Sentiment Analysis Using Maximum Entropy on Application Reviews (Study Case: Shopee on Google Play)

ABSTRACT


INTRODUCTION
Shopee is top five shopping application and most popular in Indonesia.Most Shopee application users who provide reviews in the form of praise or criticism.A review is critical to know the feedback from the product [1].Currently there are a lot of Shopee application reviews in Indonesia on Google Play.It makes difficult for the company to obtain information as a whole because it takes a long time to read the reviews that enter Google Play one by one.Sentiment analysis can be used to resolve this problem [2].
This research aims to analyze the sentiment of Shopee application reviews on Google Play.The analysis can classify whether a review is positive or negative [3].By using sentiment analysis the company can find out positive and negative reviews from users discuss the application.Then the result of both sentiments will be visualized in the word cloud.Next look for word associations based on the workload.From the word associations the company will get information to what things discussed by user based on positive and negative sentiment.When the company knows what things discussed in positive sentiment, the company can also improve the quality so that users satisfaction fulfilled.Meanwhile, when the company knows the information related to negative sentiment then the company can quickly take action to mitigate it.Sentiment analysis is the process of understanding, extracting and processing textual data automatically to get information [4].
Research of sentiment analysis using Chinese reviews took a long time in pre-processing data [5].While in this study using reviews of Shopee applications that use Indonesian, so that pre-processing does not take a long time because there is already a package to help to pre-process the Indonesian reviews.Maximum entropy is very useful in several natural language processing applications [6].So, our method to be used for Vol. 5, No. 1, Juni 2019, pp. 44-49 Sentiment Analysis Using Maximum Entropy on Application Reviews... (Ulinnuha Rhohmawati, et al) sentiment analysis is maxent.Distribution search which results in maximum entropy value aims to get the best probability distribution that is closest to reality [7], [8].

RESEARCH METHOD
Shopee application reviews on Google Play from 16 th February to 8 th March 2019 used in this research.Data retrieval using scrapping techniques with extensions from Google Chrome, namely Data Miner [9].After the data is collected then pre-processing or cleaning data [10].After that, do manual labelling and divide the data into training and testing [11].Then using maximum entropy method to sentiment analysis the followings are [12] : a. Identify specific words in the document (sentence).b.Form a matrix that contains the value of the occurrence of these specific words with the following index: c. Make a maximum entropy model with training data that j a is counting for each class with a generalized iterative scaling procedure The next steps calculate the accuracy from the classification using the confussion matrix.In the classification evaluation process, that are four possibilities to occur from the process of classifying a row of data [13].If the positive data and positive predictions will be counting as true positive, but if the data is predicted to be negative then it will be counted as false negative.If the data is negative and negative predictions will be counted as true negative, but if the data is positive predicted it would be counted as a false positive [14], [15].The confussion matrix shown in Table 1.The final steps are visualization and word association the positive and negative reviews.The research step can be simplified using a flowchart in Fig. 1.

RESULTS AND DISCUSSION
Review data on Shopee application users on Google Play from 16 th February to 8 th March 2019 using scrapping technique with Data Miner obtained 1305 reviews.Then, do pre-process that starts with translation, spelling normalization, case-folding, tokenizing, and filtering.The example of pre-processing shows in Fig. 2.

Fig. 2. Pre-processing result of Indonesian reviews
Fig. 2. it is known that input data Shopee application review data from Google Play as input data, then after pre processing so that results such as output data.Then perform manual labelling into positive or Before Pre-processing Very fast delivery of goods and online shopping chat response Shopee is very good.Order products according to choice, cheaper prices and faster delivery.
After Pre-processing fast shipping goods online shopping response chat very good order products according to low prices fast delivery Vol. 5, No. 1, Juni 2019, pp. 44-49 Sentiment Analysis Using Maximum Entropy on Application Reviews... (Ulinnuha Rhohmawati, et al) negative.Review data of Shopee application users who entered the positive sentiment class amounted to 891, while those the negative sentiment class amounted to 414.

Accuracy testing of maximum entropy method
The accuracy test of the maximum entropy method using a comparison of training and testing data at 80%: 20% [15].The training data in this research are 1044 reviews, while for the testing data amount to 261 reviews.The test parameters used is accuracy.The test parameters can be calculated using the confussion matrix.The result of the confussion matrix in this research shown in Table 2. Table 2, it is known that the total data is 261, where there are 254 review data classified into the appropriate sentiment class, and there are 7 review data that have not classified into the appropriate sentiment class.The accuracy has achived of 97.32% using (11).

Visualization and associations of positive reviews
Extraction of information on positive reviews repeated.So, that information about the positive reviews of Shopee application users is most often reviewed.A positive review can identify based on word frequency of words in the review.The visualization of information extraction result obtained from positive reviews of users as follows: Fig. 3.Most frequent word in positive Fig. 3. shows some of the most frequently occurring words, namely the word "application" 627 times, the word" like" 523 times.The words that appear in Fig. 3. have positive sentiments and also are the topic of positive reviews that are most widely reviewed by Shopee application users.Then, the association words are shown in Table 3.  3, the word associations related to the word "application", there is information obtained that many users who access the Shopee application are overall good and great.Then, the word "discount" also give information on how the Shopee application provides large discounts to users.After that, the word "satisfy" informs that the friendliness and response of the seller and the customer service in the Shopee application are very satisfying for the users and the security in shopping also satisfying for users.After that, the word "promo" give information that the promo held by Shopee is very attractive for users, so Shopee application users are interested in shopping at the Shopee application.
d. Find the joint probability p (a, b) for testing data } the topic of testing data document by looking at the largest a * value in a

Table 3 .
Association Positive WordBased on Table.3, we obtained several word associations in the classification of positive sentiment.Table