The mediating role of brand equity in corporate social responsibility and customer purchase intention based on text-CNN model

Abstract

In the digital era, corporate social responsibility (CSR) has become a crucial factor influencing consumer purchase behavior. To investigate the mechanism through which CSR affects customer purchase intention (CPI) and to examine the mediating role of Brand Equity (BE), this study integrates deep learning techniques with social identity theory to construct a comprehensive research framework supported by multi-source data. A two-stage approach is employed. In the first stage, an optimized Text-Convolutional Neural Network model is applied to perform sentiment analysis on consumer reviews from the publicly available Yelp dataset, encompassing sectors such as food services, hospitality, and beauty. This model effectively captures consumer sentiment toward CSR-related initiatives. In the second stage, a structured questionnaire is developed based on social identity theory to quantitatively examine the relationships among CSR, BE, and CPI. The empirical results reveal that CSR exerts a significant positive influence on CPI and that BE plays a partial mediating role in this relationship. These findings suggest that CSR activities enhance purchase intention by fostering stronger brand identification. Notably, CSR efforts related to environmental protection and public welfare demonstrate a particularly strong influence on consumer behavior. Theoretically, this study extends the application of social identity theory in the CSR domain by introducing an integrated framework that combines sentiment recognition with structural equation modeling. Methodologically, the study advances a cross-disciplinary approach by merging natural language processing with survey-based research to facilitate the deep integration of multi-source data. Practically, this study provides data-driven evidence and strategic insights for enterprises aiming to design effective CSR strategies and enhance brand management, offering valuable implications for managerial decision-making.

Introduction

Amid accelerating digital transformation and rising awareness of sustainable consumption, corporate social responsibility (CSR) has evolved from a traditional ethical obligation into a strategic instrument that significantly influences consumer behavior and corporate Brand Equity (BE). Increasingly, consumers are attuned to companies’ social performance in domains such as environmental protection, public welfare, and employee rights, and they adjust their purchasing preferences accordingly. Long et al. (2023) showed that young consumers were more inclined to support brands that demonstrate social responsibility. This trend compelled enterprises to engage in socially responsible practices and also to effectively communicate these efforts to consumers, thereby fostering deeper emotional connections in the market¹.

Although the influence of CSR on consumer behavior has become a prominent topic in marketing and organizational behavior research, several limitations persist in existing studies. Most existing studies have relied on structured questionnaires and experimental designs emphasizing rational attitude formation^2–5. These approaches often highlight consumers’ judgment processes while overlooking their natural emotional responses in real contexts. With the rapid growth of unstructured data such as online reviews and social media texts, a pressing challenge is how to capture consumers’ genuine psychological and emotional reactions to CSR through innovative methods^6–8. Prior research has confirmed that CSR influences customer purchase intention (CPI) through multiple dimensions⁹. Lin et al. (2011) showed that purchase intention was indirectly affected by corporate ability, negative publicity, and CSR via trust and emotional identification¹⁰. Gatti et al. (2012) emphasized that CSR and corporate reputation enhanced perceived product quality and positively affect purchase intention, implying a potential mediating role of brand-related assets¹¹. Zhang and Ahmad (2021) demonstrated that CSR awareness moderated the CSR–purchase intention relationship, supporting the role of brand cognition and emotion in this process¹². Al-Haddad et al. (2022) further found that, with the rise of social media, CSR in ethical and environmental domains significantly boosted customer engagement and purchase intention¹³.

Despite these insights, the mediating role of BE has not been systematically examined. Within a data-driven research framework, it remains unclear how CSR shapes consumers’ brand cognition and emotions and how these factors translate into purchase intention. BE, as a key intangible asset, reflects brand image, reputation, and recognition in consumers’ minds^14,15. Theoretically, CSR enhances reputation and brand image, fosters trust and emotional identification, and thereby influences purchase intention^16,17. Incorporating BE into the CSR–CPI model is therefore well justified: CSR activities elevate BE, which in turn generates positive cognition and emotional resonance, indirectly stimulating purchase intention. This mechanism clarifies CSR’s influence pathway and highlights BE’s strategic role in CSR practices.

Despite the well-established importance of emotional factors in shaping consumer decisions, there remains a lack of systematic research exploring how emotional signals—particularly those derived from textual analysis—intervene in the mechanism by which CSR influences CPI. This research gap underscores the need for an integrated framework that combines sentiment recognition with theoretical modeling. In recent years, with the rapid development of deep learning, particularly in the field of natural language processing (NLP), sentiment recognition has achieved significant breakthroughs. NLP has been widely applied in areas such as machine translation, spam detection, information extraction, text summarization, medical analysis, and question–answering systems¹⁸. Prior research has shown that NLP can effectively process large-scale text data and extract sentiment features from diverse sources such as consumer online reviews and social media posts, providing reliable tools for understanding consumer attitudes^19,20. At the corporate sustainability level, NLP enables the automated extraction and analysis of large volumes of enterprise text data, where topic identification and sentiment analysis reveal the priorities and trends of corporate behavior and CSR practices^21,22. Deep learning models have demonstrated exceptional capabilities in analyzing consumer-generated content with high precision, enabling the capture of subtle emotions and latent attitudes that are often inaccessible through traditional survey methods. Compared with conventional techniques, deep learning enhances the accuracy and granularity of sentiment analysis, providing deeper insight into consumers’ authentic emotional responses and behavioral motivations. This, in turn, supports a more comprehensive and dynamic understanding of how CSR activities influence consumer behavior²³. The Text-Convolutional Neural Network (CNN) model, due to its strong performance in feature extraction and text classification, has become one of the mainstream deep learning approaches for sentiment analysis. It is now widely applied in tasks such as text classification, sentiment analysis, public opinion monitoring, question–answering systems, and information retrieval. However, its application in corporate research remains limited, with most studies focusing on enterprise management^24–27. Existing sentiment recognition models generally emphasize word-level features while overlooking the value of character-level information in capturing semantic nuances and subtle emotional variations, thereby constraining the effectiveness of fine-grained sentiment analysis. Although some studies have employed deep learning models to analyze consumer reviews, they have largely concentrated on improving classification accuracy and have rarely attempted to systematically embed the recognition results into CSR research frameworks. Furthermore, research platforms tend to be relatively homogeneous, often neglecting the dynamic reflection capabilities of multi-source data such as e-commerce reviews and social media texts, which limits the generalizability of the research findings²⁸.

To address these gaps, this study introduces methodological and theoretical innovations. From a methodological perspective, the study leverages recent advances in NLP by employing an optimized Text-CNN model to perform fine-grained sentiment analysis on consumer-generated reviews sourced from platforms such as Yelp. Compared with traditional survey-based data, deep learning methods provide a more accurate means of identifying consumers’ implicit attitudes and semantic distinctions related to CSR^29–31. To overcome the limitations of word-level models, character-level embeddings are incorporated to enhance the detection of emotional nuance and semantic complexity³². By integrating unstructured review data with structured questionnaire responses, the study captures quantifiable indicators related to CSR, BE, and CPI. This multi-source data approach improves the reliability and validity of the variables used, enriches understanding of the CSR influence mechanism, and provides a robust empirical foundation for subsequent structural equation modeling (SEM), thereby enhancing the overall rigor and practical relevance of the analysis.

