ANALYSIS SENTIMENT OF NESTLE BEAR BRAND DURING THE COVID-19 PANDEMIC ON SOCIAL MEDIA TWITTER

Position and brand image are important and crucial in marketing strategy. Brand position and image form strong associations with targeted consumers used to differentiate a brand from its competitors. Companies must understand good marketing strategies to improve their brand position in society based on consumer perspective. During the Covid-19 pandemic, especially in 2021, there was a unique phenomenon, whereby the demand for Bear Brand Milk was greater than the market leader, namely Ultra milk, while Ultra milk was even cheaper than the Bear Brand milk. This phenomenon is discussed on social media, especially Twitter regarding opinions from consumers in purchasing the products. This study aims to determine whether online text reviews can provide an overview of the position and brand image of Bear Brand Milk using LIWC (Linguistic Inquiry and Word Count) sentiment analysis and PCA (Principal Component Analysis). Based on our analysis, we can conclude that LIWC can provide an overview regarding the brand image and brand positioning of Bear Brand milk and Ultra milk. Brand image is obtained from variables that describe psychological variables when using or imagining the brands. Brand Position through PCA analysis describes the difference in gain between the dominant variables in the two brands.


Introduction
Nowadays brand position is very important and crucial in marketing strategy.Consumers' experiences with products are part of their information and can be considered important in their evaluation of a brand (Rao and Monroe, 1989).From the buyer's point of view, a clear brand position can reduce uncertainty and provide a strong reason for buyers to buy a product (Carpenter and Nakamoto 1989).Companies with a good brand position can be first movers, earn buyers' trust, and become the industry standard of their business (Carpenter, and Nakamoto 1989).The brand position forms strong brand associations to targeted consumers and is used to differentiate a brand from its competitors (Keller et al., 2002).Companies must understand consumer assessments of their products so that they can carry out good marketing strategies to improve their brand position in the public.
Social media is a place where consumers can express their opinion on a brand, including one type of eWOM (Electronic Word of Mouth).eWOM in consumer reviews itself usually consists of words uploaded on social media (Berger, 2020).These words are in the form of consumer opinions on their experience when using a brand or product.Negative articulation influences consumer behavior more strongly than positive articulation, and the same pattern exists in the online environment (Hennig-Thurrau et al, 2014).Purchase changes based on the articulation of opinions can influence online and offline buying behavior (Hennig-Thurrau et al, 2014).eWOM can ultimately provide consumers with alternative sources of information, thereby reducing the ability of brands to influence consumers through traditional marketing and advertising (Duan et al, 2008).
During the Covid-19 pandemic, especially in 2021, there is a unique phenomenon, namely the demand for Bear's Milk (Bear Brand) is greater than the market leader, Ultra Milk, even though both are sterile milk in packs and the price for Bear's Milk is much higher Rp.4,000 compared to Ultra Milk in normal conditions.
However, during rare conditions such as a pandemic, the highest price for Bear's Milk when it is rare can reach Rp. 50,000 per can of 250 ml with a difference of up to Rp. 44,000 from its competitor, Susu Ultra.In some circulating news it is also said that the stock of Bear Brand Milk in stores is very scarce and the price increases when purchased on the online market (tribunnews.com, Fimela.com, cnbcindonesia.com).Purchasing decisions by these consumers are of course based on brand position and certain brand images so that many people choose Bear Brand Milk over Ultra Milk even though Ultra Milk is the leader in market share and the price is cheaper in Indonesia.These associations and brand images can be searched using Big Data because in this modern era, many consumers share their experiences when using products on social media such as Twitter.
Researchers chose Twitter as a place to find data because Twitter is a social media that has users who are free to express their opinions daily compared to other social media.Besides that, Twitter is used by those who communicate the latest news when events occur (Ahmed, 2015), Twitter also has a strong hashtag culture that makes it easy to collect, sort, and broaden searches when collecting data.Therefore, Twitter data is easier to retrieve when things go viral or are widely discussed on Twitter because these conversations tend to focus on hashtags.Twitter API to access tweet data is more open and easier to access compared to other social media platforms.This makes Twitter easier for researchers to access and provide data.Tweets shared by these users can be used to see sentiment and position as well as the brand image of the product so that it can provide an overview of how the market views Bear Brand Milk and Ultra Milk products during a pandemic.Several studies have been used to look at brand image and product brand positioning through sentiment analysis on social media, namely a research written by Alzate et al in 2015 with the title "Mining the text of online consumer reviews to analyze brand image and brand positioning".This study proposes an integrated, structured, and easy-to-implement procedure for online review text analysis with the ultimate goal of studying brand image and brand positioning.Text mining analysis is based on a lexicon-based approach, Linguistic Inquiry and Word Count (LIWC) (Pennebaker et al., 2007), which provides researchers with insight into emotional and psychological brand associations.The second study was written by Pathak, Sema et al in 2017 with the title -Sentiment analysis of virtual brand communities for effective tribal marketing.‖In this paper, the authors have shown how combining netnographic data with sentiment analysis using a spreadsheet-based tool helps in an in-depth examination of emotions.that unfolds in virtual communities.Companies should focus on non-commercial initiatives that speak to the specific interests and goals of different brand communities.These two studies are used as a reference in this research because they can see the brand image and brand position on sentiment in a text which can be used as a reference in acting through a brand's marketing actions.

