Quantifying customer sentiment for automobile brand perception analysis using machine learning on Twitter

Abstract

Social networking sites provide a platform for individuals to express their opinions publicly. Brand managers actively use these platforms to gain insights into brand perceptions, as users often share their views on products and services. In this study, we use sentiment analysis to assess customer sentiment towards five leading automobile brands, analyzing text content shared on Twitter(or X). The research models the ’Brand Polarity Score’, which indicates whether customers perceive the brand positively or negatively. This score is further weighted based on the tweet’s influence, characterized by the engagement metrics of the tweet and the author’s follower count. We also demonstrate how this brand polarity score can effectively communicate near real-time brand positioning, providing a valuable tool for monitoring brand sentiment over time. The proposed Brand Polarity Score (BPS) not only gauges brand perception but also serves as a reliable tool for progressive and competitive analyses, contributing to a comprehensive understanding of brand dynamics. A comprehensive validation strategy—including event-sensitivity analysis, cross-model convergence checks, and stability assessments—demonstrates the robustness of the proposed BPS system.

Introduction

Brand perception refers to how consumers perceive and interpret a particular brand. It encompasses the thoughts, feelings, attitudes, and overall impressions individuals associate with a brand based on their experiences, interactions, and exposure to its marketing and products. Brand Image as a whole is a hot research topic because of its strong association with different marketing constructs^1,2. Researchers have given several definitions to brand image. According to³, “brand image is largely a subjective and perceptual phenomenon that is formed through consumer interpretation, whether reasoned or emotional”. In⁴, brand image is defined as “a set of associations, usually organized in some meaningful way” and in⁵ it is defined as “perceptions about a brand as reflected by the brand associations held in consumer memory”. Brand image is perceived differently by the consumers and the company itself. The organization/company wants to know how the consumers perceive their brand². A Positive brand perception fosters customer loyalty, influences purchasing decisions, and enhances a brand’s market position, financial performance, and overall competitiveness. Conversely, negative brand perception can lead to customer dissatisfaction, reduced loyalty, and adverse effects on a brand’s market share and profitability. Effective brand management involves consistently delivering positive experiences, aligning with consumer expectations, and actively shaping perceptions through strategic marketing and customer engagement efforts.

Traditional customer perception analysis primarily uses techniques such as focus group discussions, surveys, and in-depth interviews. They often employ structured questionnaires to collect quantitative and qualitative data, aiming to gauge customer satisfaction, preferences, and overall perceptions. However, as explored in the study⁶, traditional methods have certain limitations. For instance, consumer’s responses to survey questions can sometimes be misleading. This can occur when consumers express opinions about a product they haven’t purchased yet, or when they comment on a brand about which they have no knowledge. Additionally, they can be time-consuming and expensive to administer, limiting the frequency and scale of data collection. Survey results may lack real-time relevance, making it challenging to capture dynamic shifts in customer sentiment. Moreover, survey fatigue can lead to decreased respondent engagement, impacting the quality and reliability of the collected data. Lastly, the fixed nature of survey questions may not capture nuanced or spontaneous aspects of customer perception, limiting the depth of insights obtained through this traditional analysis method.

Social media interactions, particularly on platforms like Twitter and Facebook, offer a dynamic and valuable source of data for understanding and shaping brand perception and has shown extremely valuable to marketing research companies, public opinion organizations and other text mining entities^7–9. 77% of businesses use social media to reach customers, and 90% of users follow at least one brand on social media¹⁰. Users frequently express their opinions, experiences, and sentiments about brands in real time, creating an unfiltered stream of information. Analyzing these interactions allows brands to gauge the immediate and evolving reactions to their products and services. The most interesting fact is that 71% of consumers who have had a positive experience with a brand on social media are likely to recommend the brand to others¹¹. Through sentiment analysis and machine learning models, the content shared by users on Twitter can be assessed for positive, negative, or neutral tones, providing a quantifiable measure of brand sentiment. The real-time nature of social media interactions enables brands to respond promptly to feedback, address concerns, and engage with their audience, fostering a more positive perception. Additionally, by monitoring patterns over time, brands can identify emerging trends, assess the impact of marketing initiatives, and make data-driven decisions to actively shape and enhance their overall brand image.

However, despite the richness of social-media data, existing approaches to brand-level sentiment measurement often lack a unified, interpretable, and statistically validated metric that captures both short-term shifts and long-term trends in customer perception—highlighting the need for a robust, systematically designed Brand Polarity framework.

The present study addresses this gap by focusing on the systematic analysis of customer perception polarity toward leading automobile brands. While the empirical data in this work is sourced from Twitter, the methodology is platform-agnostic and can be readily extended to other channels such as Facebook, Instagram, YouTube comments, or product-review sites. Moreover, this approach is not limited to automobile brands and can be extended to different brand domains. Sentiment analysis is employed to quantify a Brand Polarity Score (BPS), and we introduce a longitudinal analysis of brand polarity progress, enabling brands to respond promptly to changes in perception polarity. Furthermore, we propose a novel Influence-weighted Brand Polarity Score () that not only accounts for the sentiment of user content but also considers the influence generated by the content.

From a managerial standpoint, the proposed Brand Polarity metrics— BPS, BPPI, and —offer practical decision-support value beyond their analytical formulation. BPS quantifies and monitors immediate fluctuations of consumer mood and early detection of reputation risks; highlights which issues are amplified by high-influence voices and therefore merit priority attention; and BPPI provides a long-term trajectory of brand sentiment that managers can use to evaluate campaign effectiveness, track competitive positioning, and guide strategic communication and customer-engagement decisions. Together, these indices allow marketing and customer-experience teams to translate dynamic social sentiment into actionable insights for resource allocation, campaign tuning, and crisis management.

This work is distinctive not only because it quantifies brand polarity through the proposed BPS and its cumulative and influence-weighted variants, but also because it offers a unified framework suitable for both longitudinal tracking and cross-brand comparisons. Furthermore, the robustness of these measures is demonstrated through a multi-layered validation strategy—including event-sensitivity analyses, cross-model sentiment convergence, stability and noise-robustness tests, and volatility comparisons—thereby establishing a systematic foundation for understanding the dynamics of brand perception.

This research contributes in the following ways:

• Proposes a suite of polarity measures—including the Brand Polarity Score (BPS), its cumulative form (BPPI), and influence-weighted variants (, )—to quantify consumer sentiment in both cross-sectional and longitudinal settings.