In terms of theoretical construction, this study adopts social identity theory proposed by Tajfel and Turner as the foundational framework³³. This theory emphasizes the psychological mechanisms through which individuals seek affiliation and value alignment within social groups, which closely parallels the ways in which CSR initiatives communicate corporate values and foster consumer identification. Compared with stakeholder theory³⁴ or social moral identity theory³⁵, social identity theory offers a more nuanced explanation of how socially responsible behavior by corporations fosters group identification and, subsequently, purchase intention. It thus provides a solid theoretical basis for understanding the role of BE as a mediator in the CSR–CPI relationship.

Accordingly, this study seeks to explore how CSR influences CPI through the mediating role of BE, using emotional signals extracted from real-time consumer text. Two key research questions guide this investigation:

1. How do consumers emotionally respond to CSR initiatives as reflected in authentic online reviews?

2. How does CSR influence CPI through its impact on BE?

Based on these inquiries, the study makes several contributions. Theoretically, it introduces social identity theory to explain consumer-brand identification and purchase behavior in response to CSR, addressing the lack of psychological mechanism analysis in prior CSR research. In terms of knowledge advancement, it proposes an interdisciplinary framework that integrates deep learning–based sentiment recognition with SEM, expanding methodological approaches in CSR studies. Methodologically, it enhances the granularity and precision of sentiment analysis through the optimization of the Text-CNN model with character-level embeddings. Practically, it identifies the pathway through which CSR activities foster brand identification and thereby influence purchase intention, providing a data-driven foundation for evaluating CSR communication effectiveness. From a managerial perspective, the findings suggest a strategic integration of CSR and brand management, leveraging emotion-driven data insights to better align social value with business performance.

The structure of the remainder of the study is as follows: Sect. 2 presents the literature review. Section 3 outlines the research methodology. Section 4 details the data sources, empirical results, and discussion. Section 5 concludes with the main findings and suggestions for future research.

Literature review

Research progress on the relationship between CSR and consumer behavior

CSR refers to the proactive assumption of responsibilities toward the environment, employees, communities, and society as a whole while pursuing economic benefits³⁶. The core feature of CSR lies in balancing shareholder interests with broader social interests. In this study, CSR is divided into five dimensions: environmental responsibility, employee rights protection, consumer rights protection, philanthropic responsibility, and economic responsibility. Together, these dimensions constitute a comprehensive framework of CSR³⁷. With growing global attention on CSR, research examining its impact on CPI has steadily expanded³⁸. Barlas et al. (2023), in a questionnaire-based study of a Greek mobile telecommunications firm, confirmed that CSR significantly contributed to improved financial performance and enhanced consumer trust. The study emphasized that successful companies—both in theoretical models and practical operations—not only prioritized sales and profitability but also focused on enhancing their social impact by giving back to society and generating broader social value through corporate initiatives³⁹. Nguyen-Viet et al. (2024) found that companies adopting socially responsible practices strengthened their BE, which in turn influenced consumer purchase decisions. Their findings highlighted a preference among consumers for socially responsible brands during the brand selection process⁴⁰. Similarly, Cuesta-Valiño et al. (2024) demonstrated that CSR activities enhanced corporate image, thereby increasing purchase intention and strengthening customer loyalty⁴¹. Ciu and Wijayanti (2024) reported that high BE enabled firms to occupy advantageous market positions, while simultaneously promoting purchasing behavior through enhanced emotional identification with the brand⁴². In a study focused on Generation Y consumers in Ghana, Amankona et al. (2024) examined the effects of digital CSR on purchase intention and identified brand loyalty and customer engagement as mediating variables. Their results indicated that CSR initiatives significantly boosted purchase intention by fostering stronger brand loyalty and increasing consumer participation⁴³.

Although these studies collectively affirm the link between CSR and consumer behavior, most have relied on structured questionnaires or experimental designs that emphasize cognitive constructs such as brand attitude and customer satisfaction within controlled models. However, these approaches often neglect the affective dimension of consumer responses, failing to fully capture the emotional impact of CSR as expressed in authentic, real-world contexts.

CSR’s influence pathways on BE

BE refers to the overall value assessment formed by consumers based on brand associations⁴⁴. Scholars generally recognize that BE comprises multiple dimensions, including brand awareness, brand loyalty, brand trust, and brand attachment. These dimensions reflect consumers’ rational evaluations of a brand as well as their emotional identification and attachment, which directly or indirectly influence purchase intention⁴⁵. In the field of brand management, the positive effect of CSR on BE has been well established. Golob et al. (2008) argued that CSR initiatives were often perceived by consumers as reflective of shared moral and social values, thereby reinforcing brand trust and emotional attachment⁴⁶. Gálvez-Sánchez et al. (2024) similarly concluded that CSR practices enhanced both the cognitive and emotional dimensions of BE, which subsequently influenced consumer purchase intention⁴⁷. However, much of the extant literature treats the relationship between CSR and BE as a direct and linear association, neglecting the mediating function of BE in the broader pathway linking CSR to consumer behavior. This oversight results in an incomplete understanding of the CSR–BE–CPI mechanism and limits insight into the underlying psychological processes that drive consumer decision-making in socially responsible contexts.

The application potential of sentiment analysis techniques in CSR research

Sentiment analysis, also known as opinion mining, is a key research direction in NLP. It aims to automatically identify, extract, and quantify individuals’ subjective emotions and attitudes from unstructured text⁴⁸. The fundamental process generally involves three stages. First, text preprocessing converts raw corpora into a computable form through operations such as tokenization, stop-word removal, and part-of-speech tagging. Second, feature representation transforms textual content into numerical form using methods such as bag-of-words models or deep learning–based word embeddings. Finally, classification and modeling employ machine learning or deep learning approaches to recognize and predict sentiment polarity (positive, negative, neutral) and its intensity^49,50. Compared with traditional questionnaire-based attitude measurement, sentiment analysis can more directly capture consumers’ natural emotional responses to corporate behavior in authentic contexts, thereby providing richer and more dynamic data support for CSR research⁵¹.

Recent developments in affective computing and NLP have opened new avenues for investigating the mechanisms through which CSR affects consumer behavior⁵². Giacomini et al. (2021) applied sentiment analysis techniques to study public responses on social media to CSR initiatives undertaken by Italian companies during the early stages of the COVID-19 pandemic. Their findings revealed the role of external environmental factors in shaping CSR strategies⁵³. Fan et al. (2024) employed a Text-CNN model for sentiment analysis of consumer reviews, demonstrating its effectiveness in capturing fine-grained emotional cues and improving classification accuracy⁵⁴. Further advancing this field, Xing et al. (2025) enhanced granular sentiment recognition by integrating the Text-CNN model with multi-head encoder-sharing mechanisms, particularly excelling in the analysis of complex emotional expressions within reviews⁵⁵. Despite these technological advancements, the application of sentiment analysis in CSR-related behavioral research remains limited. Most studies to date have concentrated on improving the technical accuracy of sentiment classification, with minimal emphasis on how emotional responses—particularly those articulated in consumer-generated text—can be systematically extracted and incorporated into CSR research frameworks. As a result, the affective dimension of consumer reactions to CSR has yet to be meaningfully embedded in empirical models, leaving a critical gap in both theoretical explanation and practical application.