Literature Review
Literature review consists of several parts, namely the concept of brand image and brand position in relation to marketing strategy, methods for measuring brand image and position, NLP (Natural Language Processing), and sentiment analysis.The broad picture of this chapter is to explain the concept of reputation and brand position in purchasing packaged ready-to-drink milk, especially during Covid-19 and can be illustrated through big data processing.

Concept of Brand Image and Brand Position
Fombrun and Van Riel (in Stenger, 2014) define reputation as "The overall award in which a company is held by its constituents".Furthermore, Stenger (2014) suggests that reputation can also be considered an intangible asset.Fomburn and Van Riel's definition is in line with Kumbhar's (2012) research which reveals that brand reputation is part of the service quality dimension evaluated by service consumers.

Brand Image
Brand image is an exogenous perception experienced by customers about a brand (Chakraborty and Bhat 2018).Building a positive brand image is very important as it helps the organization to stand out from its competitors.Understanding the content and structure of a brand's image is important because they influence what comes to consumers' minds when they think of a brand.Brand image content is related to the types of associations (e.g., product-related and non-product-related associations, perceived usefulness of brand use, overall attitude toward the brand) that consumers hold in their memories, the liking of these associations (positive and negative associations), and their uniqueness ( Gensler et al, 2015)

Brand Position
Positioning -refers to how customers think about the proposed and/or present brand in the market (Perreault and McCarthy 1999).‖Through brand positioning, companies try to build a sustainable competitive advantage on product attributes (tangible or intangible) in the minds of consumers.These advantages are designed to appeal to one or more segments within the product category.Developing a successful positioning strategy is not easy, marketers must consider four things: the target market, how the product is different or better than competitors, the value of the difference to the target market, and the ability to demonstrate or communicate this difference to the target market.These elements relate to the components of brand positioning as described by Aaker (1996): target audience, part of the identity/value proposition, creating profit, and communicating actively.Brand identity and positioning are critical to developing a strong customer base and brand equity.

Changes in Purchase Patterns during COVID-19
In the case of Covid-19, or the novel coronavirus-19, the general public in every country reacts within approximately two weeks of becoming aware of the presence of the virus in their country, to start stockpiling their goods (Russel, 2021).When this preparedness mentality kicks in, the following categories take priority: medical supplies, rubbing alcohol, antibacterial wipes, first aid kits, antiseptics, cold and flu medications, and cough drops (Nielsen 2020).With many consumers feeling left behind with these purchases, it can provide a perspective for other individuals as a reference for goods that will remain readily supplied by the population in the future (Russel, 2021).

