help

Deep Learning for Sentiment Analysis Based on Customer Reviews

B. Seetharamulu

1 , B. Naresh Kumar Reddy

1 and K. Bramha Naidu

1

1 Faculty of Science and Technology,

1 ICFAI Foundation for Higher Education, Hyderabad, India.

Abstract - Online reviews became popular as people are taking decisions with the help of them. In this context, the purpose of this project is to develop a deep learning based framework that can be used to classify customer reviews into positive or negative. This process is known as sentiment analysis. It is based on the supervised learning mechanisms where a classifier is built with knowledge of training data and then it is used to classify testing data. A prototype application is built to demonstrate proof of the concept. The success of deep learning highly relies on the availability of large-scale training data. A novel deep learning framework for review sentiment classification which employs prevalently available ratings as weak supervision signals. An algorithm by name Deep Learning based Sentiment Analysis (DL- SA) is proposed and implemented to achieve this. A deep learning framework is proposed and implemented. A prototype application is built to demonstrate proof of the concept. The empirical study revealed that the proposed system is better than the state of the art.

Index Terms - Deep learning, customer reviews, machine learning, sentiment classification

I. INTRODUCTION

Since Web 2.0 and e- commerce booming, yet more people begin shopping online and post comments on merchant / review websites regarding their buying experiences. Such opined details are important resources for both future decision- making customers and retailers to improve their products/goods and services. Moreover, as the amount of reviews is rapidly growing, people face a serious problem of overloading the information. A variety of opinion mining strategies were suggested to mitigate the problem, e.g. opinion overview, comparative studies, and opinion polling. A vital component of such opinion mining approaches is a natural sentence classifier for feelings. Common methods of classification of emotions typically fall into two classifications: (1) Methods based on lexicons, and (2) methods of machine learning. Generally, consider the approach of first constructing a sentiment lexicon of opinion terms (such as e.g. “nice”, “bad"), and then developing category rules based on the words of opinion and previous syntactic information that occur. Notwithstanding productivity, these types of approaches require considerable effort in the creation of lexicons and the design of laws. In addition, lexicon-based methods can't manage implicit opinions well, i.e. objective observations like "I bought the mattress a week ago, and today a valley appeared." Deep learning has appeared in recent years as a powerful means of solving the challenges of classifying

emotions. The deep neural network algorithm implicitly learns to interpret the data at a high level and thus prevents laborious tasks like feature engineering. Second benefit is that the exponentially stronger expressive power of deep models than the shallow models. Yet deep learning success depends strongly on the existence of the data regarding large-scale training. It is still very laborious to build large-scale named training data sets for the detection of sentence level sentiment. The provided contributions of this paper are illustrated below: 1. An algorithm named Deep Learning for Emotion Recognition (DLER) is proposed and implemented. 2. An application which is a prototype is built to illustrate the proof of a concept. 3. The algorithm is evaluated and the results are compared with the state of the art.

The organization of this paper is illustrated as follows. Section II provides literature survey of the review. Section III provides the in detail proposed application. Section IV illustrates the implementation of the application. Section V contains results and section VI contains conclusion of the paper and future work.

II. RELATED WORK

This section provides review of relevant literature. The effectiveness of algorithms regarding machine learning actually depends on the representation of data, and we theorize that it's because different interpretations will interact more or less with and cover behind the data the various understanding factors of variance. While domain specific knowledge may be used to aid design depictions, having to learn with generalized priors could also be used, and the search for AI motivates the creation of more efficient algorithms of representation learning that incorporate these priors [1].Conceptual results indicate that deep architectures may be needed to learn the complex functions that can depict high-level abstractions (such as e.g., vision, language, and other AI level tasks). Deep architectures were generally made up of numerous levels of nonlinear activities, for example in neural networks with several hidden layers or in complex propositional formula that reuse many sub formulas. Looking for parameter space of the deep architectures is a challenging task, yet learning algorithms like Deep Belief Networks has currently been introduced to address this issue with significant results, exceeding the state of the art in some areas [2]. Rather than leveraging carefully