How emotion recognition supports CSR mechanism studies

Emotion functions as a pivotal mediator in the relationship between CSR and consumer behavior, yet its role in evaluating the effectiveness of CSR communication remains insufficiently explored. Kaur and Sharma (2023) employed Long Short-Term Memory (LSTM) models to extract sentiment orientations from unstructured consumer texts. Their study demonstrated that such methods not only captured authentic emotional feedback on corporate activities but also enabled firms to better understand consumer behavior and preferences, thereby offering valuable insights for strategic decision-making⁵⁶.

Jang and Kang (2023) utilized topic modeling and other NLP techniques to analyze open-ended responses from 703 U.S. consumers regarding acquisition, usage, and disposal behaviors in the fashion industry. Their analysis focused on emotional expressions related to consumer social responsibility and sustainable practices. By applying deep learning to extract sentiments and perceptions of responsibility across various stages of consumption, the study highlighted the crucial role of emotion recognition in revealing the micro-level mechanisms underlying individual CSR-related behaviors. This research not only enriched the psychological dimensions of CSR studies but also expanded the methodological toolkit available for investigating consumer responsibility⁵⁷.

Similarly, Galiano-Coronil et al. (2024) applied data mining and sentiment analysis techniques to assess CSR-related content posted by Spain’s ten most socially responsible companies on Twitter. Their findings revealed that the emotional tone embedded in CSR communications significantly shaped public engagement and response. The study suggested that sentiment recognition—particularly when powered by deep learning—played a critical role in identifying effective patterns in CSR messaging, thereby providing empirical evidence to guide the optimization of corporate social marketing strategies⁵⁸.

Further analysis of the relationship between social identity theory and CSR indicates that CSR initiatives can enhance consumers’ group identification, thereby improving their brand attitudes and behavioral intentions^59,60. At the same time, advances in sentiment analysis have extended its application to cross-cultural emotion recognition, brand sentiment monitoring, and consumer psychology prediction^61,62. For instance, Kayakuş et al. (2024) examined user reviews of the iPhone 11 on Trendyol, the largest e-commerce platform in Türkiye. They also employed sentiment analysis combined with machine learning methods to investigate the relationship between brand reputation, consumer trust, and loyalty. Their findings showed that sentiment analysis could accurately identify positive, negative, and neutral emotions, thereby revealing how brand reputation shaped consumer trust and loyalty. This demonstrates that sentiment analysis was an effective tool in CSR research for capturing consumers’ emotional responses to corporate behavior and brand practices. It provided valuable data support and decision-making insights for optimizing CSR strategies, enhanced brand reputation, and strengthened customer loyalty⁶³. Similarly, Anderson et al. (2024) utilized social media data and applied machine learning and deep learning techniques to conduct sentiment analysis that quantified public attitudes toward environmental sustainability initiatives. The results indicated that sentiment analysis effectively uncovered consumers’ or the public’s genuine emotions, thereby supplying CSR research with reliable data that supported the improvement of CSR strategies and strengthened the effectiveness of social responsibility practices⁶⁴. More recently, Belarmino et al. (2025) found that sentiment analysis could also reveal consumers’ negative emotions toward CSR activities. Their study further showed that applying emotional and cognitive information frameworks could mitigate negative perceptions, thereby influencing consumer attitudes toward CSR⁶⁵.

In this study, emotions are conceptualized as consumers’ immediate psychological responses triggered by the perception of CSR activities, representing an attitudinal tendency elicited by external stimuli. Conceptually, emotions differ from BE; the former reflects short-term, dynamic affective experiences, whereas the latter represents a comprehensive evaluation of the brand formed through long-term consumer interactions. Although emotions are not a direct component of BE, they play a crucial role within the CSR–BE–CPI pathway. On one hand, positive emotional responses can enhance consumers’ favorable perceptions of the brand, thereby promoting the development of BE. On the other hand, emotions can indirectly influence consumers’ purchase intentions through affective identification⁶⁶. In this study, emotions function both as a key output of the sentiment analysis, reflecting consumers’ authentic feedback from online reviews, and as a precursor in the structural equation model, influencing BE to clarify their role in the CSR–CPI relationship. The inclusion of emotions thus enriches the psychological mechanism perspective in CSR research while providing a more dynamic and data-driven explanation of the mediating effect of BE.

Research gaps and innovation positioning of this study

Most existing studies still rely on surveys and interviews, which are limited by sample representativeness and social desirability bias and also struggle to capture consumers’ authentic attitudes and emotional responses toward CSR. Prior research often remains at a macro-level linear inference, failing to deeply uncover the psychological mechanisms—such as emotional identification and value resonance—through which CSR influences consumer decision-making. To bridge these gaps, this study proposes an integrated analytical framework that combines deep learning–based sentiment recognition with the prediction of purchase behavior. By leveraging large-scale online review data, the study captures consumers’ emotional responses to CSR at a micro level and investigates the potential pathways through which these emotional cues influence consumption decisions. This approach addresses the limitations of traditional CSR studies by enhancing both the richness of data and the technological sophistication of analysis. It also provides actionable, data-driven insights and robust empirical evidence to support the optimization of CSR strategies and digital marketing practices.

A summary and critical evaluation of representative CSR-related studies is presented in Table 1.

Table 1.

Summary and evaluation of prior CSR research.

Category	Representative studies	Methodology	Key findings	Limitations
Impact of CSR on Consumer Behavior	Barlas et al.39; Nguyen-Viet et al.40; Cuesta-Valiño et al.41; Ciu and Wijayanti42; Amankona et al.43	Questionnaire surveys, SEM	CSR enhances brand trust and purchase intention, with brand loyalty acting as a mediating variable	Heavy reliance on subjective surveys; limited analysis of authentic emotional responses in natural consumer contexts
CSR’s Influence on BE Pathways	Golob et al.46; Gálvez-Sánchez et al.47	Conceptual models, empirical investigation	CSR strengthens emotional attachment and brand trust, thereby increasing BE	Limited exploration of BE as a mediating factor; psychological mechanisms remain underdeveloped
Applications of Sentiment Analysis	Giacomini et al.53; Fan et al.54; Xing et al.55	Text-CNN, Sentiment Classification	NLP models effectively capture fine-grained emotions and improve sentiment recognition accuracy in CSR contexts	Focus primarily on emotion classification accuracy; limited integration of sentiment results into behavioral analysis
Emotion-Driven Mechanisms in CSR Research	Kaur and Sharma56; Jang and Kang57; Galiano-Coronil et al.58.	LSTM, Topic Modeling, Sentiment Mining	Deep learning uncovers how emotional responses shape perceptions of CSR and influence behavioral outcomes	Studies often limited to single platforms or specific contexts; lack of generalizable and integrative modeling frameworks

Table 1 summarizes and evaluates the main categories of recent CSR-related studies, including representative literature, research methods, key findings, and limitations. The studies were selected based on systematic retrieval and screening, prioritizing those with strong academic influence, close relevance to CSR, and methodologically representative approaches. A comparative analysis reveals that, although existing research reflects the mainstream directions in CSR studies, notable gaps remain in methodological approaches and the exploration of psychological mechanisms. In particular, the limitations in exploring emotional mechanisms and integrating multi-source data are evident. This highlights a theoretical gap and providing a clear entry point for the present study’s framework, which combines sentiment recognition with the prediction of consumer purchase behavior.