Brand Image and Brand Positioning Measurement Method
The measurement method that will be used in this research is a combination of qualitative and quantitative methods.The use of these measurements is because quantitative measures focus on the relationship between brands and attributes without considering the true meaning of why this can occur from brand image (Brand et al, 2011).So, if we want to know and analyze whether the attribute is important or not, which attribute is the first or second-order attribute, or the type of relationship between the attributes, we must use both methods.The merging of these methods can occur because this study analyzes consumer opinions directly through the text of the opinions they assume on the internet (qualitative methods) while the amount of data to be processed requires researchers to use text mining as a tool to process data where text mining itself uses calculations statistics to benefit from quantitative methods.
Text mining processes unstructured text consists of various types, namely documents, comments, reviews, or other types of information.This technique allows us to extract the content of texts -such as reviews -and classify them into positive or negative ratings based on their polarity (Casalo, Flavian, Guinal u, & Ekinci, 2015), as well as gain insight into consumers, attitudes, opinions, and emotions as a whole.Text mining typically includes text categorization, text grouping, and sentiment analysis, among others (Srivastava & Sahami, 2009).

NLP (Natural Language Processing)
NLP defines the expression of subjectspecific sentiments and classifies the polarity of the sentiment lexicon.NLP can identify text fragments with subject and sentiment lexicon to perform sentiment classification, instead of classifying the sentiment of the entire text based on certain subjects (Nasukawa and Yi, 2003).
This approach focuses on general text, and gets better results by removing some words, such as ambiguous sentences, or sentences that lack sentiment.Machine learning and previous NLP studies in sentiment analysis for text may not be suitable for sentiment analysis for tweets, because the structure between tweets and text is different.

Sentiment Analysis
Sentiment analysis is an automated process of understanding opinions on a particular subject from written or spoken language.Sentiment analysis is also known as opinion mining which is an area that includes natural language processing that extracts opinions hidden in text (Kawashima et al., 2013).There are three attributes in extracting the expression: a) what kind of polarity is expressed by the customer in his review, it can be positive, negative, or neutral.b) subject-the thing being talked about.c) opinion holders-customers who express opinions about a product through reviews.
Today, sentiment analysis is a premium subject and a tremendous improvement because it has many functional applications.As data that can be accessed openly and confidentially via the internet continues to grow, a large number of writings that communicate feelings are accessible on review sites, discussion sites, web journals, and social media.Sentiment testing frameworks can assist in turning this unstructured data into organized information from popular sentiments about products, administrations, brands, legislative issues, or any subject that individuals may express (Russom et al., 2011).This information can be useful for business applications such as display inspection, advertising, item surveys, net advertiser scoring, item input, and client administration.
Sentiment analysis that researchers use is through the LIWC or Linguistic Inquiry and Word Count program.LIWC can show that language can provide very rich insights into their psychological state, including their emotions, thinking styles, and social concerns which in turn can provide an overview of brand positioning and image.

Research Methodology
The author collects data on Twitter and then cleans it up in Python.Data that is clean from noise will be included in the LIWC program to generate sentiment values for each data.The data will be further processed so that it can present the percentage value and sentiment statistics.Sentiment data that has been processed will be entered in R so that it can produce a brand position map.Figure 1 shows the stages in data collection and analysis carried out by the author

Data Acquisition
Python connects Twitter API with help of snscrape library.First, you must make a Twitter API creation request which will later have a token used in coding.Second, install the snscrape library in Python and create the variables in Python.After that we can search for tweets based on certain users or keywords with a certain number randomly so that later data will be obtained based on random sampling of at least 1000 twitter posts which will later be processed with Python.The explanation above the sampling method used is the random sampling method where data collection is taken randomly without any specific criteria with the amount we want.