• Introduces a mathematical and computational framework for tracking brand perception dynamics, enabling both instantaneous sentiment monitoring and long-horizon trend evaluation- the first systematic exploration of polarity-based long-horizon brand tracking using cumulative sentiment indicators

• Develops a comprehensive validation strategy combining event-sensitivity tests, cross-model sentiment convergence, bootstrap-based uncertainty quantification, noise-robustness evaluation, and volatility comparison across the proposed indicators.

• Conducts an empirical comparison of pre-trained sentiment classification models on Twitter data, identifying the most suitable model for polarity estimation under real-world constraints.

• Demonstrates the practical applicability of the proposed indicators, showing how firms can utilise BPS/BPPI/ for brand monitoring, campaign evaluation, and competitive benchmarking.

Literature

Use of social media data in research

Social media content has been effectively utilized in research across various domains¹² that include healthcare¹³, information systems¹⁴, education¹⁵, marketing¹⁶ etc. This body of work demonstrates that user-generated content provides valuable insights into public opinion and consumer attitudes.

Brand perception studies based on attribute identification

There are several directions of research when trying to model brand perception. Most of the research focused on scoring the brands based on chosen attributes using social media data. In¹⁷, the authors proposed a method for inferring attribute-specific brand perception ratings by mining the brand’s social connections on Twitter. The approach measures the similarity between a brand’s Twitter account and a set of exemplar accounts representing a perceptual attribute. The work in¹⁸, proposed a framework to quantify the Online Brand Image(OBIM) from consumer reviews of mobile phone brands. They considered three attributes, favourability, uniqueness and strength, computed using sentiment and co-word network analysis. and the OBIM score is calculated as the multiplicative sum of these values. The work presented in¹⁹ focuses on modeling brand personality from social media content. Even though the work primarily focuses on a brand’s personality, they have also showed that the same framework can be easily adapted for finding the brand’s perceived personality i.e., how consumers perceive the brand. They analyze Tweet texts and returns scores for five brand personality dimensions: Sincerity, Excitement, Competence and Ruggedness. They proposed a hybrid machine learning method, LDA2Vec to label tweets with corresponding personality dimensions and then trained a Roberta model for classifying the tweets. In these works, the customer’s polarity towards the chosen attributes is not considered rather the polarity is inferred from the sentiment of the attribute itself. This creates a gap in identifying the customer’s sentiments towards the brand and focuses on identifying the alignment of the brands towards particular attributes.

Lexicon or dictionary based sentiment and brand evaluation approaches

In²⁰, the authors proposed a brand reputation tracker where the sentiment of tweet text is used to formulate the brand reputation score for 3 drivers and 11 sub drivers. They formulated a dictionary of positive and negative words for each driver. This dictionary is used to find positive and negative tweets belonging to each driver. The brand reputation score for each driver is calculated as the ratio of the number of positive tweets to the number of negative tweets. In⁶, they collected comments on 6 mobile phone brands from e-commerce platforms, where consumers in Turkiye shared their opinions after the purchase, and used aspect-based opinion mining methods to determine the product features that help to compare brands in perceptual maps. They calculated sentiment scores for each comment using lexicon-based sentiment analysis approach to compare the brands based on different mobile phone features. Zhao et al.²¹ proposed a method called the dynamic-Brand-Topic Model to get relevant topics and their sentiments from product reviews. They use a statistical model called the Gaussian state space model to show how sentiment and topic change over time. Similarly, the work in²² used sentiments of tweets to model brand reputation among popular mobile brands across a time period of 7 days and in the work⁹, a random sample of 3516 tweets was used to evaluate consumer’s sentiment towards well-known brands using an expert-predefined lexicon that includes around 6800 seed adjectives with known orientation. In the current literature, The works that use sentiments to model customer perception mainly rely on statistical models and predefined dictionaries to find the sentiments. The semantics of the whole tweet are not considered in such cases. Further a careful formulation of the dictionary is needed for ensuring the reliability.

Deep-learning sentiment classification approach

Sentiment classification approaches are mainly classified into (a) lexicon-based and (b) Machine-Learning based^23,24. Recent studies on sentiment analyses have shown that deep learning approaches provide more accuracy than shallow models^25,26. Especially transformer based models like BERT demonstrate state of art performance in text classification^27–29.

In the proposed work, we utilize the potential of deep learning to classify tweet texts and model customer polarity towards a brand. We consider the engagement of the tweet and the influence of the author to model a novel and reliable Brand Polarity Score. Further, we propose different longitudinal analysis techniques to keep track of customer polarity and also for performing a competitive analysis among different brands. As of our knowledge, this is the first work that uses a deep learning model to score and track the customer polarity of a brand.

Methodology

This section provides a comprehensive overview of our approach to modeling brand perception polarity using social media data. We delve into the intricacies of data extraction and sentiment detection, elucidating each step in detail. Furthermore, we outline our methodology for mathematically modeling the Brand Polarity Score (BPS) based on tweet sentiments and utilizing its cumulative measure as a Brand Polarity Position Indicator (BPPI). Additionally, we introduce the concept of the Influence-weighted BPS (), which not only considers the sentiment of the tweet but also factors in its influence. This enhancement adds an extra layer of reliability to our scoring technique, ensuring a more nuanced assessment of brand perception.

Data extraction

We retrieve data from Twitter (or X), a prominent social media platform, using its API³⁵. All tweets were collected via the search_recent_tweets endpoint under the Basic API plan, which permits up to 10,000 tweets per subscription month. Our collection window spanned November 30th to December 31st, during which the Basic-tier quota reset once mid-month. This provided two separate monthly allowances (20,000 tweets total), within which our final dataset of 15,765 tweets remained fully compliant. Tweets referencing any of the top five automobile brands identified in³⁶—BMW, Mercedes-Benz, Porsche, Tesla, and Toyota—were extracted using brand-specific queries of the form:

(e.g., @Toyota lang:en -from:Toyota).

The lang:en filter restricted results to English-language tweets, while the-from: operator ensured that tweets authored by the official brand account were excluded to focus solely on consumer perspectives. Since these constraints were enforced directly at query time, no additional cleaning was required for language or author removal. Please refer to Table 1 for a breakdown of the selected brands, the number of tweets extracted for each brand, and sample tweets. During extraction, we collected all tweet and author-level fields needed for analysis, including tweet text, timestamps, full public metrics (likes, retweets, quotes, bookmarks, impressions), and author information such as account creation date and follower statistics. These fields enabled both sentiment classification and the construction of influence-weighted polarity measures.