Research model

Construction of the sentiment analysis model of CPI based on Text-CNN

This study employs a deep learning–based Text-CNN model for textual sentiment analysis. Text-CNN is a high-performing CNN architecture widely used in NLP tasks, which extracts local text features via multi-scale convolution kernels and performs sentiment classification through max-pooling operations⁶⁷. The basic architecture of the Text-CNN model is illustrated in Fig. 1⁶⁸.

Fig. 1.

The Text-CNN model structure.

In the Text-CNN model structure shown in Fig. 1, different colors are used to distinguish the convolutional kernel branches and their corresponding feature processing flows. The original input data on the left (orange block) is divided into three parallel processing branches, marked in blue, yellow, and green, each corresponding to a convolutional kernel of different sizes to extract local features of varying lengths within the text. Within each colored branch, the large colored blocks represent the outputs of the convolution operations, while the subsequent smaller blocks indicate the results of the pooling operations. On the far right, the multicolored small blocks represent the concatenated features from the three branches after pooling, integrating multi-scale features extracted by different convolutional kernels.

To enhance the model’s ability to capture fine-grained emotional characteristics, an improved two-dimensional (2D-) Text-CNN model is developed by integrating both character-level and word-level features, thereby extending the traditional Text-CNN framework⁶⁹. Figure 2 depicts the structure of this optimized 2D-Text-CNN model.

Fig. 2.

Architecture of the 2D-Text-CNN model.

The 2D-Text-CNN model mainly comprises an embedding layer, a 2D convolutional layer, a pooling layer, a fully connected layer, and an output layer. Initially, the original text is encoded into two input channels corresponding to character-level and word-level embeddings. The embedding matrices and are generated, where and denote the number of characters and words, respectively, and and represent their respective embedding dimensions.

Each channel then extracts local features using distinct 2D convolution kernels . The convolution operation is defined as Eq. (1)⁷⁰:

1

Here, denotes the activation function; is the bias term; represents a character or word channel.

The convolution outputs undergo a max-pooling operation to extract the most salient local emotional features, calculated as Eq. (2):

2

Subsequently, the pooled features from both character and word channels are concatenated into a unified feature vector:

3

This combined vector passes through a fully connected layer, and the predicted probability distribution over emotion categories is obtained via a Softmax activation function:

4

In Eq. (2), is the activation function; is the weight matrix of the classification layer; is the corresponding bias vector.

Given a training sample size , the true labels (one-hot encoded), and the predicted probabilities , the cross-entropy loss is defined as Eq. (5):

5

Empirical research on CPI

Building upon the sentiment recognition model, this study constructs a SEM to investigate the mechanisms through which CSR influences consumer purchase intention, with a particular focus on the mediating role of BE. SEM is well-suited for analyzing complex relationships among variables, as it allows for the simultaneous estimation of multiple causal paths and latent constructs while maintaining strong theoretical alignment and statistical rigor.

The theoretical model is grounded in social identity theory, which posits that individuals derive part of their self-concept from membership in social groups. When individuals identify with a social entity—such as a corporation—and align with its behaviors and values, they develop a sense of belonging and collective identity. This psychological identification can significantly shape consumer attitudes and behaviors. social identity theory emphasizes that identification with social categories not only affects interpersonal dynamics but also plays a critical role in consumer behavior, particularly in fostering brand affiliation and loyalty. However, the specific cognitive mechanisms through which social identity shapes consumer responses to CSR activities—and how these responses influence purchase intentions—remain insufficiently understood. Factors such as the depth and persistence of identification, as well as the interaction of multiple social identities, may vary across cultural and market contexts. Empirical research on these contextual variations is still limited. Moreover, discussions on negative social identity—such as consumer disassociation from brands perceived as socially irresponsible—have received inadequate attention in current CSR literature^71,72. In this study, the proposed model draws on the central principles of social identity theory, particularly the link between psychological identification and behavioral motivation, while acknowledging the need for deeper theoretical refinement in the CSR domain. This theoretical foundation supports the empirical analysis and offers valuable insights for future research directions.

The proposed research model is illustrated in Fig. 3.

Fig. 3.

Theoretical model framework.

The variables and their operational definitions are presented in Fig. 4. CSR is measured through consumer evaluations of CSR practices, including environmental protection, employee welfare, and community development. These evaluations are captured via questionnaire items addressing corporate public welfare efforts, environmental policies, and related initiatives⁷³. BE reflects consumers’ positive emotional evaluations and trust in a brand and is assessed through dimensions such as brand awareness, loyalty, and satisfaction⁷⁴. High levels of BE are associated with stronger purchase intentions and represent a strategic asset in brand management. Finally, CPI refers to consumers’ willingness to make repeat purchases and recommend the brand to others, serving as a key indicator of marketing effectiveness⁷⁵.

Fig. 4.

Definitions of research variables.

In this study, BE is explicitly defined as a mediating variable that bridges CSR and CPI. Specifically, BE captures consumers’ perceptions and evaluations of a company’s CSR practices and functions as a key psychological mechanism through which CSR shapes purchasing behavior. By enhancing BE, companies can reinforce their socially responsible image, foster consumer trust, and cultivate favorable brand attitudes, thereby promoting purchase intention. As such, BE is not positioned as a moderating variable but rather as a pivotal mediator that transmits and amplifies the positive influence of CSR on consumer behavior.

The theoretical framework of this study is grounded in social identity theory. CSR activities enhance consumers’ social identification with a brand by strengthening BE, including trust, loyalty, and attachment, thereby promoting CPI. The CSR → BE → CPI pathway reflects the sequence of “group identification → attitude formation → behavioral intention.”

According to social identity theory, individuals are inclined to align themselves with organizations that reflect their values and social ideals⁷⁶. When companies actively engage in CSR activities, consumers tend to integrate these firms into their psychological “in-groups,” which fosters stronger identification, trust, and affiliation⁷⁷. Empirical evidence supports that CSR initiatives significantly enhance purchase intention and brand preference⁷⁸, particularly when companies demonstrate strong performance in areas such as environmental protection and public welfare, which are more likely to attract consumer support⁷⁹. Based on this, the study proposes hypothesis 1:

Hypothesis 1

CSR has a positive and significant impact on CPI.

BE reflects consumers’ cognitive assessments, emotional attachment, and trust in a brand, forming a fundamental component of BE. According to BE theory, elevated BE enhances brand loyalty, increases purchase frequency, and raises consumers’ willingness to pay a premium⁸⁰. Studies have shown that higher BE correlates with greater likelihood of sustained purchase behavior⁸¹ and more frequent engagement in positive word-of-mouth communication⁸². Based on this, the study proposes hypothesis 2:

Hypothesis 2

BE positively and significantly influences CPI.