Data Cleaning and Filtering
The data cleaning process begins by eliminating @ through the help of the pandas library and RegEx as a library.RegEx is used to remove and search for certain characters in a sentence, namely @ and the sentence that follows it in the data frame.Furthermore, the authors reduce the letters in the data because it can interfere with the results of data processing using the string function on the dataframe by changing all letters to lowercase with "lower()".Next, the author does a link cleaning but first duplicates the reduction column in order to review the data that was previously cleaned whether there are errors or not.The column is then processed again to remove the link or url through the help of ReGex where if there is a sentence with http or https it will be removed by ReGex.Furthermore, the researcher will change the language of the data from Indonesian to English because the sentiment and topic modeling carried out can only be used in English because there is no sentiment dictionary in Indonesian in the LIWC.However, because the psychological score is calculated based on the occurrence of words, the wrong word order in the translated text will not have much effect (Tiffany and Desi, 2017).

Sentiment Analysis Using LIWC
Linguistic Inquiry and Word Count (LIWC) stems from decades of scientific research showing that each individual's language can provide enormous insight into their psychological state, including their emotions, thinking styles, and social concerns (Tausczik & Pennebaker, 2010).Linguistic Inquiry and Word Count (LIWC; Pennebaker, Booth, & Francis, 2007) is a wordcounting software program that references a dictionary of grammatical, psychological, and content word categories.LIWC has been used to efficiently classify texts along psychological dimensions and to predict behavioral outcomes, making it a text analysis tool widely used in the social sciences.
The purpose of using LIWC is that LIWC can cover aspects that other programs have not been able to analyze such as standard LIWC reads a specific text and compares each word in the text with the dictionary word list it has and calculates the percentage of total words in the text that match each dictionary category.For example, if LIWC analyzed a single utterance containing 1000 words using the built-in LIWC dictionary, LIWC might find that 50 words are related to positive emotions and 10 words are related to affiliation.LIWC will convert these figures into percentages: 5.0% positive emotion and 1.0% affiliate.This value comes from the assumption that 50 positive words out of 1000 words are processed by means of calculations, namely the 50 words earlier divided by 1000 words, namely 0.05 multiplied by 100% produces a value of 5% positive words out of a total of 1000 words processed.
LIWC provides a total of about 90 output variables.Some of them are general descriptors (e.g., word per sentence), others are standard linguistic dimensions (e.g., pronouns, articles, adverbs, prepositions and auxiliary verbs) and 53 variables belonging to a group called -Psychological Processes‖.In this group of 53 variables, there are 10 general variables or categories and 43 more specific variables or subcategories.In this study, the researcher used a set of 26 variables, which fall into 7 general variables or categories under the "Psychological Process" group: influence, social processes, perceptual processes, biological processes, drives, relativity, and personal attention.
LIWC review scores are aggregated to the brand average.N represents the number of brands analyzed.Because it represents the brand average, the minimum statistic belongs to the brand with the minimum average score on each text variable and the maximum statistic belongs to the brand with the highest average score on each feature.For example, the PosEmotions variable measures the level of positive emotion associated with any given brand, a high score in this category means that consumers associate the brand with positive experiences; The Power variable measures the number of power-related words that appear in online reviews of any given brand, and a high score in this category means consumers associate the brand with strength.