Table 1.

Selected brands with the total number of tweets extracted for each brand and sample tweets.

Brand Name	Total Tweets	Sample Tweets
BMW	3163	RT @BMW: A luxury view no matter which seat.
		@BMW Ahh, not such good news, it’ll be worth the wait though.
Mercedes-Benz	3094	Compared with other vendor’s 70% degrade warranty, Mercedes is taking care of their customers + the environment.
		Mercedes-Benz Blainville sent a bailiff to my home yesterday (November 30th 2023), which intimidated and harassed my family, as well as my landlord and his ringing the doorbell and pounding at both our homes non stop. Is this normal behaviour?
Porsche	3157	@patrickcandoit @Porsche These look awesome!!
		Bro, I don’t have this much CR @Porsche
Tesla	3076	No amount of words in my vocabulary will ever be able to describe the amount of excitement I have for the future right now! @Tesla is incredible @Tesla_Optimus is incredible. This is the future!!!!
		@Tesla Why don’t you respond to your customers and especially not reassure them about the future features you previously had?
Toyota	3167	RT @Toyota: Howdy partner! #Camry #LetsGoPlaces
		@Toyota I have 2013 rav4 and when it rains floods on the passenger side possibly from the sun roof any possible solutions

To ensure comparability across brands and days, we imposed a uniform cap of 100 tweets per brand per day imposed by setting max_results=100 for each (brand, day) query window. This fixed daily quota provided balanced exposure across brands and consistent temporal coverage throughout the collection period. This design minimizes the over-representation of brands with higher daily activity and enables fairer cross-brand comparisons while ensuring compliance with monthly rate limits. Although this approach yields a quota-based (non-probability) sample—since tweets are retrieved in reverse-chronological order dictated by the API,it is appropriate for our study because the goal is not to produce population-level sentiment estimates but to introduce a generalisable polarity-estimation methodology that organisations can employ to track evolving consumer sentiment towards their brands under realistic data-access and cost constraints.

All time stamps were obtained from the API in UTC. Since our focus is on relative day-to-day changes and cross-brand comparisons within a common temporal frame, we treat the UTC day boundary as a consistent and acceptable time scale, while acknowledging that local-time alignment for different markets is an avenue for future refinement.

Retweets or quote tweets were not removed beyond the exclusion of brand-authored content. Conceptually, we treat retweets and quotes as valid indicators of sentiment amplification; their presence reflects the degree to which sentiment-bearing content spreads across the platform.

Sentiment classification

Understanding public sentiment on Twitter requires the use of reliable and domain-appropriate sentiment classification tools. In this section, we discuss how we chose a sentiment classification tool to identify the sentiment of the extracted tweet data. To ensure that the sentiment labels applied in subsequent analyzes were derived from a well-performing and Twitter-optimized model,we selected five pre-trained sentiment analysis tools for evaluation, as summarised in Table 2.

Table 2.

Details of chosen sentiment classification tools.

Tool Name	Category	Description
Vader30	Lexicon and rule-based	VADER (Valence Aware Dictionary and sEntiment Reasoner) is a lexicon and rule-based sentiment analysis tool that is specifically attuned to sentiments expressed in social media.
TextBlob31	Lexicon and simple Machine Learning	TextBlob is a Python library for processing textual data. It provides a simple API for diving into common natural language processing (NLP) tasks such as part-of-speech tagging, noun phrase extraction, sentiment analysis, classification, and more.
Amazon comprehend32	Machine Learning based	AWS cloud NLP service that uses Machine Learning to perform tasks such as Sentiment Analysis, Entity Extraction, Topic Modeling, and more.
BerTweet33	Deep Learning/ Transformers based	BERTweet is the first public large-scale language model pre-trained for English Tweets. BERTweet is trained based on the RoBERTa pre-training procedure. The corpus used to pre-train BERTweet consists of 850M English Tweets containing 845M Tweets streamed from 01/2012 to 08/2019 and 5M Tweets related to the COVID-19 pandemic.
TweetNLP34	Deep Learning/ Transformers based	A Python library that provides a collection of useful tools to analyze/understand tweets such as sentiment analysis, emoji prediction, and named entity recognition, powered by state-of-the-art language models specialised on Twitter.

We first conducted a benchmarking step using the publicly available Sentiment140 corpus - a labelled dataset of 1.6 million tweets annotated as 0(negative) and 4(positive)³⁷. This dataset provides a large, human-labelled, in-domain reference set that enables objective comparison across sentiment analysis tools under consistent conditions. To enable computationally tractable benchmarking while maintaining representativeness, we used a randomly sampled subset of 10,000 tweets from Sentiment140. This approach follows standard practice for large-scale NLP datasets, which often contain significant redundancy. The sampled subset exhibited a near-balanced, binary class distribution (5039 positive, 4961 negative), allowing fairer performance comparison across tools.

Each tool’s predictions were compared with the benchmark dataset’s labeled values to compute overall accuracy. Among them, TweetNLP achieved the highest accuracy (84%) on this balanced test set, demonstrating its robustness for binary sentiment classification tasks³⁴. TweetNLP utilizes a RoBERTa-base model trained on a massive dataset of tweets (~120 Million tweets collected from January 2018 to December 2021) and fine-tuned for sentiment analysis using the TweetEval benchmark^38,39.

TweetNLP tool⁴⁰ was therefore selected for labelling the extracted automobile-related tweets in this study, as it combines strong accuracy, tweet-specific training, and open-source availability. Details of preprocessing, including removal of mentions and URLs, are provided in sentiment tool analysis section. Since our collected automobile tweets share the same linguistic characteristics as those in the benchmark dataset—short, informal, and often containing emojis, the two-stage evaluation approach (benchmarking on Sentiment140 followed by application to our dataset) ensures that the chosen model is both validly selected and generalizable to the target data.