CSR efforts not only enhance a company’s public image but also contribute to the development of intangible assets such as brand trust and emotional identification⁸³. Stakeholder theory suggests that by fulfilling social responsibility expectations, companies gain emotional endorsement from consumers and broader social groups, thereby strengthening favorable brand perceptions⁸⁴. Empirical studies consistently confirm the positive impact of CSR on brand awareness, brand loyalty, and overall BE^85–87. Based on this, the study proposes hypothesis 3:

Hypothesis 3

CSR has a positive and significant impact on BE.

CSR initiatives indirectly shape consumer decision-making by elevating BE⁸⁸. When consumers perceive alignment between their values and a company’s CSR efforts, they develop deeper emotional attachment and trust, which in turn enhances purchase intention. This mediating pathway aligns with the “cognition–affection–behavior” model, in which CSR (cognition) enhances BE (affection), ultimately resulting in consumer action (behavior)⁸⁹. Existing research further validates BE’s role as a significant mediator in the relationship between CSR and consumer behavior^90,91. Based on this, the study proposes hypothesis 4:

Hypothesis 4

BE plays a mediating role between CSR and CPI.

Based on the above hypotheses, the structural equation model is specified as follows:

6

7

Here, denotes the path coefficient from to ; represents the path coefficient from BE to CPI; indicates the direct effect of CSR on CPI; and are the corresponding error terms.

Experimental design and performance evaluation

Datasets collection

The data for this study were sourced from two types of multi-source information: first, the Yelp review dataset, and second, the CSRHub corporate social responsibility dataset. The Yelp dataset, publicly released by the Yelp platform, contains authentic user reviews across various service industries, including restaurants, hotels, and beauty services. It includes fields such as review text, ratings, timestamps, and user and business information, making it a reliable resource for research in natural language processing, sentiment analysis, and recommendation systems. The CSRHub dataset, launched by CSRHub in 2007, compiles global corporate social responsibility performance data across dimensions such as environmental impact, employee rights, and philanthropy, providing a robust evaluation tool for CSR research.

To achieve effective integration of Text-CNN and SEM, this study first constructed a CSR-themed keyword lexicon based on CSRHub data, covering three core dimensions: environmental responsibility, employee rights, and philanthropy. Next, Yelp reviews were matched with corporate CSR information, and only reviews containing CSR-related content were retained. This ensured that the input text for the Text-CNN model reflected corporate social responsibility rather than general service quality or consumer experience. The filtered review data underwent standard text preprocessing. Specifically, non-English reviews, special characters, URLs, and HyperText Markup Language tags were removed; stop words and short reviews with fewer than 10 characters were discarded; and tokenization and lowercasing were applied to guarantee high-quality, consistent training corpora.

At the text representation level, this study adopts a dual-channel embedding strategy combining character-level and word-level features: character vectors have a dimension of 100, and word vectors have a dimension of 300. The word vectors are pre-trained on large-scale corpora and fine-tuned to better capture the fine-grained sentiment nuances in Yelp reviews. Sentiment labels are automatically generated by mapping Yelp star ratings (1–2 stars as negative, 3 stars as neutral, 4–5 stars as positive) and then fed into the 2D-Text-CNN model for sentiment classification. The model is trained using a 70% training set, 10% validation set, and 20% test set split, combined with five-fold cross-validation to ensure stable performance and good generalization. After model output, the results are aggregated at the firm level by CSR dimension to form a corporate-level CSR sentiment index, which is aligned with CSRHub data. This operation establishes an auditable text-to-model integration pipeline, providing exogenous variable inputs for subsequent SEM analysis.

To complement the sentiment analysis model and support SEM, a formal questionnaire survey was conducted. Prior to full-scale deployment, a pilot test was carried out with 25 respondents who had prior consumer experience. The pilot assessed item clarity, response comprehensibility, and logical coherence. Feedback indicated that the questionnaire was well-designed, requiring only minor wording refinements. The final version of the questionnaire was developed based on the theoretical model proposed in this study, encompassing three primary constructs: CSR, BE, and CPI. Measurement items were adapted from well-established scales and revised to ensure appropriateness for the Chinese linguistic and cultural context. A five-point Likert scale (1 = strongly disagree; 5 = strongly agree) was used for all items, striking a balance between cognitive ease for respondents and measurement granularity. This format is also consistent with prevailing consumer response tendencies in similar contexts⁹².

The questionnaire was administered online via the Wenjuanxing platform using a non-probability convenience sampling approach. Survey items were specifically designed for Brand A to ensure that the measurement of CPI corresponds directly with Yelp reviews and CSRHub corporate data, thereby enhancing the behavioral relevance and validity of the assessment. The target population comprised mainland Chinese consumers aged 18 or older with prior online shopping experience. To improve sample diversity and representativeness, demographic balance was maintained across gender, age, education level, and geographic location. Quality control measures, including response time restrictions and reverse-coded items, were employed to ensure data validity. A total of 350 questionnaires were distributed, yielding 298 valid responses and an effective response rate of 85.1%. Validity criteria included completeness, internal consistency, and the elimination of extreme or patterned responses. Reliability and validity assessments confirmed data quality, with a Cronbach’s alpha of 0.906 and composite reliability (CR) values exceeding 0.70 for all constructs.

Details of the measurement items, their sources, and final wordings are presented in Table 2^93–108.

Table 2.

Measurement model validation results.

Variable name	Original Item (English)	Revised Item (Final wording)	Reference
Economic Responsibility (ER)	The company provides quality products and services at reasonable prices.	The company provides high-quality products and services at reasonable prices.	Maignan and Ferrell93
	The company pays taxes in accordance with national regulations.	The company pays taxes in compliance with the law and fulfills its economic duties.	Avi-Yonah94
Consumer Rights Protection (CRP)	The company guarantees consumers’ rights to be informed and safe.	The company ensures consumers’ rights to information and product safety.	Scruggs and Ortolano95
	The company responds to customer complaints promptly and fairly.	The company handles customer complaints in a timely and fair manner.	Brown and Dacin96
Employee Rights Protection (ERP)	The company provides safe and fair working conditions for employees.	The company offers employees safe and fair working conditions.	Shaharuddin et al.97
	The company respects employees’ rights and promotes their development.	The company respects employees’ rights and supports their career development.
Environmental Responsibility (ENV)	The company actively takes steps to protect the environment.	The company actively implements environmental protection measures.	Dixon et al.98
	The company implements environmentally friendly practices in production.	The company adopts environmentally friendly practices in its production processes.	Peng et al.95
Philanthropic Responsibility (PHI)	The company donates to charities or community development programs.	The company participates in philanthropy or supports community development.	Brammer and Millington100
	The company organizes or sponsors public welfare activities.	The company organizes or sponsors social welfare activities.
BE	I recognize this brand and can easily recall it.	I can identify this brand and recall it easily.	Yoo and Donthu101
	I trust this brand.	I trust this brand.	Bernarton et al.102
	I feel emotionally connected with this brand.	I have an emotional attachment to this brand.	Chaudhuri and Holbrook103
	I am loyal to this brand and prefer it over others.	I am loyal to this brand and prioritize it when choosing products.	Alhaddad104
CPI	I intend to purchase products from this brand in the future.	I intend to purchase products from this brand in the future.	Kitzmueller and Shimshack105
	I would recommend this brand to others.	I would recommend this brand to others.	Boulding et al.106
	I will choose this brand when purchasing similar products.	I will prefer this brand when purchasing similar products.	Younus et al.107
	I plan to continue buying this brand’s products.	I plan to keep purchasing this brand’s products in the future.	Mirabi and Akbariyeh108

