الفهرس 
AbstractOnly 14 pages are availabe for public view
 |  | from | 85 |  |  |
| | | from | 85 | | |
Abstract
Currently, a huge number of people are using social media to express their opinions on various subjects, as a result of which a vast amount of unstructured opinionated data has become available. By analyzing this data for opinion mining, we can infer the public’s opinion on several subjects and thus the derived conclusions from it could be used to make informed choices and predictions concerning these subjects. Moreover, this result becomes very important in Arabic because Arabs using social media is more common than any other website online. Therefore, the opinion mining can be used for marketing a particular idea or particular trend such as political trends, social trends, etc. Researches have shown that opinion mining in Arabic is a difficult task because of its rich morphology, and it is leaded to difficulty in pre -processing for this language. It is considered a complex task because of these informalities. This is the main holdup for planning a text-mining system. Although opinion mining gives more pores and befits for Arabs users, Arabic Processing is still a challenging task. According to mining view, ignoring the neutral class is not accurate and some studies were recommending using it in the future works to in use the third class (neutral class) for better results and higher accuracy. This thesis presents proposing a new framework to improve classification methods in opinion mining of Arabic social media network data using neutral training examples in learning based on machine learning techniques. In addition, this thesis has shed light on the importance of pre-processing, in which the text pre-processing is the key factor to the sentiment analysis and classification, especially for highly complicated languages (with rich morphology), such as the Arabic language. When the tweets have various approaches to pre-processing, the result will show dissimilar levels of accuracy and the importance of using neutral training examples to facilitate learning. To prove this, we present a proposed framework for sentiment analysis in Arabic tweet. A data collected from 2000-labeled tweets (1000 positive tweets and 1000 negative ones) on different subjects are used such as: politics and arts, then apply several classifier algorithms for two levels. For the first level (positive, negative), we got the best accuracy 88% using MLP algorithm. For the second level (positive, neutral, negative) when the classification is based on three classes, we got the best accuracy 58% using LR algorithm which shows different levels of accuracy when adding neutral training examples and experimental results showed this indeed.