Public Opinion Analysis Of Amharic News Using Deep Learning

By Author: GEREMEW ASRAT
Total Articles: 1
Comment this article

Background of the Study
Sentiment analysis is a method of computing and satisfying a view of a person given in a piece of a text, to identify persons thinking about any topic is positive or negative. Sentiment analysis of Facebook data is providing an effective way to expose user opinion which is necessary for decision making in various fields [1]. Sentiment analysis is a fundamental branch in natural language processing (NLP) [2]. It is the process of understanding sentiment in user-generated opinionated data in social media, product feedback or blogs. Sentiment analysis or opinion mining aims at determining the attitude of a speaker, writer or another subject with respect to a certain topic or event. [3].Sentiment analysis or opinion mining is the computational study of people’s opinions, sentiments, emotions, appraisals, and attitudes towards entities such as products, services, organizations, individuals, issues, events, topics, and their attributes[4].

Recently, deep learning has shown remarkable improvements in the sentiment analysis field in the English language[3]. However, a research has been done on ...
... using deep learning approach in the Amharic sentiment analysis[9]. Amharic is widely spoken language in Ethiopia. Amharic has own written system with a version of the Ge'ez script known as Fidel. It is the second most-spoken Semitic language in the world, after Arabic[10], and the official working language of the Federal Democratic Republic of Ethiopia.
Social media has given web users a venue for expressing and sharing their thoughts on different events[1], Face book is one of them and used as a famous social media platform through which users can express their opinions on various events or objects.
The social media site Facebook is our targeted website for this study. This is because the AMC Facebook page has many members or followers and vast user-generated data is available. The objective of this work to use word2vec technique to automatically classify Amharic as positive, negative or neutral using deep learning classifiers, LSTM and GRU.

Statement of the Problem
Amhara Media Corporation (AMC) is one of the official media services in Ethiopia, and it covers social, economic, cultural and political subjects. AMC is distributed their programs in the way of different mediums Television, radio, magazine, and social media sites Facebook, Twitter, YouTube, Telegram with local and international languages.
In recent years[6], the exponential increase in the Internet usage and exchange of public opinion is the driving force behind opinion mining today. AMC is one of Amhara regional state governmental media organization that deliver relevant information for the population among the organization’s news service is common. Currently the corporation receives comment from customers for their services are through social media sites Facebook, Twitter, YouTube, and Telegram. And from social sitesby reading that much comments which is difficult to addressall customers’ opinions. The analysis of this data to extract latent public opinion and sentiment is a challenging task.

Deep learning has shown great success in the field of sentiment analysis and is considered as the state-of-the-art model in various languages[3]. The Amharic language imposes many challenges, due to its complex structure, various dialects, in addition to the lack of its resources.

We have been proposed corpus-based sentiment analysis for Amharic language using deep learning approach
This research will answerthe following questions: -
 Which text representation technique applied for Amharic comments is used?
 Which deep neural networks classifiersfor Amharic commentsare used?
 What tools appropriate for extracting Amharic comments on social media are used?

Methodologies
Data Source and Data Collected Technique
In order to analyses the opinions of the users to collect the user-generated content from AMC used web scraping. Web scraping is the process of automatically mining data or collecting information from the World Wide Web. Such type of web scraping method is used to collect data from social media, Facebook Graphic API tool, Face-pageris used to collect public posts from Facebook page [21]. This study the primary data source was conducted from Amhara Media Corporation (AMC)Amharic Facebook page, because of it is legal under the Facebook company terms and condition.We focused on special attention to socio-politics domain. The reason behind choosing these domains are the availability of user generated content in Amharic language pretty good domain.

Text Preprocessing and Representation Techniques
Text Preprocessing:Texts generated by humans in social media sites contain lots of noise that can significantly affect the results of the sentiment classification process.The Preprocessing is increasing the data quality to some extent and also needed to transform the data raw data into a coherent format. In this study, the Preprocessing techniques data cleaning:removing non-Amharic text and symbols, numbers and punctuation, stop-words removal, Tokenzation and normalization are used in the collected datasets.

Deep Learning Approach
Sentiment analysis algorithms fall into one of three loads: Rule-based, Automatic or machine learning, and hybrid approaches.This study has been used automatic approach. Automatic methods, systems rely on machine learning techniques to learn from data. In classification algorithms or models have been selected neural network algorithm. Deep learning approaches are part of machine learning which refers to deep neural Networks. Deep neural networks consist of many networks[16], such as CNN (Convolutional Neural Network), DBN (Deep Belief NeuralNetwork), RNN (Recurrent Neural Network), Recursive Neural Network, Long Short term Memory (LSTM), Bidirectional Long Short-Term Memory (BI-LSTM), RNDM (Recursive Neural Deep Model), and RNTN (Recursive Neural Tensor Network). We usedLSTM and GRU models for training and testing the dataset.

Design Science Approach
Sentiment classification model of Amharic text, we used deep neural network with word embedding technique. To realize this, data preparation like separation of data into training and test sets that is used training-test-split method, loading the data, and cleaning the data to remove punctuation and numbers used preprocessing techniques, and defining a vocabulary of preferred words are play vital role for this work. Then, we train the model using the Keras deep learning library with LSTM and GUR classifiers as it confirmed to be successful at classification problems.

Tools and Techniques
In this study have been used tools that are Tensorflow deep learning library, Keras deep learning library, Scikit learn machine learning library and python with Anaconda navigator. And also, we used web scrapping methods; the methods are export comment site and Facepager tool to extract users’ comments from AMC public Facebook page and to represent the text we used word embedding technique.

Related Work
In this session briefly discussed of the sentimental analysis related works for some languages. This paper has been seen the three mostly employed approaches for sentiment analysis i.e., Rule-Based, Machine Learning and Deep learning are reviewed.
As summarized,the previous researchers the majority works done on Local language and International Languages, and as well as various domain dataset and different text granularity. Almost their works are using Machine learning and Deep learning approaches. In Amharic Language using deep learning approach is work done very rare. Hence our study is focused on Amharic Language sentence level sentiment classification using deep leaning approach, in order to improve the Amharic sentiment analysis accuracy.

Proposed Approach
The users’ comments were collected from Amhara Media Corporation official Facebook page posted Amharic news is considered for our analysis. Our proposed approach for Sentiments Analysis for Amharic text using deep learning approach is presented in the following section. Our work divides into four main stages. Those are: firstly, we collect unstructured users’ comments need to be used to feed our network. Secondly, we used the preprocessing method since the users’ review data are unstructured. This data to be preprocessed using various preprocessing techniques like data cleaning, URL, HTML tag removal, punctuation and number removal, and stop word removal and so on. Thirdly,the preprocessed texts are represented using word embedding representation model, word embedding to obtain the feature vectors. Finally, configure the LSTM and GRU (Gated Recurrent Units)model to train and classify the text sentences as positives, negatives or neural.