Through the aforementioned procedure, this study achieves multi-source integration of textual data, CSR metrics, and survey responses. Yelp reviews are processed by the Text-CNN model to extract CSR-related emotional features, which are then combined with brand-specific CPI data from the survey to construct the SEM model. In this way, the CSR latent variables not only reflect consumers’ authentic emotional responses but also correspond to the brand- or company-level context, thereby enhancing the empirical validity and inferential credibility of the theoretical framework.

Experimental environment and parameter settings

Regarding the experimental setup, this study was conducted on a Windows 10 operating system with an Intel Xeon Gold 6226R processor, ensuring computational efficiency for model training and data processing. The deep learning framework selected for model construction and training was PyTorch 2.0, while statistical analyses were performed using SPSS 26.0.

For data processing and partitioning, after text cleaning and filtering, a total of 52,184 valid reviews were obtained, of which 16.8% (8,763 reviews) were related to CSR themes and used as the core input for sentiment modeling. Sentiment labels were derived from star ratings: negative (1–2 stars), neutral (3 stars), and positive (4–5 stars), resulting in a distribution of 28.6% negative, 18.9% neutral, and 52.5% positive, maintaining a relatively balanced dataset. The vocabulary was constructed with a minimum frequency threshold of 5, yielding approximately 38,000 unique tokens.

The model employed a dual-channel embedding strategy (character vectors of 100 dimensions and word vectors of 300 dimensions, pre-trained using Word2Vec and fine-tunable). The convolutional layer included three kernel sizes (2 × 2, 3 × 3, and 4 × 4), each with 128 filters. Convolution outputs were processed through ReLU activation and 2 × 2 max pooling, followed by a Dropout layer with a rate of 0.5. Finally, a Softmax layer was applied for sentiment classification.

The training parameters were set as follows: cross-entropy loss function, Adam optimizer with an initial learning rate of 0.001, batch size of 64, and 30 training epochs. The dataset was split into 70% training, 10% validation, and 20% test sets, with five-fold cross-validation applied within the training set. An Early Stopping mechanism was also employed to prevent overfitting. This consistent partitioning and validation process ensured transparency and reproducibility throughout the study.

To evaluate the model’s effectiveness in identifying CSR-related comments, a separate assessment was conducted on the CSR-specific subset. The 2D-Text-CNN achieved a precision of 0.893, recall of 0.891, and an F1-score of 0.897, demonstrating stable recognition of CSR-related sentiment signals. These results outperformed comparison models on the CSR subset, reinforcing the model’s credibility and providing interpretable external input for subsequent SEM analysis.

To comprehensively evaluate the performance of the 2D-Text-CNN model in recognizing sentiment related to CPI, this study compares it with several mainstream sentiment analysis models. The benchmark models include: Deep Pyramid Convolutional Neural Network (DPCNN), Standard Text-CNN, Text-based Recurrent Neural Network (Text-RNN), Bidirectional Encoder Representations from Transformers with CNN (BERT-CNN), and Convolutional Recurrent Neural Network (CRNN). These comparative models represent a range of text representation and feature extraction strategies, encompassing deep convolutional structures, temporal sequence modeling, pre-trained transformer-based embeddings, and hybrid convolutional–recurrent architectures. This comparative analysis provides a robust foundation for assessing the accuracy, robustness, and fine-grained recognition capabilities of the proposed model.

Performance evaluation of the 2D-text-CNN model in CSR sentiment recognition and consumer behavior analysis

To ensure that the sentiment data input into the SEM model are reliable, this study validated the accuracy of the Text-CNN model in sentiment recognition. Figure 5 presents a comparative analysis of the performance of various algorithms across different types of text representations.

Fig. 5.

Comparison of recognition performance across different algorithms and text representation types.

As shown in Fig. 5, the 2D-Text-CNN model employed in this study outperforms comparative models such as BERT-CNN, Text-CNN, and DPCNN across character-level, word-level, and multi-feature fusion sentiment recognition tasks. Metrics including accuracy, precision, recall, F1-score, and AUC-ROC are consistently higher for 2D-Text-CNN. By integrating both character- and word-level embeddings, the model effectively captures fine-grained emotional cues while leveraging multi-scale convolutional kernels to extract both local and global features, thereby enhancing sentiment discrimination. Appropriate regularization strategies, such as Dropout and early stopping, prevent overfitting and improve model generalization and robustness. Compared with traditional single-feature or single-channel models, the 2D-Text-CNN consistently outperforms across multiple evaluation metrics. This performance demonstrates its ability to comprehensively and accurately capture sentiment in complex and diverse user-generated reviews, thereby exhibiting strong practical utility and considerable potential for broader application.

To assess the reliability and validity of the latent constructs, confirmatory factor analysis (CFA) was conducted using AMOS 26.0. The results are summarized in Tables 3 and 4.

Table 3.

Measurement model validation results.

Latent variable	Indicator	Average factor loading	Average variance extracted (AVE)	CR
CSR	ER	0.78	0.59	0.85
	CRP	0.75
	ERP	0.77
	ENV	0.8
	PHI	0.76
BE	—	0.82	0.66	0.88
CPI	—	0.85	0.71	0.91

Table 4.

Model fit indices.

Fit index	Value	Recommended threshold
Chi-square/degrees of freedom (χ2/df)	2.12	< 3
Comparative Fit Index (CFI)	0.958	≥ 0.90
Tucker-Lewis Index (TLI)	0.951	≥ 0.90
Root Mean Square Error of Approximation (RMSEA)	0.056	≤ 0.08
Standardized Root Mean Square Residual (SRMR)	0.043	≤ 0.08

As shown in Tables 3 and 4, the measurement model employed in this study demonstrates overall good performance. The indicators for each latent variable effectively capture their corresponding constructs, indicating that the scale design is reasonable and possesses strong explanatory power and internal consistency. Moreover, the overall model fit meets commonly accepted academic standards, reflecting high reliability and stability. Collectively, these validation results provide a solid foundation for subsequent analysis of path relationships, ensuring that the study’s conclusions are both scientific and credible.

To examine whether the data are subject to significant common method bias, this study employed both Harman’s single-factor test and the CFA marker variable approach. The results are presented in Table 5.

Table 5.

Common method bias test results.

