thesis, we study, develop, and evaluate R2L, DANA, DANAg, and AdDANAg, a family of novel algorithms for extracting the main content of web documents. The main concept behind R2L, which also provided the initial idea and motivation for the other three algorithms, is to use well particularities of Right-to-Left languages for obtaining the main content of web pages. As the English character set and the Right-to-Left character set are encoded in different intervals of the Unicode character set, we can efficiently distinguish the Right-to-Left characters from the English ones in an HTML file. This enables the R2L approach to recognize areas of the HTML file with a high density of Right-to-Left characters and a low density of characters from the English character set. Having recognized these areas, R2L can successfully separate only the Right-to-Left characters. The first extension of the R2L, DANA, improves effectiveness of the baseline algorithm by employing an HTML parser in a post processing phase of R2L for extracting the main content from areas with a high density of Right-to-Left characters. DANAg is the second extension of the R2L and generalizes the idea of R2L to render it language independent. AdDANAg, the third extension of R2L, integrates a new preprocessing step to normalize the hyperlink tags. The presented approaches are analyzed under the aspects of efficiency and effectiveness. We compare them to several established main content extraction algorithms and show that we extend the state-of-the-art in terms of both, efficiency and effectiveness. Secondly, automatically extracting the headline of web articles has many applications. We develop and evaluate a content-based and language-independent approach, TitleFinder, for unsupervised extraction of the headline of web articles. The proposed method achieves high performance in terms of effectiveness and efficiency and outperforms approaches operating on structural and visual features.