Mining the impact of social media information on public green consumption attitudes: a framework based on ELM and text data mining

China’s drive toward carbon neutrality has intensified the need for low-carbon, green transformation. Because consumption drives economic growth, promoting green consumption from the demand side is crucial. Multiple societal actors—government, enterprises, and social organizations—use information strategies to persuade the public toward green behaviors, yet adoption remains challenging. Key questions include whether such information attracts attention, increases understanding, and effectively shapes attitudes. Social media offers real-time, large-scale data on consumer sentiment and behavior, enabling analysis of how information dissemination and social influence affect green consumption attitudes. Existing behavior theories (TPB, SDT, TAM, Uses and Gratifications, Diffusion of Innovations) have limits in explaining how external information alters attitudes through different cognitive processing routes. ELM, which accounts for both high-level and low-level processing and individual differences in motivation and ability, is well-suited to explain attitude formation and change in response to information. This study uses ELM with text mining to categorize information from different stakeholders, analyze public reactions, and quantitatively assess shifts in attitudes using large-scale social media text, filling gaps in prior small-sample, survey-based research.

The review covers two areas: (1) Text mining of green consumption attitudes: Traditional surveys/interviews face limitations in sample size and representativeness, while text mining offers objective insights from large-scale online comments (e.g., LDA topic modeling, sentiment analysis) to assess public attention, emotions, and attitudes. (2) Influencing factors: Internal factors (subjective norms, perceived behavioral control) and external factors (policy, economic incentives, technology, social pressure) shape green attitudes and willingness. Prior work often isolates single information sources (policy, enterprise, or social), missing multi-source exposure that includes cognition, emotion, and behavioral tendencies. This study aggregates news and social interactions from multiple stakeholders to mine multi-level content on cognition, emotional preferences, and behavioral tendencies, enabling a multi-dimensional exploration of public attitudes and addressing representativeness by leveraging large online samples. The review also notes that ELM is underused in green consumption studies relative to ABC, TPB, TRA, KAB, and Affect-Cognition theories; combining ELM with text mining enables dynamic tracking of attitude change and information processing.

The study integrates ELM with text mining in a four-step framework: (1) Data collection: Retrieve official news and online comments related to “green and low-carbon consumption” (2019–2022) from Baidu News, Toutiao, and Weibo, focusing on public-related content. Total: 6425 news posts; 12,696 Weibo posts/comments randomly sampled for in-depth analysis. (2) Preprocessing: Remove noise, stop words, symbols; tokenize and filter low-frequency words. (3) Text mining: - Subject classification via text CNN into three categories (government policy, corporate measures, media promotion); labeled dataset included 3177 government, 5058 corporate, 3177 media posts; CNN accuracy 66.7%. - Topic modeling via LDA to identify subtopics under each subject. - Sentiment analysis via LSTM to classify public emotions (positive/negative), followed by a fully connected layer for visualization; co-occurrence semantic networks constructed. - Demographics: infer commenter gender and region, grouped into seven Chinese macro-regions (NE, N, C, E, S, NW, SW). (4) ELM-based attitude modeling: Construct three ELM-consistent factors from comments using LIWC and unsupervised algorithms: Attention and Understanding (Cogmech), Participation Mode (Certain), and Belief/Attitude Change (Affect). Quantify factors by lexicon-based counts. Incorporate policy environment (Policy Modeling Consistency, PMC) measured by policy intensity, goals, and measures (2019–2022: PMC 16, 35, 40, 49) and economic environment via Consumer Confidence Index (CCI). Analyze with Generalized Linear Mixed Models (GLMM) due to count data and non-normality, following an incremental modeling strategy. Hypotheses: information strategies influence attitudes; processing routes differ; attitudes vary by region/gender and evolve; positive linear relations among Cogmech, Certain, Affect; and environmental fluctuations modulate responses.

