Date of Award
Doctor of Philosophy (PhD)
The amount of text data produced in science, finance, social media, and medicine is growing at an unprecedented pace. The raw text data typically introduces major computational and analytical obstacles (e.g., extremely high dimensionality) to data mining and machine learning algorithms. Besides, the growth in the size of text data makes the search process more difficult for information retrieval systems, making retrieving relevant results to match the users’ search queries challenging. Moreover, the availability of text data in different languages creates the need to develop new methods to analyze multilingual topics to help policymakers in governmental and health systems to make risk decisions and to create policies to respond to public health crises, natural disasters, and political or social movements. The goal of this thesis is to develop new methods that handle computational and analytical problems for complex high-dimensional text data, develop a new query expansion approach to enhance the performance of information retrieval systems, and to present new techniques for analyzing multilingual topics using a translation service.
First, in the field of dimensionality reduction, we develop a new method for detecting and eliminating domain-based words. In this method, we use three different datasets and five classifiers for testing and evaluating the performance of our new approach before and after eliminating domain-based words. We compare the performance of our approach with other feature selection methods. We find that the new approach improves the performance of the binary classifier and reduces the dimensionality of the feature space by 90%. Also, our approach reduces the execution time of the classifier and outperforms one of the feature selection methods.
Second, in the field of information retrieval, we design and implement a method that integrates words from a current stream with external data sources in order to predict the occurrence of relevant words that have not yet appeared in the primary source. This algorithm enables the construction of new queries that effectively capture emergent events that a user may not have anticipated when initiating the data collection stream. The added value of using the external data sources appears when we have a stream of data and we want to predict something that has not yet happened instead of using only the stream that is limited to the available information at a specific time. We compare the performance of our approach with two alternative approaches. The first approach (static) expands user queries with words extracted from a probabilistic topic model of the stream. The second approach (emergent) reinforces user queries with emergent words extracted from the stream. We find that our method outperforms alternative approaches, exhibiting particularly good results in identifying future emergent topics.
Third, in the field of the multilingual text, we present a strategy to analyze the similarity between multilingual topics in English and Arabic tweets surrounding the 2020 COVID-19 pandemic. We make a descriptive comparison between topics in Arabic and English tweets about COVID-19 using tweets collected in the same way and filtered using the same keywords. We analyze Twitter’s discussion to understand the evolution of topics over time and reveal topic similarity among tweets across the datasets. We use probabilistic topic modeling to identify and extract the key topics of Twitter’s discussion in Arabic and English tweets. We use two methods to analyze the similarity between multilingual topics. The first method (full-text topic modeling approach) translates all text to English and then runs topic modeling to find similar topics. The second method (term-based topic modeling approach) runs topic modeling on the text before translation then translates the top keywords in each topic to find similar topics. We find similar topics related to COVID-19 pandemic covered in English and Arabic tweets for certain time intervals. Results indicate that the term-based topic modeling approach can reduce the cost compared to the full-text topic modeling approach and still have comparable results in finding similar topics. The computational time to translate the terms is significantly lower than the translation of the full text.
Al Shanik, Farah Mahmoud, "Enhancing the Performance of Text Mining" (2021). All Dissertations. 2942.