Brand polarity score

Following sentiment classification, we obtained a labelled dataset of tweets mentioning each brand. Neutral tweets were excluded from the polarity analysis for both methodological and practical reasons. First, neutral posts do not convey a clear evaluative stance toward the brand and often consist of factual statements, generic mentions, or context-free references⁴¹. Neutral content on Twitter is also known to exhibit higher semantic ambiguity and lower inter-annotator agreement, which introduces additional noise into polarity computations. Second, the benchmarking step used to select the sentiment model (TweetNLP) was conducted using the Sentiment140 dataset, which contains only positive and negative ground-truth labels. The model evaluation and selection were therefore based exclusively on distinguishing positive from negative sentiment. To maintain methodological consistency—and because the neutral class could not be validated using the benchmark dataset—we restricted all downstream polarity metrics to positive and negative tweets identified by the selected model. This ensures alignment between the model-validation procedure and the sentiment labels used in the study.

After sentiment filtering, the study focused on evaluating the distribution of positive and negative tweets as an initial assessment of brand-level polarity. The filtering process resulted in 1875 tweets for BMW, 1933 for Mercedes-Benz, 1469 for Porsche, 1708 for Tesla, and 1888 for Toyota, which formed the basis for the subsequent polarity analysis. Figure 1 presents the total counts of positive and negative tweets for each brand during December, illustrating that BMW received the highest overall volume of favourable polarity-bearing tweets, followed by Mercedes-Benz and Toyota. Figure 2 further visualises the temporal distribution of positive and negative tweets for Toyota across the month.

Fig. 1.

Distribution of positive and negative tweet counts among top automobile brands.

Fig. 2.

Day-to-day distribution of positive and negative tweet counts for Toyota.

However, this initial analysis cannot provide a comprehensive understanding of overall customer polarity. To address this, we devised the Brand Polarity Score given by,

1

where is the total positive tweet mentions and is the total negative tweet mentions over a period.

The formulation of the Brand Polarity Score(BP Score) is grounded in established sentiment-indexing practices used in computational social science and marketing analytics which relies on forms of net favourability⁴². The BP Score follows this theoretical principle by capturing the relative balance of favourable versus unfavourable sentiment while adjusting for the total volume of sentiment-bearing mentions. BP Score is a numerical value ranging from – 1 to 1, where a score greater than 0 signifies a positive brand polarity and a score less than zero indicates a negative polarity. This score reflects the net polarity of consumer opinion and provides a balanced view of brand perception relative to competitors.

Figure 3 presents the aggregated BP Scores for the selected brands across the study period. Table 3 summarizes these results and additionally reports 95% bootstrap confidence intervals for each brand’s overall polarity. These intervals quantify the statistical uncertainty in the aggregated BP Score estimates by repeatedly resampling the tweet-level sentiment labels with replacement. Including 95% CIs clarifies whether observed differences between brands reflect genuine polarity shifts or fall within the range of expected sampling variability.

Fig. 3.

Aggregated Brand Polarity Scores for the top automobile brands during the study period.

Table 3.

Brands with counts of positive and negative tweets with corresponding BP Score and Bootstrap Confidence Intervals.

Brand	Positives	Negatives	BPS	95% CI
BMW	1581	294	0.6864	[0.65, 0.72]
Mercedes-Benz	1439	494	0.4889	[0.45, 0.53]
Porsche	1260	209	0.7155	[0.68, 0.75]
Tesla	1030	678	0.2061	[0.16, 0.25]
Toyota	1376	512	0.4576	[0.42, 0.50]

Contrary to the initial analysis, based on BP Score, Porsche exhibits the most favorable overall sentiment, followed by BMW. Unlike the initial analysis, where the brand positioning is determined by considering the positive tweets only, BP Score provides a comprehensive understanding of the brand’s polarity positioning because it jointly incorporates both supportive and critical sentiment. It is noteworthy that all brands exhibit a positive social polarity, which is unsurprising given their status as leading automobile brands.

A longitudinal analysis provides a more comprehensive understanding of time-sensitive brand perception, empowering brands to promptly respond and enhance their brand value. Figure 4 illustrates the BP score graph for Toyota throughout the study period. Notably, Toyota consistently maintains above-average BP scores on most days, except for December 2nd and December 26th. An analysis on polarity fluctuations on specific dates is provided in section where we discuss the practical implications of these scores. The interpretability and reliability of the BP Score are reinforced by the additional validation analyses conducted in this study and presented in validation section.

Fig. 4.

BP Score graph for Toyota throughout the study period.

Brand polarity position indicator

Competitive analysis of the daily BP Scores for all brands (Fig. 5) reveals short-term fluctuations that reflect daily shifts in consumer sentiment. As demonstrated in previous section, the BP Score performs well for cross-sectional comparisons and for capturing day-to-day movements. However, our validation analyses also show that BP Score exhibits high short-horizon volatility—an expected consequence of its sensitivity to daily tweet composition, sampling variation, and sudden bursts of positive or negative attention.

Fig. 5.

Daily Brand Polarity (BP) scores for chosen brands in longitudinal analysis.

Sudden spikes in daily polarity (positive or negative) may represent short-term reactions to specific incidents, but they rarely shift long-term perception unless such changes are sustained over time. Based on the rationale that brand perception typically changes gradually, as it reflects accumulated consumer experiences and collective sentiment rather than isolated events, we address the limitation of BP score by introducing the Brand Polarity Position Indicator (BPPI), a cumulative measure that aggregates daily polarity signals to produce a stable long-term trajectory. BPPI retains the responsiveness of the underlying BP Score while substantially reducing variance, thereby enabling clearer interpretation of medium to long-term sentiment trends, strategic inflection points, and competitive positioning.

The Brand Polarity Position Indicator (BPPI) illustrates the current positioning or progress of a brand relative to an initial reference point in time. This is a cumulative measure of the BP score calculated from an initial time and most reliably shows how the brand polarity changes at each time step with respect to that of the previous time step. By assessing the BPPI, analysts can discern whether the brand polarity exhibits an upward or downward trend, facilitating a thorough analysis of brand progress over time.

The Brand Polarity Position Indicator is given by,

2

The Brand Polarity Position Indicator () at time t is calculated as the ratio of the cumulative difference between total positive () and negative () tweets from the initial time to time, t, to the cumulative sum of positive and negative tweets during the same period. This indicator provides insight into the brand’s polarity evolution over time, where a positive value suggests a predominantly positive sentiment, and a negative value indicates a predominance of negative sentiment.