Perceptual map using PCA
Principal Component analysis (PCA) is a dimension reduction method often used to reduce the dimensions of large data sets, by converting a large set of variables into smaller ones that still contain most of the information in the large set.
Researchers propose to use the linear method.Principal Component Analysis (PCA) to project data linearly into spatial dimensions.PCA can reduce the dimension that can be used to visualize and explore the structure in the data and extract plots that better depict the data (Cunningham & Ghahramani, 2015;Gwin & Gwin, 2003).
Brand mapping consists of a graph of competitors' brand positions in the market based on their location space, as defined by key dimensions (Tirunillai and Tellis, 2014).Overall, PCA is a technique touted as correlated variables (p) into a smaller number of k (k < p) uncorrelated variables called principal components, while maintaining as much variation as possible in the original data set.
Complex data causes a data to have many PCs or main components.Principal components are constructed in such a way that the first principal component accounts for the largest possible variance in the data set.The second principal component is calculated so that the second principle is uncorrelated with the first principal component or perpendicular to the first principal component and the second highest variance.This continues until the total p principal components have been calculated, equal to the sum of the initial variables.
In this study, PCA was run in R software using the -factoextra‖ package (Kassambara and Mundt, 2020).26 variables (p) generated from LIWC are used as input variables to obtain a brand perception positioning map using Principal Component Analysis (PCA).PCA is usually used when variables are correlated.Therefore, with our data, the first step is to analyze the correlations between the text variables, where some positive correlations are high (Family/PosEmotions, Space/Work, Body/Power and Reward/Affiliate, NegEmotions/Risk and Family /Home) can be observed.Therefore, the use of PCA is justified to avoid the potential problem of multicollinearity

Results and Findings
This research aims to provide an overview of the viral phenomenon of Bear Brand Milk and its relation to brand position and image during the pandemic, especially in 2021.Changes in the current pandemic pattern that are less stringent and clearer regulations allow companies to further improve their brand image and position in order to increase their market share.Product and brand management can add or maintain one of its brand images and positions in order to provide insight and understanding about the position and brand image of a product by using sentiment analysis.Aspects that influence brand image and brand positioning can be taken using PCA analysis through LIWC and sentiment analysis which can provide support for an in-depth description of brand image and positioning.PCA is used to describe the position and image of the brand within a year so that it can support and complete sentiment analysis.

Description of Sentiment Analysis using LIWC
The research was divided into two processes, namely sentiment analysis using LIWC and LIWC analysis using PCA depiction.This research uses data that has been obtained in the amount of 8154 data consisting of 4175 tweets about Bear Brand Milk in 2021 and 3979 tweets about Ultra Milk in 2021.The data consists of several words that have meaning and sentiment value and will be analyzed in depth using sentiment analysis.
Visualization and data exploration on sentiment analysis will describe the psychological level of the data based on the LIWC Program which contains a dictionary of about 4500 words covering a number of dimensions.Through a word counting strategy, the text is analyzed on a word by word basis, each word being compared to a pre-dictionary.The linguistic indicator score for each LIWC variable is calculated as the percentage of words that match a predefined dictionary.To measure the level of positive emotion in online reviews, for example, LIWC counts the number of times words defined in the dictionary as related to positive emotions (e.g., "love", "good", and "beautiful") appear in reviews and divide the result by the number total words in the text.The words in a sentence affects the value of sentiment that often appears, but because we have so much data and use the average value of the sentiment data this effect can be reduced.

Sentiment Analysis of Bear Brand
To measure the degree of variable in online tweets, LIWC counts the number of times words defined in the dictionary are associated with one of positive emotions (e.g., "love", "good", and "beautiful") appear in reviews and divide the result by the number total words in the text.
The statistical results of Bear's Milk sentiment data can be seen in table 1

