Search

There are two problems in using words to represent document contents and query in information retrieval: ambiguity and different words which represent the same concept. These problems can be addressed by using query expansion. We focused on analysing the implementation of query expansion, word sense disambiguation (WSD), iterated relevance feedback, and some retrieval variations to retrieval performance. In this paper, WSD is implemented in Lucene using query expansion with thesaurus and relevance feedback. Extended Lesk algorithm was re-implemented to disambiguate the query using WordNet. Expansion terms were limited up to 20 words chosen from expansion term candidates from disambiguated query's senses information, co-occurrence terms, and most frequent terms using Kullback-Leibler Distance. We iterated the process to find the best number of expansion iteration. We found that the method using WSD to query can extend search process time to 161 times longer at worst. Query expansion using disambiguated sense information did not affect the performance much while using information from relevance feedback did. This experiment provides better understanding of WSD in information retrieval system performance.

Time constraints often lead a reader of scientific paper to read only the title and abstract of the paper, but reading these parts is often ineffective. This study aims to extract information automatically in order to help the readers get structured information from a scientific paper. The information extraction is done by rhetorical classification of each sentence in a scientific paper. Rhetoric information is the intention to be conveyed to the reader by the author of the paper. This research used corpus-based approach to build rhetorical classifier. Since there was a lack of rethorical corpus, we constructed our own corpus, which is a collection of sentences that have been labeled with rhetorical information. Each sentence represented as a vector of content, location, citation, and meta-discourses features. This collection of feature vectors is used to build rhetorical classifiers by using machine learning techniques. Experiments were conducted to select the best learning techniques for rhetorical classifier. Training set consists of 7239 labeled sentences, and the testing set consists of 3638 labeled sentences. We used WEKA (Waikato Environment for Knowledge Analysis) and LibSVM libraries. Learning techniques being considered were Naive Bayes, C4.5, Logistic, Multi-Layer Perceptron, PART, Instance-based Learning, and Support Vector Machines (SVM). The best performers are the SVM and Logistic classifier with accuracy of 0.51. By applying one-against-all strategy, the SVM accuracy can be improved to 0.60.

Content-targeted advertising is an ads placement technique which associates ads to a web page relative to (based on) the content of the web page (web page content). It introduces a challenge about how to settle the conflict of interests by selecting advertisements that are relevant to the users but also profitable to the advertisers and the publishers. This paper proposes an approach to associate ads with web pages using Genetic Programming (GP). GP is an extension of genetic algorithm in which the individual is not a stream of character but rather a program (function).