BPPI extends the principle of BP Score by aggregating daily polarity values over time, analogous to cumulative-sum (CUSUM) trajectories and long-horizon sentiment measures used in behavioural and brand-equity research. Unlike the daily BP Score, which reflects short-term fluctuations, BPPI highlights consistent improvement or decline rather than transient variations. This aligns with the psychological and behavioral nature of consumer sentiment, which evolves through repeated exposure and reinforcement. Therefore, BPPI serves as a meaningful measure of brand polarity position that reflects the stability and inertia inherent in public perception.

Figure 6 shows the BPPI calculated from the starting time of our study. We can see that the progress is better modelled using the BPPI.

Fig. 6.

Brand Polarity Position Indicator of chosen brands during the study period.

Brand polarity weighted with tweet influence

Through a comprehensive study of social media platforms, we recognize that the engagement level of each tweet significantly influences the perception it generates among users. A positive tweet with higher engagement has the potential to influence a larger audience’s perception. To capture this asymmetry in influence, the BPS score is refined to incorporate tweet-level engagement and author reach. The influence of a tweet can be evaluated based on its engagement metrics, as well as the follower reach of the author. The engagement metrics of a tweet include , , , , and . The author‘s follower reach is measured based on his/her . This adjustment ensures that the polarity metric reflects not only the sentiment distribution but also the differential visibility and diffusion potential of individual tweets.

Drawing from our investigation, we introduce a weighting system for each tweet based on two factors: Tweet Rank and Author Rank. The Tweet Rank is computed as follows:

3

where and TweetAge represent the time difference between the moment the tweet was extracted and its creation time, measured in minutes. This calculation yields the rate of tweet engagement, indicating the number of engagements per minute. Likewise, the Author rank is given by

4

where AuthorAge denotes the difference between the date the tweet was extracted and the author’s date of account creation, measured in days. This calculation provides the rate of follower growth for the author, representing the number of followers gained per day.

TweetRank captures the engagement a tweet has accumulated at the time of extraction, whereas AuthorRank represents the structural influence of the author based on follower reach and account maturity. Their magnitudes differ substantially across tweets and authors. To ensure that both components contribute comparably to the influence weight, we transform each into a normalized percentile rank within each day. After this transformation, and lie on a common [0,1] scale, where 0 denotes the lowest and 1 denotes the highest influence within that day. For simplicity of notation, we continue to use and to denote these normalized values in the influence-weight formulation that follows. Building on these normalized components, the combined influence weight for each tweet is defined as:

5

Here, represents the influence weight of tweet T, while and are the normalized TweetRank and AuthorRank, respectively. Tweet rank represents a tweet’s content-level visibility and is computed as the normalized average of key engagement metrics. Author rank represents the author’s credibility and reach, derived from the follower count standardized by the account’s age. Normalizing both components before aggregation ensures they lie on comparable scales.

Following the equal-weighting principle widely used in composite index construction, we assign the two components equal weight in the absence of strong empirical evidence that one dimension should dominate. This balanced formulation keeps the influence measure interpretable, avoids arbitrary weighting schemes, and ensures that incorporates both message-level and author-level impact in a principled manner.

The Brand Polarity Score weighted by Tweet Influence for a particular time slice is explained as follows: Let,

• a specific date in the analysis period,

• number of positive tweets on day d

• number of negative tweets on day d

• influence weight of the positive tweet on day d,

• influence weight of the negative tweet on day d.

then,

6

In this equation,

• represents the sum of influence weights for all positive tweets on day d and

• represents the sum over all negative tweets on day, d.

The distribution across the chosen brands is shown in Fig. 7. We considered extracting tweet metrics data and author metrics along with tweets only after 6th December. So the results shown in figure considers the initial time point as 7th December 2023.

Fig. 7.

Distribution of for the chosen brands.

Similarly, The cumulative Influence-Weighted Brand Polarity Position Indicator up to day D is defined as:

7

In this equation,

• the outer summation aggregates daily influence-weighted polarity scores from the start of the analysis period up to day D, where D denote the final day up to which cumulative polarity is computed.

• and denote the number of positive and negative tweets on day d respectively.

• and represent the influence weights of individual positive and negative tweets on day d.

In this formulation, the inner fraction computes the daily influence-weighted polarity (), while the outer summation accumulates these values across the time horizon from day 1 to day D to produce a smoothed, long-horizon signal.Because influence-weighted polarity moderates the volatility introduced by low-activity users, yields a trajectory that captures durable shifts in public sentiment amplified by influential voices, making it particularly suitable for managerial monitoring and campaign evaluation. The resulting distributions are visualized in Fig. 8.

Fig. 8.

distribution across the brands.

Ethics approval

This article does not contain any studies with human participants or animals performed by any of the authors.

Validation

In this research, we focused on analyzing tweets mentioning the top 5 automobile brands and developing a Brand Polarity Score based on the sentiments expressed in these tweets. Additionally, we introduced a novel variant called the Influence Weighted Brand Polarity Score, which takes into account both the sentiment of the tweet and its influence. Furthermore, we presented a variant of our analysis methodology called the Brand Polarity Position Indicator, designed for long-term progress tracking of brands.

This study adopts a multi-layered validation strategy to evaluate the reliability, responsiveness, and stability of the proposed polarity metrics—BPS, , BPPI, and . Our aim is not only to examine whether these indices capture meaningful sentiment variation but also to assess whether they produce stable estimates, converge with independent sentiment systems, and remain robust under noise or random perturbations. The following subsections describe the validation procedures and findings in detail, demonstrating that the polarity framework is both empirically grounded and methodologically sound.

Sentiment tool analysis

As the initial step, the extracted tweets required categorization into positive and negative sentiments. To accomplish this task, we employed machine learning-based pre-trained models specifically designed for sentiment analysis on tweet data, as well as those techniques intended for sentiment classification on general text. Through a thorough review of existing literature, we curated a selection of available pre-trained models for sentiment classification and conducted a comparative analysis of their performance on a benchmark dataset. Additionally, we leveraged the sentiment classification tool provided by AWS-Amazon Comprehend in conjunction with the selected models. The details of models and tools employed in the sentiment analysis are given in Table 2. This includes the sentiment classification by the TweetNLP package, BERTweet sentiment classification, and widely-used text classification tools such as Vader and TextBlob.