Indicator	Eigenvalue	Variance explained (%)	Cumulative variance explained (%)	Standardized factor loadings (CFA marker variable)
All Measured Items Combined	6.74	27.15	27.15	-
ER	3.21	12.84	39.99	0.72–0.84
CRP	2.47	9.94	49.93	0.68–0.81
ERP	1.91	7.68	57.61	0.70–0.82
ENV	1.35	5.44	63.05	0.66–0.79
PHI	1.02	4.09	67.14	0.65–0.77
BE	0.87	3.48	70.62	0.74–0.88
CPI	0.81	3.24	73.86	0.71–0.85

As shown in Table 5, the assessment using Harman’s single-factor test and the CFA marker variable approach indicates that the study data do not exhibit significant common method bias. The first factor accounts for only 27.15% of the variance, and the high correlations among latent variables primarily reflect genuine relationships rather than measurement artifacts. This demonstrates that the self-reported survey data are robust and interpretable, providing a reliable foundation for the subsequent SEM path analysis.

Figure 6 presents the descriptive statistics derived from the 2D-Text-CNN-based empirical study on CPI.

Fig. 6.

Descriptive statistics of the empirical study.

Figure 6 illustrates that the means of most variables are centered around 3, indicating that respondents generally hold a neutral-to-slightly-positive evaluation of these factors. CSR and economic responsibility show slightly higher means, suggesting that participants generally consider these responsibilities important. BE has a slightly lower mean, reflecting more cautious assessments. CPI is at a moderately high neutral level overall. Standard deviations indicate some variability in respondents’ perceptions across certain variables, particularly in employee rights and environmental protection dimensions. Skewness and kurtosis results show that most variables exhibit relatively flat distributions close to normal, indicating that while scores are generally concentrated, differences among respondents exist.

To examine consumers’ emotional responses to CSR and their effects on BE and purchase intention, the study employed the 2D-Text-CNN model for sentiment classification of online reviews. The specific results of this sentiment analysis are presented in Fig. 7.

Fig. 7.

Sentiment analysis results.

As shown in Fig. 7, consumers generally hold a positive attitude toward corporate CSR, with positive emotions predominating (approximately 45%–52%), neutral emotions accounting for roughly one-third, and negative emotions representing a relatively small proportion. Among the various CSR dimensions, ENV and PHI exhibit the highest proportions of positive emotions, indicating strong consumer recognition of corporate efforts in environmental protection and public welfare. In contrast, ERP shows a relatively higher proportion of negative emotions, suggesting that consumer perceptions in this area remain more divided. Furthermore, BE and CPI also display evident positive emotional responses, indicating that CSR enhances consumer identification with the brand and also effectively translate into purchase intention. These findings support the core logic that the various dimensions of CSR influence CPI through the mediating role of BE.

Figure 8 illustrates the results of the Pearson correlation analysis conducted on the variables used in the 2D-Text-CNN-based empirical study.

Fig. 8.

Correlation analysis results of the empirical study.

As shown in Fig. 8, CSR and its individual dimensions are all significantly positively correlated with CPI, indicating that the more comprehensive a company’s CSR practices, the higher the consumer’s willingness to purchase. Among the dimensions, ENV, PHI, and ERP have the most pronounced effects on CPI. Overall CSR scores exhibit a close relationship with CPI. Additionally, BE shows the strongest correlation with CPI, suggesting a potential mediating role between CSR and customer purchase intention.

Based on the CSR sentiment indices generated by the 2D-Text-CNN model and survey data, a complete SEM was constructed to examine the influence pathways of CSR and its dimensions on CPI and to test the mediating effect of BE. The detailed results are presented in Tables 6 and 7.

Table 6.

Structural model fitting index.

Fit Index	Value	Recommended threshold	Result
χ²/df	2.18	< 3	Good
CFI	0.956	> 0.90	Good
TLI	0.949	> 0.90	Good
RMSEA	0.057	< 0.08	Good
SRMR	0.045	< 0.08	Good

Table 7.

Structural model path coefficient and mediating effect (standardization).

Independent Variable (X)	Mediator (M)	Dependent Variable (Y)	Standardized Path (c’)	Standardized Indirect Effect (a×b)	95% Bootstrap CI	Total Effect (c)
CSR	BE	CPI	0.361***	0.226***	[0.190, 0.265]	0.587***
ER	BE	CPI	0.192***	0.143***	[0.115, 0.173]	0.335***
CRP	BE	CPI	0.177***	0.136***	[0.110, 0.165]	0.313***
ERP	BE	CPI	0.201***	0.131***	[0.105, 0.160]	0.332***
ENV	BE	CPI	0.219***	0.144***	[0.117, 0.177]	0.363***
PHI	BE	CPI	0.213***	0.137***	[0.110, 0.169]	0.350***

As shown in Table 6, the overall model exhibits good fit, meeting the requirements for SEM analysis. The R² values for the endogenous variables indicate that BE accounts for 62.1% of the variance, while CPI accounts for 73.8% of the variance, demonstrating that the model effectively explains the relationships among the variables.

As shown in Table 7, CSR and its individual dimensions all exhibit significant positive direct effects on CPI. BE plays a significant partial mediating role across all paths, with indirect effects accounting for approximately 38–44% of the total effects, indicating that BE serves as a crucial bridging mechanism in the CSR→CPI pathway. Notably, the mediating effects of the ENV and PHI dimensions are the most prominent, highlighting the strong influence of corporate actions in environmental protection and social welfare on enhancing consumer purchase intention.

The results provide empirical support for the hypothesized relationships:

• CSR has a significant positive impact on CPI (supporting H1).

• BE positively influences CPI (supporting H2).

• CSR significantly enhances BE (supporting H3).

• BE partially mediates the relationship between CSR and CPI (supporting H4).

Overall, the findings validate the theoretical framework grounded in the cognition–affection–behavior model. CSR initiatives enhance consumers’ cognitive evaluations of the firm, elevate emotional attachment via BE, and ultimately increase behavioral intentions such as product purchases. This study offers robust theoretical and empirical evidence to guide enterprises in designing effective CSR strategies and optimizing brand development to foster long-term consumer engagement.

The structural model path diagram of this study is presented in Fig. 9.

Fig. 9.

Structural model path diagram.

Figure 9 visually illustrates the influence pathways of CSR and its five dimensions on CPI, while highlighting the mediating role of BE. In the figure, the blue area represents CSR and its dimensions, the magenta area denotes the mediator BE (R² = 0.621), and the orange area represents the dependent variable CPI (R² = 0.738). Solid lines indicate the indirect paths from CSR/dimensions → BE → CPI, and dashed lines represent the direct paths from CSR/dimensions → CPI. All paths are labeled with significant positive coefficients (***p < 0.001). The diagram also confirms good model fit and shows that the mediating effects are validated through Bootstrap testing, clearly demonstrating the core mechanism whereby CSR enhances CPI by strengthening BE.

Discussion

This study innovatively integrates the Text-CNN deep learning model with SEM to examine the impact of CSR on CPI and the mediating role of BE. Unlike prior research that primarily relies on survey data and linear statistical models, this study overcomes methodological limitations by leveraging Text-CNN to extract fine-grained emotional features from consumer online reviews and incorporating them into the SEM framework for systematic analysis. This dual-method approach enhances the precision of theoretical validation and improves the depth of empirical analysis. In addition, it fosters the integration of deep learning techniques with behavioral science, addressing a previous gap in which these approaches were often applied separately without meaningful interaction.