IEEE - 49239

11th ICCCNT 2020 July 1-3, 2020 - IIT - Kharagpur

Kharagpur, India

integrated manmade input functionalities for each assignment, our system gets to know internal depictions based on enormous amounts of training data that is unlabeled[3]. The site includes a multitude of product reviews but it is a challenging task to sift through them. Preferably, an opinion mining tool will process a collection of search results of a given item, produce a collection of product attributes (quality, functionality, etc.) and aggregate opinions (poor, mixed, good) on each of them [4]. One of essential forms of web based information is the views expressed in the content generated by the user, e.g. product reviews, forum posts, and forums. In this article we rely on product reviews from buyers. In general, we are studying the issue of deciding the semantic orientations (positive, negative or neutral) of opinions expressed within comments regarding product features. There are many solutions to this issue, such as opinion mining, summarization and search [5]. The literature found in [6]-[9] revealed that the deep learning mechanisms can still be exploited for opinion mining. As an improvement of the existing techniques, A novel deep learning framework for review sentiment classification is proposed in this paper. Which employs prevalently available ratings as weak supervision signals

III. PROPOSED SYSTEM

Customer reviews, of late, are playing vital role in making well informed decisions. The purpose of this paper is to propose and implement a deep learning based framework for sentiment analysis. Sentiment analysis results in identifying reviews as positive or negative. The solution is based on supervised learning method which needs training data. The aim of this paper is to develop a deep learning based framework that can be used to classify customer reviews into positive or negative. It Describes the Architecture diagram, it contains Web Database, Remote User, Admin web server modules are there. The purpose of remote user is to search the friends and posts and to send the request. The purpose Admin Web Server is to add the products and list all the posts with

ranks. As presented in Figure 1, there are many functionalities in the system associated with different roles. Based on this, different modules are identified. They are as follows. In the admin module, the Admin has to login by using valid user name and password. After login successful he can do some operations such as add categories, add posts, list of all posts, list of all recommended posts, view good reviews, view bad reviews, list of all reviewed posts, list of users, list of all search history, update posts, lists of bad reviews by date wise, list of good reviews by date wise. In the add products to posts module, the admin can add the post by including product name, price, description and corresponding product image.

In the view all posts module, the admin can view the post by searching keyword and can get all the information about the product like product name, price, description and corresponding product image. In the sentiment analytics module, the admin can analyse the sentiment based on

products from positive sentiment words, products from negative sentiment words, products from neutral sentiment words and View Products Rating based on sentiment words.

Fig.1. Overview of the proposed system

In the user module, there are n numbers of users are present. User should register before doing some operations. After registration successful he has to login by using authorized user name and password. Login successful he will do some operations like view user details, search for products posts, view my search history, view recommended, search for top N posts and logout. In the search for good reviews and bad reviews model, user searches for reviews for the post and can get the following information like product name, price, description and corresponding product image. The user can recommend the product and can give review using sentiment words (such as good or bad product like that) based on brand, Quality, Price. A. Proposed Algorithm

Algorithm: Deep Learning based Sentiment Analysis

Inputs: Review dataset D Output: Sentiment classification results R 1. Start 2. Load dataset D 3. Pre-processing to generate training and testing data 4. Learn Convolutional Neural Network (CNN) classifier

using training data 5. Take testing set for classification 6. Update R 7. Prune sentiments 8. Return R 9. Stop

IEEE - 49239

11th ICCCNT 2020 July 1-3, 2020 - IIT - Kharagpur

Kharagpur, India

Algorithm 1: Deep learning based sentiment analysis

A deep learning based algorithm is proposed to achieve the aim of the proposed research. It takes review dataset as input and provides sentiment classification as intended output as shown in the Algorithm 1, the input is Review Dataset D, and the Output is Sentiment Classification Results R. The CNN based neural network as part of deep learning will have a learning process using training data prior to actual prediction activity.