Before evaluating the models, we applied preprocessing steps appropriate to each sentiment classifier. For transformer-based models such as TweetNLP, BERTweet, and AWS Comprehend, mentions and URLs were replaced with placeholders to retain structure while reducing noise. For lexicon-based tools like VADER and TextBlob, we removed mentions/URLs entirely and converted emojis to sentiment labels so their rule-based scoring remained valid.

For benchmarking, all models were tested using a subset of 10,000 randomly selected tweets from the Sentiment140 dataset, which contains 1.6 million tweets labelled as 0 (negative) or 4 (positive). This subset had a balanced distribution of 5,039 positive and 4,961 negative samples. Since the dataset does not include ground truth for neutral sentiment, any tweets predicted as “neutral” or “mixed” by the models were excluded from accuracy calculations. This ensured that the evaluation reflected only valid comparisons between model predictions and the available binary ground truth.

Performance metrics for each model are summarized in the Table 4. The results indicate that transformer-based models such as TweetNLP and BERTweet, as well as Amazon Comprehend, outperform lexicon-based tools like VADER and TextBlob. TweetNLP, in particular, exhibited the most consistent and accurate sentiment predictions across the evaluation set. Consequently, TweetNLP was chosen for labelling the automobile-related tweets used in subsequent stages of our analysis.

Table 4.

Accuracy values of different sentiment classification tools on test data.

Sentiment Classification Tool	Accuracy
Tweetnlp - Robertabase	84.41
Amazon comprehend	82.81
Vader	71.19
TextBlob	67.75
BerTweet	83.17

Event sensitivity through permutation-based significance testing

A central requirement for any sentiment-based brand-perception measure is its ability to reflect genuine shifts in consumer attention when real events occur. To evaluate this property, we conducted a daily event-sensitivity analysis using permutation significance testing. For every brand and each date in the study period, we computed the observed day-to-day change in BPS and compared it against a null distribution generated via 5000 random permutations of tweet sentiment labels for that day. This procedure yields p-values indicating whether the polarity movement is statistically distinguishable from what would be expected under random fluctuation. The results reveal a clear pattern: sentiment spikes that coincide with observable external events tend to show unusually low p-values.

For Mercedes-Benz, we observed the sentiment increased sharply on 18 December following the announcement of conditional approval for Level-3 autonomous vehicle testing in Beijing. The BPS shift ( = + 1.245, p-right = 0.0314) indicates a statistically atypical rise, and the shows comparable evidence of a positive movement ( = + 1.244, p-right = 0.0426). Porsche experienced a modest positive shift on 21 December when news emerged regarding the introduction of a powerful hybrid engine for the 911 lineup. While the unweighted BPS provides only weak evidence of a non-random increase (p-two-tailed = 0.2326), the captures a more pronounced directional signal (p-right=0.0366), indicating that reactions with higher engagement or influential authorship drove the sentiment change. A significant sentiment spike also appears for Tesla on 14 December, coinciding with reports that more than two million vehicles may require an Autopilot software recall. Despite the negative nature of the event, both metrics detect an extreme shift (BPS = + 1.07; = + 1.15), likely reflecting polarized public discourse; the weighted score again accentuates the shift captured in the raw sentiment data.

For Toyota, the 26 December scandal involving safety-testing irregularities at Daihatsu produces a discernible negative shift. The unweighted BPS shows a decline that borders on statistical significance ( = – 0.778; p-left = 0.0646), whereas the influence-weighted version clearly identifies a significant negative deviation ( = – 0.834; p-left = 0.0402). This suggests that the downturn in sentiment was led disproportionately by tweets with higher visibility or authored by influential accounts—a pattern that the unweighted score does not fully capture.

Taken together, these results demonstrate that the proposed polarity measures are sensitive to real-world shocks and respond in interpretable ways to brand-relevant news events. Moreover, the influence-weighted variant consistently enhances the clarity and statistical strength of event-related signals, revealing sentiment movements that are either attenuated or missed entirely by the unweighted metric. This highlights the practical value of incorporating engagement and author influence in sentiment-based perception tracking, particularly for events where high-attention discourse drives public perception.

Convergent validity using cross-model sentiment agreement

To ensure that polarity outcomes are not tied to a single sentiment classification system, we examined the degree of alignment between TweetNLP-based scores and scores derived from AWS Comprehend, a widely adopted commercial sentiment analysis platform which also ranked next to TweetNLP in our sentiment tool analysis. For every tweet, we recomputed its sentiment using AWS Comprehend and recalculated corresponding polarity scores for each brand. We then computed Pearson and Spearman correlations between BP score computed using TweetNLP and AWS comprehend across all dates.

The resulting correlations ranged from moderate to very strong (Pearson r = 0.57–0.88; Spearman = 0.45–0.83), with p-values consistently below conventional significance thresholds. This level of agreement is notable given the architectural differences between the models and provides compelling evidence that the observed polarity trajectories are not artifacts of a specific classifier. Instead, the convergence indicates that the polarity signals arise from consistent sentiment patterns detectable by multiple independent NLP systems. This cross-method convergence establishes a foundation for construct validity and supports the robustness of BPS (and its derivative indices) as generalizable sentiment indicators. Table 5 presents the results of the same.

Table 5.

Cross-Model Correlation Matrix (TweetNLP vs AWS Comprehend).

Brand	Pearson r	p	Spearman	p
BMW	0.57	6.15E-04	0.45	1.06E-02
Mercedes-Benz	0.69	1.33E-05	0.48	5.95E-03
Porsche	0.79	6.21E-08	0.75	9.57E-07
Tesla	0.88	1.04E-10	0.83	9.31E-09
Toyota	0.80	3.08E-08	0.82	9.74E-09

Stability and noise-robustness assessment

To establish the reliability of the proposed polarity metrics beyond their point estimates, we conducted a series of stability and robustness analyses. These evaluations assess how sensitive the daily polarity signals are to sampling variability, temporal fluctuations, and controlled noise. The framework includes: (i) bootstrap confidence intervals to quantify statistical uncertainty in daily BPS values; (ii) LOESS smoothing to extract underlying sentiment trends; and (iii) noise-perturbation tests to examine robustness under synthetically induced label noise. Together, these analyses demonstrate that the proposed polarity measures remain stable, interpretable, and resilient to perturbations in the underlying data.

