OPEN DOMAIN TARGETED SENTIMENT POLARITY DETECTION FOR MICRO-BLOGS IN SOCIAL MEDIA. PhD thesis, King Fahd University of Petroleum and Minerals.

Preview

PDF (PhD Dissertation: Shadi Abudalfa) Shadi_Abudalfa_PhD_Dissertation.pdf Download (5MB) | Preview

Arabic Abstract

يتوفر هذه الأيام كم هائل من الموضوعات والآراء المتاحة على شبكة الإنترنت من خلال استخدام وسائل التواصل الاجتماعي، مما يدلل على أن هذه الآراء تلعب دورا هاما في حياتنا، وتؤثر على سلوك المجتمعات والحكومات. وفرة هذه البيانات على وسائل التواصل الاجتماعي يفتح الباب أمام العلماء لتطوير النظم الآلية المناسبة لتحليل المشاعر والآراء. تم تطوير العديد من الأدوات لكشف الآراء آليا عبر النصوص القصيرة ولكنها لا تزال محدودة الكفاءة. في هذا البحث، قدمنا العديد من التقنيات الجديدة والمناسبة لتحليل المشاعر والآراء على وسائل التواصل الاجتماعي. واقترحنا نظام جديد لتحليل النصوص القصيرة وإستكشاف سياق الكلام بالكشف عن الموضوعات المطروحة واستخلاص قطبية الرأي تجاهها بخصوص التصنيفات الإيجابية والسلبية والمحايدة. يتركز بحثنا على الكشف عن أي موضوع مطروح في النص ومن ثم تحديد قطبية الآراء تجاهه، والذي يشار إليه بالنطاق المفتوح لتحليل المشاعر للموضوع المستهدف. قدمنا آليات جديدة لتحليل المشاعر في النطاق المفتوح للمواضيع المستهدفة وتصنيف الآراء استنادا إلى تقنيات تعليم الآلة بالإعتماد على عدد محدود من البيانات المؤشرة. ويشمل عملنا البحثي كذلك إقتراح طرق جديدة تزيد الدقة في إستكشاف قطبية الآراء. وأجرينا كذلك العديد من التجارب والتحليلات لتوضيح قوة الآليات المقترحة بالمقارنة مع الدراسات السابقة التي تم أنجازها في هذا المجال.

English Abstract

Tremendous amount of opinions, regarding almost every topic, is available on the internet these days through social media. Evidences show that such opinions play important role in our life and affect behavior of individuals, communities, industries, and governments. Availability of such wealth of opinions in the social media motivated researchers to develop automated systems for opinion mining, also known as sentiment analysis. A sentiment represents a polarity of opinion that is typically expressed as one of three classes: positive, negative, or neutral. Many tools are currently available for sentiment mining in short text, referred to as micro-blogs, for different languages but their efficacies are still limited. Such limitations include and not limited to dealing with specific domains and providing limited performance. In this work, we developed an approach for topic identification and polarity classification of opinions offered in the form of micro-blogs. We propose a new context-based analysis system for detecting targets among a set of micro-blogs and detecting sentiment polarities towards categorized topics that describe the targets. Our literature review revealed that the research direction has been originally focusing on classifying sentiments polarities towards specific targets, i.e., topics, in the micro-blogs. A more recent direction currently addresses the problem of detecting the target and identifying the sentiment polarity toward it. While, the former direction is referred to as target-dependent sentiment classification, the latter one is referred to as open domain targeted sentiment classification. Our literature review also revealed that majority of the state of the art approaches use supervised learning techniques for both target-dependent and open domain targeted sentiment classification. Such techniques need a huge amount of labeled data for increasing classification accuracy. However, preparing labeled data from social media needs a significant effort and may cause inaccurate results if some micro-blogs are annotated incorrectly. For that matter, we propose new techniques to employ semi-supervised learning methods for improving the performance of both target-dependent and open domain targeted sentiment classification by using partially labeled data. Additionally, we propose new supervised techniques for improving the performance of both target-dependent and open domain targeted sentiment classification. Numerous experiments are conducted to show that our proposed techniques outperform prominent ones available in the literature. A comparison framework and statistical analysis are included in this work as well to validate experiment results.

Item Type:	Thesis (PhD)
Subjects:	Computer
Department:	College of Computing and Mathematics > Information and Computer Science
Committee Advisor:	Ahmed, Dr. Moataz
Committee Members:	Alshayeb, Dr. Mohammad and El-Bassuny, Dr. Tarek and El-Maleh, Prof. Aiman and Sait, Prof. Sadiq
Depositing User:	SHADI ABUDALFA (g201305390)
Date Deposited:	08 Aug 2018 05:04
Last Modified:	31 Dec 2020 05:54
URI:	http://eprints.kfupm.edu.sa/id/eprint/140788