FEATURES DEVELOPMENT AND FEATURE EXTRACTION METHODS FOR IDENTIFICATION OF SCIENTIFIC RELATION SCHEMES BASED ON RHETORICAL CITATION

• Main Author: Sibaroni, Yuliant
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49322
• Institution: Institut Teknologi Bandung

This dissertation research discusses the identification of scientific papers relations based on rhetorical citation obtained by analyzing the citation context contained in a citation sentence. This approach is known as a citation context-based approach, where this approach is more detailed compared to the previous two approaches namely content-based and citation analysis-based approaches. The latter two relations approaches can only be used to identify similarities relation between papers. At present, the schema of scientific paper relations developed based on citation context is only explicitly carried out by Wang et al, where the relations produced are extend, criticize, and compare relations. The main feature used by Wang to identify this paper relation is quite simple, namely the cue phrase feature. The focus of this dissertation research is to develop a feature extraction method and produced a feature set of paper relations that can identify Wang's paper relations better. The identification of paper relations is done by classifying each sentence using a supervised machine learning approach. The feature development process is carried out in stages, starting from the extend relation, the critique relation, and finally the compare relation. The results showed that each type of paper relations has special and different features. In the extend relation, several important features were obtained, namely the phrase combination feature and the n-gram feature with top-N correlation. In criticize relations, there are 5 groups of important features, namely the adaptation feature of extended relations, the combination of cue phrases with citation, the combination of cue phrases with previous citation, the combination of cue phrases, and conjunction of some basic features. In the compare relation, there are three important groups of features produced, namely proportionWord feature, probabilityWord feature, and cuephraseWord feature. The feature development process is done by observing the patterns that appear in each relation sentence. Although compared to the baseline, the proposed feature has better performance, but there are still some problems such as the high false prediction values that still appear, the missing of the citation context sentence (co-reference), and so forth.. The increase of F-Measure obtained ranges from 15-40% compared to the baseline feature