- Topic structures (LDA): Government topics: environmental plastic reduction, anti-food waste, garbage sorting, low-carbon travel, vehicle carbon reduction, green clothing. Corporate topics: express carbon reduction, eco-friendly food delivery, plastic reduction, green home appliances, green office. Media topics: advocate low-carbon life, plastic reduction, express carbon reduction, takeaway tableware, food waste, energy-saving appliances, green transportation. - Sentiment (LSTM): Government (positive/negative): plastic reduction 54.49%/44.38%; anti-food waste 46.02%/53.27%; garbage sorting 50.63%/45.63%; low-carbon transport 67.32%/31.68%; vehicle carbon reduction 26.68%/71.57%; green clothing 25.00%/74.07%. Corporate: express carbon reduction 68.04%/28.77%; eco-friendly food delivery 68.90%/28.65%; plastic reduction 64.28%/34.18%; green office 83.93%/16.07%; green home appliances 70.09%/27.68%. Media: advocate low-carbon life 40.39%/57.78%; plastic reduction 45.19%/53.56%; express carbon reduction 37.54%/58.86%; takeaway tableware 25.00%/75.00%; food waste 41.11%/57.98%; energy-saving appliances 70.09%/27.68%; green transportation 60.38%/37.26%. Overall, public sentiment is positive toward government and corporate content but negative toward media promotion. - Demographics: Women participate more; East China shows the highest attention to all subjects; Northwest the lowest. North and South China also show notable attention depending on subject. Corporate topics attract attention in North, Central, and Southwest China; media knowledge dissemination is most engaged by Central and North China. - ELM factor relationships: Positive linear relationships exist among Cogmech, Certain, and Affect, strongest between Cogmech and Certain. - GLMM results: Policy topics: Cogmech significantly predicts Certain and Affect; time effect notable in 2021 for Cogmech; CCI and Certain generally not significant for Affect. Corporate topics: Higher CCI negatively regulates Cogmech and Certain; Cogmech positively predicts Certain but negatively relates to Affect; Certain positively predicts Affect; time positively associated with Affect (notably in 2021). Media topics: Cogmech increased in 2019–2020; CCI negatively affects Cogmech and Certain; Certain positively predicts Affect; enhanced Cogmech does not translate into higher Affect; regional and gender differences significant for participation and attitude change. - Overall: Women and residents in economically advanced regions are more engaged; public responds positively to governmental and especially corporate actions; media campaigns often elicit negative sentiment. Attention/understanding and participation influence attitudes, but effects vary by subject and context.

The findings support the ELM framework: information processing (attention/understanding and participation) relates to attitude formation, with relationships contingent on source and content. Government policy information improves attitudes primarily through enhanced attention and understanding, but mere participation (sharing) does not change attitudes. Corporate measures garner strong positive sentiment; participation (supporting/forwarding) improves attitudes, whereas increased attention/understanding alone may not shift attitudes due to perceived costs and habit changes. Media dissemination boosts awareness and participation over time, but often evokes skepticism and negative emotions, limiting attitude change. Economic conditions (higher CCI) tend to dampen attention, participation, and attitude change toward green consumption, suggesting that favorable economic sentiment may reduce perceived need or efficacy of green behaviors. Policy environment intensity (PMC) shows limited direct influence on attitude, consistent with top-down, advisory-level policies lacking specific green consumption laws. Practical implications: enhance targeted, detailed information matching consumers’ cognitive levels; leverage “Internet +” channels and regular news integration of green narratives; deploy celebrity endorsements to build trust and emotional engagement; provide transparent product and process details; and design incentive systems (e.g., green points, personal carbon accounts) to strengthen perceived efficacy and bridge the attitude-behavior gap.

This study proposes and validates a novel framework integrating ELM with text mining to monitor and explain public green consumption attitudes using large-scale social media data. By classifying subjects (government, corporate, media), extracting topics and sentiment, and modeling ELM-derived cognitive, participatory, and attitudinal factors with GLMM, the study offers actionable insights: women and residents in advanced regions are more attentive; public sentiment is positive toward government and corporate actions but negative toward media campaigns; attention/understanding and participation shape attitudes with source-specific effects; economic optimism may reduce green attitude shifts, and policy intensity alone is insufficient. The framework provides a scalable alternative to traditional surveys and can be generalized to other domains where public attitudes evolve with information exposure. Future work should extend cross-nationally, broaden and diversify samples beyond social media users, and refine factor dictionaries and algorithms to capture additional drivers, enhancing predictive power for attitude and behavior change.

- Sample scope: Focus on Chinese online users limits generalizability; cross-country validation is needed. - Representation: Social media users may not represent all age and demographic groups, potentially biasing results. - Factor coverage: LIWC-based and unsupervised dictionary construction may omit relevant factors; expanding lexicons and algorithms could yield more precise quantification. - Model performance: Subject classification accuracy (≈66.7%) suggests room for improvement in NLP pipelines which may affect downstream analyses.