Uncertainty estimation with bootstrap confidence intervals

Because daily tweet volumes vary across dates and brands, we quantified the uncertainty in BPS estimates using non-parametric bootstrap sampling. For each day, we generated 1000 bootstrap resamples and computed the resulting distribution of BPS estimates. The 95% confidence intervals derived from these distributions provide a transparent measure of the reliability of each day’s estimate as shown in Fig. 9.

Fig. 9.

Daily Brand Polarity Score (BPS) with 95% bootstrap confidence intervals and LOESS-smoothed trajectories for each brand.

To further examine the behaviour of uncertainty under varying sentiment conditions, we analysed how the width of the bootstrap confidence interval (CI) correlates with core sentiment characteristics (Table 6). Across all five brands, CI width exhibited strong negative correlations with the absolute BPS value ( to ), indicating that days with clearly positive or negative sentiment produce substantially tighter CIs. CI width also decreased with greater imbalance between positive and negative tweets ( to ), and with larger daily tweet volume ( to ). These results reinforce the need for cumulative or smoothed polarity indicators such as BPPI and , which aggregate information across days and therefore exhibit substantially greater stability.

Table 6.

Correlation between daily BPS confidence interval width (CI) and sentiment characteristics for each brand.

Brand
BMW	– 0.912	– 0.970	– 0.847
Mercedes-Benz	– 0.587	– 0.891	– 0.758
Porsche	– 0.929	– 0.891	– 0.729
Tesla	– 0.706	– 0.865	– 0.925
Toyota	– 0.830	– 0.928	– 0.564

Trend extraction using LOESS smoothing

Although the daily BPS captures high-frequency sentiment dynamics, the raw signal can be noisy because social media conversations naturally fluctuate in volume, engagement, and topic relevance from one day to the next. To extract the underlying behavioural pattern without distorting the short-term responsiveness of the metric, we apply LOESS (Locally Estimated Scatterplot Smoothing)—a non-parametric regression technique that fits local polynomials across the timeline. LOESS is particularly suitable in this context because it does not assume a global functional form for sentiment evolution and adapts flexibly to irregularities in daily tweet distributions.

The resulting smoothed curves reveal medium-term sentiment trajectories that are obscured in the raw data: transient noise is attenuated, while persistent directional movements remain visible (Fig. 9). This clarified trend structure helps separate volatility driven by random day-to-day chatter from systematic changes linked to real events or campaign effects. The distinction between the high-resolution BPS signal and its smoothed representation reinforces the motivation for cumulative indices such as BPPI, which stabilize sentiment progression over longer horizons.

Robustness under noise-perturbation (“Flip-Noise”) tests

To further examine the resilience of polarity measurements, we conducted controlled noise-pollution experiments in which 5%, 10%, and 20% of daily sentiment labels were randomly flipped. We then recomputed polarity scores and measured correlations between the original and perturbed series. The results as presented in Table 7 show remarkable robustness for daily BPS: correlations remained extremely high at 5% noise and generally stayed above 0.70 even at 20% noise. In contrast, cumulative measures (BPPI) exhibited high stability under light noise but expectedly showed reduced correlation at high perturbation levels, reflecting the cumulative propagation of label flips across time. These observations align well with theoretical expectations: high-frequency polarity (BPS) is inherently resilient because random flips partially cancel each other out.

Table 7.

Noise–robustness of BPS and BPPI under 5%, 10%, and 20% random perturbations. Values show Pearson correlations between original and noise-perturbed series.

Brand	BPS_5%	BPS_10%	BPS_20%	BPPI_5%	BPPI_10%	BPPI_20%
BMW	0.910014	0.788106	0.695597	0.988078	0.739273	0.668873
Mercedes-Benz	0.986398	0.959732	0.859380	0.873138	0.565144	0.354813
Porsche	0.944876	0.811381	0.615165	0.995636	0.971167	0.918803
Tesla	0.978283	0.962771	0.819820	0.871674	0.822238	0.412840
Toyota	0.988324	0.964713	0.867932	0.995129	0.988182	0.975819

As expected for cumulative indices, BPPI remains highly stable under small perturbations (5% noise), but shows increasing divergence under larger artificial noise injections (10–20%). This behaviour is theoretically expected because cumulative curves aggregate errors from all previous days. Importantly, our core daily polarity measure (BPS) remains robust even under 20% perturbations, confirming that the foundational sentiment signal is stable.

Long-term stability of BPPI and comparison with alternative smoothing methods

To assess the temporal stability of the proposed perception indicators, we compared the volatility of five signals: the raw daily Brand Polarity Score (BPS), two standard smoothing baselines (LOESS-smoothed BPS and Exponentially Weighted Moving Average(EWMA), the cumulative Brand Polarity Position Indicator (BPPI), and its influence-weighted variant . While LOESS captures medium-term structural trends, EWMA serves as a complementary smoothing method that emphasizes recent observations through exponentially decaying weights. This makes EWMA particularly suitable for brand-perception tracking scenarios where more recent sentiment shifts should influence the indicator more strongly than older observations. By applying EWMA with a fixed span parameter(7), we obtain a principled smoothed signal that retains short-term adaptability while reducing high-frequency noise.

Volatility was quantified using the standard deviation of each series across all days in the observation window. The results are presented in Table 8. Across all brands, BPS exhibited the highest volatility (rho= 0.16–0.38), reflecting its sensitivity to day-to-day tweet composition. Smoothing methods substantially reduced this volatility, with LOESS and EWMA achieving 60–85% reductions confirming that daily polarity signals become more interpretable after smoothing.

Table 8.

Volatility of brand polarity measures (standard deviation over the observation period).

Brand	BPS	LOESS(BPS)	EWMA(BPS)	BPPI	IwBPPI
BMW	0.1570	0.0543	0.0593	0.0432	0.0533
Mercedes-Benz	0.2953	0.0544	0.0848	0.0328	0.0652
Porsche	0.1656	0.0484	0.0779	0.0639	0.0372
Tesla	0.3154	0.0808	0.0771	0.0333	0.0763
Toyota	0.3824	0.0625	0.1484	0.1096	0.0361

The BPPI demonstrated the strongest stability with extremely low volatility (rho = 0.03–0.11). This behaviour is expected because BPPI is constructed through cumulative aggregation, acting as a long-horizon sentiment integrator. Thus, BPPI is well suited for tracking multi-week brand-perception trajectories and identifying persistent directional trends rather than transient spikes. demonstrated similarly low variance, indicating that incorporating influence weights does not amplify noise. This pattern highlights its hybrid role: it preserves long-term stability while remaining sensitive to perception shifts initiated by influential users or viral content.