Sentiment Analysis of Ultra Milk
In order to make it easier to analyze the sentiment results, researchers used a radar plot that illustrates the difference in the mean value of the LIWC sentiment variable Bear Brand Milk and Ultra Milk.The results of the comparison of the two milks can be seen in table 4.3 and simplified with a radar plot in figure 4.1 where the figure illustrates that Ultra Milk (green line and green dot) has a high average value in the ingest, post emo, see, and space categories.compared to Bear's Milk.Bear Brand Milk (blue line and blue dot) has a high average value on the variables time, work, leisure, money, and neg emo.This variable illustrates that the brand image of Bear Brand Milk in 2021 is closely related to time, work, free time, money and negative emotions.This is based on the fact that in 2021 many consumers associate Bear Brand Milk with being able to improve health, even though many people disagree with this opinion.In addition, Bear Brand Milk is also often drunk at various times, related to work, often consumed during leisure time, and the price is expensive due to the large number of requests from consumers.
The brand image of Ultra Milk has something quite different from Bear Brand Milk where Ultra Milk is often drunk directly or used as a food ingredient which is of course different from Bear Brand Milk.Ultra Milk itself has a high positive emotional sentiment value compared to Bear Brand Milk, this is because Ultra Milk has a variety of tastes and according to the preferences of each consumer and likes to be drunk anywhere.This is consistent with studies that have been conducted where mood plays an important role in food selection and convenience is preferred in certain circumstances, such as when individuals are experiencing pain (Locher, Yoels, Maurer and Van Ells, 2005).Consumers choose to eat comfort food when they are sad or feeling lonely (Spence, 2017).High-calorie sweet foods, ice cream, cakes and chocolate, for example, elevate mood due to the production of serotonin and opiates (Stein, 2008).

Correlation Between Variables
A correlation test is used to assess the relationship between two or more variables.Correlation between variables plays an important role in descriptive analysis.Correlation measures the relationship between two variables, i.e. how they are related to one another.In this sense, the correlation makes it possible to tell which variables evolve in the same direction, which evolves in the opposite direction, and which are independent.
Regarding the direction of the relationship: On the one hand, a negative correlation implies that the two variables under consideration vary in opposite directions i.e. if one variable increases the other decreases and vice versa.On the other hand, a positive correlation implies that the two variables under consideration vary in the same direction, i.e., if one variable increases, the other increases, and if one decreases, the other also decreases.
PCA is usually used when variables are correlated.Therefore, with our data, the first step is to analyse the correlation between the text variables, which are shown in

PCA Analysis (Principal Component Analysis)
Principal component analysis or principal component analysis is a factor analysis approach that considers the total variance in the data.Principal component analysis (PCA) is recommended when the researcher's primary concern is to determine the minimum number of factors that will explain the maximum variance in the data used in a particular multivariate analysis.
Principal components are created from new dimensions (i.e.principal components) from the original data.Information in data is squeezed or compressed into its first components.So, the idea is 10 dimensional data gives 10 principal components, but PCA tries to include as much information as possible in the first component, then the maximum remaining information in the second and so on.The next step after entering the coding and processing the PCA is to look at the proportion of variance explained by each component.Table 5 shows the standard deviation and variance of the 26 principal components (PC) obtained from the PCA, which allows a reduction in dimensionality from 27 to 26 components (PC) while retaining the variance in the data and 70% of which is explained by only the first five components.In general, when applying the PCA method, researchers tend to choose the first two principal components to plot a two-dimensional position map.In figure 3 it can be seen in the researcher's data, PC1 and PC2 account for more than half (53%) of the variance in the data and, if we add PC3, PC4 and PC 5 then the explained variance increases to almost 70% of the total variance.
The proportion of variance explained by each PC is one of the factors that determines the number of PCs retained for interpretation.There is no general rule for determining the required minimum of the variance described from a defended PC.Nevertheless, there are some general recommendations.Samuels (2016) for example, recommends a minimum of 50% of the variance explained.In this study, we provide 26 components but can only present the first five for further interpretation.To further facilitate the interpretation of the relationship between the variables and the PCs additional rotation can be applied to PCs to produce-