IV. IMPLEMENTATION DETAILS

The word “data” is plural, not singular. In American English, periods and commas are within quotation marks, like “this period.” A parenthetical statement at the end of a sentence is punctuated outside of the closing parenthesis (like this). (A parenthetical sentence is punctuated within the parentheses.) A graph within a graph is an “inset,” not an “insert.” The word alternatively is preferred to the word “alternately” (unless you mean something that alternates). Do not use the word “essentially” to mean “approximately” or “effectively.” Be aware of the different meanings of the homophones “affect” and “effect,” “complement” and “compliment,” “discreet” and “discrete,” “principal” and “principle.” Do not confuse “imply” and “infer.” The prefix “non” is not a word; it should be joined to the word it modifies, usually without a hyphen. There is no period after the “et” in the Latin abbreviation “et al.” The abbreviation “i.e.” means “that is,” and the abbreviation “e.g.” means “for example.” An excellent style manual for science writers is [10]- [14] . As shown in the figure 2 it provides the interface for User authentication for user main. This will help in providing interface for various activities of the user. Thus the logged in user can operate with the system as per the privileges [15] –[19].

Fig. 2. User Page

Fig.3. Product Details

As shown in Figure 3, it describes the product details like Product name, price, brand, quality, data, product image, date. This product is recommendations from friends for user main. It provides results of deep learning based outcomes that are useful to users.

Fig.4. Product based on sentiment words

As shown in the Figure 4 it describes the product rating based on sentiment words like product name, positive comments, negative comments, neutral comments and rating. Product is useful in generating recommendations that help users.

IEEE - 49239

11th ICCCNT 2020 July 1-3, 2020 - IIT - Kharagpur

Kharagpur, India

Fig.5. Product based on Negative Sentiment words

As shown in the Figure 5, it describes the product based on negative comments or negative sentiment words for web server. The proposed deep learning based solution provides the sentiment analysis results. The results provide different kinds of reviews that can help users to have useful information.

V. EXPERIMENTAL RESULTS

This section provides experimental results. It evaluates the performance of proposed system and compares the same with the existing system.

% of Training Data

Accuracy (%)

WDE CNN-rand Proposed

20 0.7 0.72 0.75

40 0.8 0.77 0.81

60 0.82 0.82 0.87

80 0.84 0.87 0.93

100 0.86 0.92 0.99

Table 1: Results of accuracy against different % of training

data

As presented in Table 1, it provides the experimental results of proposed and existing system in terms of accuracy against different % of training data.

As shown in Figure 6, it describes the experimental results of proposed and percentage on multiple values. The results revealed that the training data affects accuracy of the classifiers. The proposed system showed better performance

over the existing.

Fig.6. Accuracy against different % of training data

Margin Parameter

Accuracy (%)

WDE CNN-rand Proposed 0 0.873 0.848 0.9 5 0.88 0.848 0.91 10 0.873 0.848 0.89 15 0.87 0.848 0.88 20 0.845 0.848 0.87 25 0.847 0.848 0.86 30 0.847 0.848 0.85

Table 2: Accuracy against margin parameter

As shown in Table 2, it provides accuracy of different methods against the given margin parameter.

Fig.7. Results of Proposed and Existing system on multiple values.

As shown in Figure 7, it describes the experimental results of proposed and existing system on multiple values. The margin parameter value is taken in horizontal axis and the accuracy of the methods is shown in vertical axis. The results revealed that the accuracy of the proposed system is better than that of existing.