At the theoretical level, this study expands the application of social identity theory in the context of CSR. The results indicate that consumers’ identification with corporate CSR activities manifests not only at the level of rational cognition but also through emotional responses. The mediating role of BE in the CSR–consumer behavior relationship highlights the dual mechanisms of cognition and emotion, providing a more comprehensive framework for understanding consumer reactions to CSR. By integrating stakeholder theory and signaling theory, the study demonstrates how CSR initiatives convey positive signals through brand image, thereby enhancing consumer trust and identification with the firm. This multi-theoretical perspective enriches the explanation of CSR’s impact mechanisms, emphasizing that CSR generates value not only at the transactional level but also in relational and reputational dimensions. Methodologically, the study achieves notable innovation. Previous applications of sentiment analysis in CSR research have largely remained at the level of basic emotion classification. In contrast, this study embeds Text-CNN–extracted sentiment features into an SEM behavioral model, illustrating how emotional responses systematically influence BE and purchase intention. This approach enhances the scientific rigor of CSR research and provides a replicable interdisciplinary paradigm for future scholars. Finally, the findings reveal dimension-specific effects of CSR. Environmental responsibility and philanthropic initiatives elicit the strongest positive emotional responses, which in turn significantly enhance BE and consumer purchase intention. This result underscores the varying contributions of different CSR dimensions to consumer behavior, offering a more nuanced theoretical entry point for future research.

At the managerial level, this study also holds significant practical value. For enterprises, the results indicate that investments in environmental responsibility and philanthropic initiatives yield the most substantial returns in terms of BE and consumer behavior. Therefore, companies should prioritize resource allocation to areas that closely align with consumer values, while also attending to employee rights and consumer protection to prevent potential negative perceptions. For brand managers and marketing professionals, the revealed mediating role of BE suggests that CSR should be closely integrated with brand communication. For instance, CSR activities can be embedded into brand storytelling, communicated transparently via social media, and involve consumers in co-creating CSR initiatives to strengthen emotional attachment and brand loyalty. Additionally, firms should leverage AI-driven sentiment monitoring tools to dynamically track consumer responses to CSR and adjust communication strategies in a timely manner. For policymakers, the findings highlight the need to establish institutional mechanisms that encourage corporate CSR engagement. This may include financial incentives for green initiatives, recognition and awards for philanthropic efforts, and stricter regulation of employee and consumer rights. Such measures can guide enterprises to actively fulfill their social responsibilities, thereby promoting sustainable development and overall societal well-being.

In summary, this study makes meaningful contributions at both theoretical and practical levels. Theoretically, it provides a deeper explanation of how CSR influences consumer behavior through BE by integrating multiple theoretical perspectives, incorporating emotional dimensions, and introducing methodological innovations. Practically, it offers actionable guidance for corporate CSR strategy implementation, brand communication, and policy design. By integrating advanced sentiment analysis techniques with the SEM framework within a multi-theoretical context, this study offers new perspectives and practical insights for understanding and optimizing the mechanisms through which CSR influences consumer behavior.

Conclusion

Research contribution

This study integrates deep learning techniques with empirical modeling by employing an optimized 2D-Text-CNN framework to conduct sentiment analysis on consumer online reviews. By incorporating both character-level and word-level features, the model enhances the granularity of textual understanding and emotional feature extraction. Grounded in social identity theory, the study constructs a SEM to investigate the relationships among CSR, BE, and CPI, with particular emphasis on the mediating role of BE in the CSR–CPI linkage. Experimental results confirm that the proposed 2D-Text-CNN model achieves superior performance across three types of feature representations. Notably, it attains an F1 score of 0.951 under the integrated feature scenario, validating its multidimensional feature extraction capability. Among benchmark models, BERT-CNN demonstrates strong semantic comprehension, achieving a precision of 0.926 in word-level scenarios. The dual-channel feature fusion strategy notably enhances the performance of 2D-Text-CNN, highlighting the benefits of combining granular and semantic representations. Empirical findings reveal that CSR exerts a significant positive effect on CPI, with environmental responsibility and employee rights protection emerging as the most influential CSR dimensions. Furthermore, BE plays a pronounced mediating role in the relationship between CSR and CPI. The results demonstrate that various CSR dimensions indirectly reinforce CPI by enhancing BE, with environmental and philanthropic responsibilities exhibiting the strongest mediating pathways.

Overall, the study confirms Hypotheses 1 through 4, substantiating that CSR initiatives not only directly affect customer behavior but also indirectly do so by fostering BE. These findings suggest that enterprises can strengthen their social image and consumer motivation by emphasizing CSR engagement, thereby achieving both reputational and market-oriented objectives.

Future works and research limitations

Although this study provides valuable insights into the impact of CSR on CPI, several limitations warrant consideration. First, despite the high accuracy achieved by the optimized 2D-Text-CNN model, NLP techniques, particularly those based on deep learning, still face challenges in interpreting complex semantic expressions such as sarcasm, irony, and implicit sentiment. These limitations may lead to misclassification of sentiment polarity, affecting the precision of the analysis. Additionally, the model’s performance is highly dependent on domain-specific corpora, which constrains its generalizability across different industries and cultural contexts. As a result, the model may struggle to maintain accuracy and interpretability when applied to more heterogeneous real-world scenarios. Second, the study adopts a cross-sectional design and thus does not capture the long-term or delayed effects of CSR on consumer behavior. In practice, the influence of CSR often accumulates over time and may exhibit lagged effects on consumer perception and intention. Future research would benefit from longitudinal data to explore the dynamic evolution of these relationships and to better understand how CSR impacts unfold over time. Third, this study does not account for potential moderating variables—such as consumer values, corporate size, or industry sector—that may condition the strength or direction of CSR’s effects. These contextual factors could play a significant role in shaping consumer responses to CSR and warrant further investigation.

Future research may advance in several directions. First, the adoption of more sophisticated sentiment analysis models—such as Transformer-based architectures or hybrid models incorporating multimodal inputs—could enhance the interpretation of complex semantics and improve robustness across domains. Second, a longitudinal design would enable the examination of temporal dynamics and cumulative effects, offering a more comprehensive understanding of CSR’s influence. Lastly, exploring potential synergistic or offsetting effects among different CSR dimensions could enrich the theoretical framework and provide more targeted practical strategies for leveraging CSR to influence consumer behavior.

Author contributions

Zhengshun Shen: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparationHuaibin Li: writing—review and editing, visualization, supervision, project administration, funding acquisitionYancai Zhang: methodology, software, validation, formal analysis.

Funding

This research received no external funding.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author Zhengshun Shen on reasonable request via e-mail zhshshen@hytc.edu.cn.

Declarations

Competing interests

The authors declare no competing interests.

Ethics

The studies involving human participants were reviewed and approved by School of Economics and Management, Huaiyin Normal University Ethics Committee (Approval Number: 2023.12500253). The participants provided their written informed consent to participate in this study. All methods were performed in accordance with relevant guidelines and regulations.