Table 4 Correlation Between Bear Brand Milk Variables (Top) and Ultra Milk (Bottom)
high-factor loadings for some of the variables.In other words, a small number of variables will be highly correlated with each PC.As a result we can analyze the load which illustrates the importance of the independent variables.PC results can be seen in tables 6 where these tables explain what variables are dominant for each PC with an asterisk (*).This aims to understand that each PC has a dominant variable, which means that these variables greatly influence the distribution of data.On PC1 the ingest variable is very dominant, which means that the ingest value is large and has an impact on the distribution of data that is also large so if we want to see the distribution of data through the dominant variable, we have to look at which PC the dominant variable is, with a note of seeing the efficiency value> 1 and the total variance of the total PC.combined values exceed 50% according to the literature.
The main purpose of PCA is to obtain a position map, where the PCA output provides the necessary information about the values of text variables in different dimensions.Figure 5 shows that Ultra Milk has several dominant components of text variables which can be indicated by blue lines, namely ingest and positive emotion.Figure 6 shows that Bear Brand Milk has more dominant time and leisure variables.
Researchers also use data on the average sentiment variable as an illustration of how clearly the position of each brand is.From Figure 7 it can be seen that Bear Brand Milk (point 1) has a brand position associated with friend, neg emo, health, time, work, money, leisure, and power which is more dominant while Ultra Milk is more dominant in the post variables emo, ingest, see, and spaces.

Discussion
In conclusion, in this study, researchers conducted an in-depth analysis of brand text variables related to Bear Brand Milk and Milk with datasets to get more in-depth results regarding what variables are in the minds of consumers using PCA modeling and sentiment analysis.The big data and machine learning approaches are applied to two milk brands, namely Susu Bear and Susu Ultra as a comparison with 8154 data from Twitter in 2021.The following are the conclusion points that answer the research objectives: 1. Bear Brand Milk has a high average value on the variables time, work, leisure, money, and neg emo compared to Ultra Milk.The Bear Brand Milk variable illustrates that the Bear Brand Milk brand image in 2021 is closely related to time, work, free time, money and negative emotions.This is based on the fact that in 2021 many consumers associate Bear Brand Milk with being able to improve health, even though many people disagree with this opinion.Besides that, Bear Brand Milk is also often drunk at various times, related to work, often consumed during leisure time, and the price is expensive because of the large number of requests from consumers.2. Ultra Milk (green line and green dot) has a higher average score in the ingest, emo post, see, and space categories compared to Bear Brand Milk.Ultra Milk's brand image is quite different from Bear's Milk where Ultra Milk, according to Twitter sentiment data, is often drunk directly or used as a food ingredient, which is certainly different from Bear's Milk.Ultra Milk itself has a high positive emotional sentiment value compared to Bear Brand Milk, this is because Ultra Milk has a variety of tastes and according to the preferences of each consumer and likes to be drunk anywhere.This is consistent with studies that have been conducted where mood plays an important role in food selection and convenience is preferred in certain circumstances, such as when individuals are experiencing pain (Locher, Yoels, Maurer, & Van Ells, 2005).Consumers choose to eat comfort food when they are sad or feeling lonely (Spence, 2017).When consuming high-calorie sweet foods, ice cream, cakes and chocolate can increase emotions and reduce negative emotions such as ultra milk which has various flavors (Macht & Dettmer, 2006) 3. Correlation means association is a measure of the extent to which two variables are related.In the data the researcher looks for a positive correlation between variables, this can show a relationship between two variables where both variables move in the same direction.PCA can describe a positive correlation between variables moving in the same direction.So that we can see a good correlation between variables with a note that PCA is used when there is a correlation between the variables.In the results of the PCA analysis it can be said that the circle is the average value of the data while the points are the values of each individual data plotted in PCA.PC1 has a very large proportion of variance values, namely 45%, so it can be used to combine with other PCs with a total proportion of the two PCs > 50% in order to provide reliable information on the data.PCA PC1 and PC2 illustrate that the position of Ultra Milk has a more dominant variable value in emotion and ingest posts, PC1 and PC 3 illustrates that the position of Bear Brand Milk has a dominant variable value in time and leisure, PC1 and PC 4 illustrates that the position of Ultra Milk has a dominant variable value in space, PC1 and PC 5 illustrates that the position of the Milk Bear has a dominant variable value in time and leisure.PCA based on the average sentiment analysis variable data can provide a clear picture of the position of each brand.From the findings it can be seen that Bear Brand Milk has a more dominant brand position associated with friend, neg emo, health, time, work, money, leisure, power which is more dominant while Ultra Milk is more dominant in the emo, ingest, see, and space post variables.This position in PCA can provide an overview of what brand images are seen by consumers and are dominant in each brand.As marketers, of course we have to be observant in seeing what advantages our brand has and in comparison with other brands.In this case sentiment analysis and PCA can be used as a reference in knowing the image and position of a brand