IEEE - 49239

11th ICCCNT 2020 July 1-3, 2020 - IIT - Kharagpur

Kharagpur, India

V. CONCLUSION AND FUTURE WORK

Sentiment analysis is one of the key challenges for mining online user generated content. In this work, we focus on customer reviews which are an important form of opinionated content. The goal is to identify each sentence’s semantic orientation. The success of deep learning highly relies on the availability of large-scale training data. Customer reviews, of late, are playing vital role in making well informed decisions. The purpose of this project is to propose and implement a deep learning based framework for sentiment analysis. Sentiment analysis results in identifying reviews as positive or negative. An algorithm by name an algorithm by name Deep Learning based Sentiment Analysis (DL-SA) is proposed and implemented to achieve this. A prototype application is built to demonstrate proof of the concept. The empirical study revealed that the proposed system is better than the state of the art. The solution is based on supervised learning method which needs training data. Experiments on reviews collected from Amazon.com show that WDE is effective and outperforms baseline methods. For future work, we will investigate applying WDE on other types of deep networks and other problems involving weak labels.

REFERENCES

[1] Bengio et al., YoshuaBengio, Aaron Courville, and Pierre Vincent. Representation learning: A review and new perspectives. IEEE TPAMI, 35(8):1798–1828, 2013.

[2] YoshuaBengio. Learning deep architectures for ai. Foundations and trends_R in Machine Learning, 2(1):1–127,2009.

[3] ChristopherMBishop. Pattern recognition and machine learning. springer, 2006.

[4] Ronan Collobert, Jason Weston, L´eonBottou, Michael Karlen, KorayKavukcuoglu, and Pavel Kuksa. Natural language processing (almost) from scratch. JMLR, 12:2493–2537, 2011.

[5] Kushal Dave, Steve Lawrence, and David M Pennock. Mining the peanut gallery: Opinion extraction and semantic classification of product reviews. In WWW, pages 519–528, 2003.

[6] Xiaowen Ding, Bing Liu, and Philip S Yu. A holistic lexicon-based approach to opinion mining. In WSDM,pages 231–240, 2008.

[7] John Duchi, EladHazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. JMLR, 12:2121–2159, 2011.

[8] Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-RuiWang, and Chih-Jen Lin. Liblinear: A library for large linear classification. JMLR, 9:1871–1874, 2008.

[9] Ronen Feldman. Techniques and applications for sentiment analysis. Communications of the ACM, 56(4):82–89, 2013.

[10] Xavier Glorot, Antoine Bordes, and YoshuaBengio. Domain adaptation for large-scale sentiment classification: A deep learning approach. In ICML, pages 513–520, 2011.

[11] Minqing Hu and Bing Liu. Mining and summarizing customer reviews. In SIGKDD, pages 168–177, 2004.

[12] NalKalchbrenner, Edward Grefenstette, and Phil Blunsom. A convolutional neural network for modelling sentences. In ACL, 2014.

[13] Yoon Kim. Convolutional neural networks for sentence classification. In EMNLP, pages 1746–1751, 2014.

[14] Bing Liu, Minqing Hu, and Junsheng Cheng. Opinion observer: analyzing and comparing opinions on the web. In WWW, pages 342–351, 2005.

[15] Bing Liu. Sentiment analysis and opinion mining. Morgan & Claypool Publishers, 2012.

[16] Andrew L Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher Potts. Learning word vectors for sentiment analysis. In ACL, pages 142–150, 2011.

[17] Min-Yuh Day ; Yue-Da Lin, “Deep Learning for Sentiment Analysis on Google Play Consumer Review,” IEEE International Conference on Information Reuse and Integration, 2017.

[18] Alhassan Mabrouk ; Rebeca P. Díaz Redondo ; Mohammed Kayed, “Deep Learning-Based Sentiment Classification: A Comparative Survey,” IEEE Access, Vol 8, 2020.

[19] Seungwan Seo ; Czangyeob Kim ; Haedong Kim ; Kyounghyun Mo ; Pilsung Kang, “Comparative Study of Deep Learning- Based Sentiment Classification,” IEEE Access, Vol 8, 2020.

IEEE - 49239

11th ICCCNT 2020 July 1-3, 2020 - IIT - Kharagpur

Kharagpur, India