Collectively, these results clarify the functional distinction between daily polarity measures (BPS), which capture short-term sentiment bursts, and cumulative perception indicators (), which capture durable shifts in brand perception. We recommend BPPI for long-term monitoring and for contexts where influential-user dynamics are relevant.

Practical implications

The cumulative polarity measures introduced in this study—BPPI and its influence-weighted counterpart—offer brand managers a structured and reliable way to track how consumer perceptions evolve over extended periods. Because brand attitudes typically shift gradually and respond to sustained communication, customer experience, and product performance, cumulative polarity scores provide a stable and strategically meaningful representation of sentiment trends. BPPI smooths daily noise and highlights enduring shifts in consumer favourability, while the influence-weighted variant further reveals whether sentiment is driven by highly visible, credible, or authoritative social-media users. These long-horizon insights help managers evaluate the effectiveness of brand-building activities, monitor changes in competitive positioning, and allocate resources toward initiatives that contribute to sustained improvements in public perception.

The inclusion of influence weighting enhances the practical utility of the cumulative polarity metrics by distinguishing between sentiment driven by general consumers and sentiment driven by high-impact voices. Understanding whether shifts in perception originate from prominent content creators, industry commentators, or influential communities allows firms to tailor communication strategies, manage influencer relationships, and design more targeted engagement campaigns. Influence-weighted BPPI, therefore, provides a richer basis for strategic planning—highlighting which audiences shape brand narratives and how their contributions evolve over time. This perspective is especially valuable in competitive markets where a small number of influential users can disproportionately drive narrative momentum.

While long-term indicators are foundational for brand strategy, the daily BPS offers a complementary short-term lens into immediate consumer reactions. BPS can help organisations detect early signals of emerging issues, gauge public response to new campaigns or announcements, and monitor shifts in conversation tone during periods of heightened visibility. Rather than prompting reactive day-to-day decision-making, daily polarity acts as an operational monitoring tool that supports preparedness, rapid situational awareness, and timely communication adjustments when necessary. Together, BPPI, , and BPS provide a multi-layered analytical framework that integrates both short-term sentiment dynamics and long-term brand perception—offering managers a comprehensive view of how their brand is experienced, discussed, and evaluated across social media.

Beyond quantitative tracking, BPS and its variants can support practical root-cause analysis (RCA) by helping managers understand why sentiment shifts occur. By combining daily polarity changes with simple qualitative tools such as word clouds, managers can quickly identify the themes driving spikes or drops in sentiment—even when the cause is unrelated to the brand’s products or services. For instance, negative polarity dip for Toyota on December 3rd was traced to unrelated event that a sports team sponsored by the brand losing a game (Fig. 10). However on Dec 26th world cloud clearly depics the sentiment outrage caused by Daihatsu’s safety-testing scandal. Similarly while extreme values for Mercedes-Benz corresponded to major announcements or viral user outrage. These examples illustrate how integrating BP score trends with contextual tweet content enables managers to rapidly diagnose sentiment movements, distinguish genuine brand issues from external noise, and respond more effectively.

Fig. 10.

a Word cloud depicting negative tweets related to Toyota on December 3rd, b Word cloud depicting negative tweets related to Toyota on December 26th, c Word cloud depicting positive tweets related to Mercedes Benz on December 18th, d Word cloud depicting negative tweets related to Mercedes Benz on December 30th.

Limitations and conclusion

This research harnessed a deep-learning–based sentiment analysis model to formulate a Brand Polarity Score for the top five automobile brands. By extracting tweets mentioning these brands and analyzing their sentiment, we introduced several variants of the polarity score that offer increasingly nuanced insights into consumer perception. The BPS provides a high-resolution view of daily fluctuations in expressed sentiment, while the BPPI captures the cumulative trajectory of polarity over time, offering a stable indicator of longer-term brand perception trends. Furthermore, incorporating tweet-level influence through the and metrics allows the framework to account for differential impact across users and posts, thereby providing a more behaviourally meaningful measure of brand polarity.

This study is distinctive in extending polarity estimation beyond traditional sentiment ratios by integrating influence-weighted components and longitudinal tracking, supported by a multi-layered validation framework. To our knowledge, prior work has not combined deep-learning sentiment models, influence normalization, uncertainty quantification, and temporal aggregation within a unified polarity-measurement system.

Nevertheless, several limitations must be acknowledged. The dataset represents only a subset of daily brand mentions and was collected using a non-probability, quota-based sampling strategy imposed by Twitter API rate limits. While this design ensures balanced and comparable exposure across brands, it may under-capture intra-day variability for high-volume brands. Additionally, the one-month collection window—constrained by API access—limits the study’s ability to infer long-horizon brand-perception dynamics. Our intention, however, is not to claim full reconstruction of brand equity, but to demonstrate a generalisable methodological framework for extracting, weighting, and aggregating polarity signals from social media data. Importantly, the cumulative indicators (BPPI and ) are explicitly designed to mitigate daily volatility and approximate gradual sentiment evolution; applying the framework to multi-month or multi-year datasets would further enhance stability and interpretability.

Future work may involve fine-tuning sentiment models using domain-specific tweet corpora, which could improve performance in brand-related contexts. Additionally, incorporating data from multiple platforms (e.g., Reddit, YouTube, Instagram) would allow broader external validation of the proposed metrics. Predictive modelling of future polarity based on brand-level or event-level covariates also represents a promising direction, enabling organisations to identify the drivers of polarity shifts and respond proactively to emerging sentiment trends.

Author contributions

S.S.M contributed to conceptualizing the study, devising the methodology, and securing funding for the project. K.H contributed to the overall project supervision with closely reviewing the experiments and write up. N.V actively conducted experiments, analyzed outcomes, contributed to the write up, and prepared visualizations. M.E.B provided critical insights during manuscript reviews and suggested enhancements.

Funding

This research work was funded by Zayed University, UAE, under grant number R23016.

Data availability

Data availability: All data included in this manuscript are available upon request by contacting with the corresponding author. Code Availability: All scripts used for data preprocessing, sentiment modelling, and computation of polarity metrics are available at our public code repository: https://github.com/NeethuVenugopal/Brand-Perception.

Declarations

Competing interests

The authors declare no competing interests.