Managerial Research Implications and Contributions
The researcher suggests that businesses can benefit from integrating sentiment analysis tools together with eWOM data.If Companies choose to use these tools, the expected increase in the speed of their decisions will likely increase their satisfaction with the overall customer experience and increase the popularity of the products and brands offered.Sentiment scores as used in our experiments will also have practical managerial value especially if one can supplement them with further text analytic techniques such as clustering.Product and brand managers can use breakdown to identify variables that appear in positive or negative reviews.Managers can then direct their efforts to strengthening the variables (e.g., product ease of use) that evoke positive sentiments while modifying the attributes of their offerings (e.g., unpalatable) that evoke negative sentiments.The potential for future research I feel is quite significant in this area of managerial decision support.
The results of the study show that data on Bear Brand Milk is closely related to negative emotions containing complaints and consumer distrust of the efficacy of Bear Brand Milk.This of course can give a negative stigma to the company and it would be better if the company could remove this stigma by conducting research that Bear Brand Milk products can at least provide health benefits in the body which will later remove this stigma and provide greater positive opinions regarding health so that it can used as a marketing object which in my opinion is quite good compared to its competitors.
Ultra Milk has a very high sentiment value in the ingest variable which shows that Ultra Milk is often associated with the consumption of food and drink or as a raw material for food or drink.This brand image can be used further to market its products in various ways, one of which is cooking recipes.This is intended so that Ultra Milk can target other market segments other than as ready-to-drink milk but as a food ingredient which of course can be beneficial for the position and brand image of Ultra Milk.

Research Limitations and Recommendations
The researcher uses tweet-based data for packaged milk products, namely Ultra Milk as a comparison because it is a product with the highest market share, while Bear Brand Milk is the focus of the research.Analysis of the broader picture of a brand's position in other brands is of course not always the same as the variables taken but depends on the product being analyzed.Therefore, each brand needs to go further by examining their case and thus draw more specific conclusions.
The use of brands in research that only consists of two brands can only describe competition between two companies.In order to be able to see the market conditions of the types of products circulating, the researcher recommends that further research should use many brands of the same type of product.This is used to further see the actual market conditions which of course can be beneficial for many parties.
The use of LIWC in the text mining literature is of course often used, the researcher recommends that further research can use other sentiment calculations based on the available lexicon.Machine learning methods for text mining can also be used by companies wishing to extract more specific aspects of online review text, provided they have the necessary resources.
In this research data is only centered on tweets, so that the data obtained is more accurate researchers can add all types of eWOM, such as reviews, social networks, and blogs.The brand association data for the researcher's analysis was obtained via a specific online source, so it would be interesting to compare brand associations across different online sources, to detect possible differences and verify whether they are platform dependent

Figure 3 .
Figure 3. Data Proportion on each PC

Figure 4 .Figure 5 .
Figure 4. PCA from PC1 and PC2 Milk data.It can be seen that the large average values from largest to smallest in LIWC Bear Brand Milk data are ingest, time, positive emotion, space, leisure, negative emotion, and money.This large average value can illustrate that Bear Brand Milk is most likely related to this variable.The small average values from smallest to largest in LIWC Bear Brand Milk data are death, sexual and religion.This small average value illustrates that the Bear Brand Milk data is most likely not related to this variable.

Table 3 Comparison of the average value of each brand
Figure 2. Radar Plot of the Average Sentiment Value of Each Variable