China’s carbon peaking policy intensity dataset across national, provincial, municipal to county-level over 2021 to 2025

Abstract

Policy texts serve as the carriers of policy content and hold unique value in the research of policy differences, policy diffusion, and inter-governmental relations. Carbon peaking is a key strategic goal for China to address global climate change. The carbon peaking implementation plan provides a carrier for exploring China’s carbon peaking path and actions. However, at present, there is relatively little research on the policy text of the carbon peaking implementation plan. In this study, for the first time, a systematic collection of publicly available carbon peaking implementation plans in China was conducted, and the policy intensity was quantified by applying a combination of natural language processing and machine learning methods. This dataset includes 316 national, provincial, municipal and county-level policies during the period from 2021 to 2025, and quantifies the policy intensity from the dimensions of three policy instruments: legal, market and supervision, as well as three policy targets: precise, general and vague. This dataset can help researchers further understand China’s carbon peaking policy system and provide basic information for subsequent policy comparative studies and policy evaluations.

Background & Summary

Under China’s carbon peaking goal, local governments at all levels have successively released carbon peaking implementation plans. As a guiding document, the carbon peaking implementation plan provides a window for understanding the carbon peaking paths, measures and specific actions of each region and even the whole country. However, in China’s multi-level government system, there is still a lack of necessary research on what characteristics the policy plans issued by different local governments have, what regional differences they show, and how to quantify these policy differences. Therefore, it is particularly important to develop a policy dataset for carbon peaking implementation plans from the national level to the county level.

Since the signing of the United Nations Framework Convention on Climate Change in 1992, over the past 30-plus years, climate policies have gradually expanded from international and national scales to administrative units such as provinces, cities, and counties. The number of climate policies has grown explosively¹, posing a huge challenge to the traditional policy research paradigm. Quantitative analysis has become the basic trend in the field of climate policy research².Wu (2024) systematically reviewed the allocation of powers and responsibilities among different levels of government in China’s climate governance, providing an institutional analytical foundation for understanding intergovernmental disparities in climate policy³. Zhao and Qi (2022) characterized the three-phase evolution of China’s climate policy, offering a theoretical framework for investigating the trajectory of climate policy targets⁴. Goldberg et al. (2025) employed a conjoint experimental design to uncover the heterogeneous effects of policy attributes on public support⁵. Wu (2025) provided a comprehensive overview of China’s climate governance landscape by tracing policy evolution, analyzing major policy instruments, evaluating policy effectiveness, and identifying ongoing challenges⁶.

Typical quantitative research on climate policies originates from econometric models in economics. Early studies predominantly employed system dynamics models⁷ and difference-in-differences (DID) frameworks⁸ to assess policy effectiveness. In recent years, more practice-oriented quantitative approaches have emerged: first, through the coding and scoring of core policy components—such as policy targets, instruments, and implementation requirements—to achieve standardized quantification of policies⁹; second, by aligning policy implementation timelines with carbon emissions data to precisely identify effective emission reduction policies and their combinations¹⁰. Overall, the existing literature remains largely focused on policy effect validation, with insufficient systematic analysis of intrinsic policy text attributes such as target differentiation and regional adaptability. This limits the capacity to support quantitative assessments of multi-level climate policy disparities. Moreover, approaches that rely on proxy variables for indirect evaluation are inherently limited in capturing the textual characteristics of policies themselves¹¹.

Policy texts serve as the carriers of policy content¹² and possess unique research value in studies such as policy differences¹³, policy diffusion¹⁴ and inter-governmental relations¹⁵. Policy targets are the textual representations of policymakers’ intentions¹⁶, while policy instruments are a series of political actions and behavioral patterns adopted by policymakers to achieve a certain policy target¹⁷. When governments choose policy instruments, they are influenced by multiple factors¹⁸. At the same time, Different policy instruments also have different impacts on the implementation effect of policies^19,20. Policy intensity is a compound indicator, calculated based on the implementing entity of the policy instrument and the accuracy of the expression of the policy target. For example, Dong et al. classified the policy targets and policy instruments of low-carbon policies into three levels, and defined the policy intensity as the product of the policy targets and policy instruments²¹; Zhang et al. defined policy intensity as the product of policy measures and policy targets²². The quantitative analysis of policy intensity directly affects the assessment of policy effects, thereby providing a basis for the continuous monitoring of policy effectiveness and the analysis of the carbon peaking policy system. China first proposed the national-level “Action Plan for Carbon Dioxide Peaking Before 2030” on October 24, 2021. With the increasing number of carbon peaking implementation plans, it is crucial to establish a carbon peaking policy database and conduct unified quantitative research on it.

From the perspective of methodological genealogy, policy text research is undergoing a paradigm shift from an interpretative to a computational paradigm²³. Early studies were predominantly interpretative, relying on individual researchers’ experiential judgments for textual exegesis—an approach that inherently limits independent replicability and verifiability²⁴. In recent years, with the development of information technology and the integration of disciplines, quantitative research on policy texts has emerged as a new research paradigm. It takes a large number of policy texts in specific fields as the research objects²⁵, enabling researchers to understand the regional heterogeneity of policies from a macro perspective, explore the evolution laws of policies²⁶, diffusion paths²⁷, etc. Capture information that is difficult to obtain in small-sample studies²⁸. More importantly, the quantitative research on policy texts is reproducible, enhancing the reliability of policy research²⁹. Early quantitative research methods for policy texts included bibliometric methods³⁰, policy modeling consistency index methods³¹, etc. In recent years, with the development of artificial intelligence, Natural Language Processing (NLP) and Machine Learning (ML) have begun to be applied to the textual quantitative research in the field of climate policy³².

This article systematically collects the carbon peaking implementation plans publicly released by Chinese governments at all levels from 2021 to 2025, and quantifies the policy differences and policy intensity by combining NLP and ML algorithms. This database provides a new tool for understanding and evaluating China’s carbon peaking policy system. The quantitative results will help researchers systematically observe the regional differences in China’s carbon peaking policies. In addition, our dataset provides detailed sources of carbon peaking policies, enabling researchers to reduce the labor and time costs of collecting such policies. Our dataset provides fundamental variables such as policy text differences and policy intensity, which researchers can further utilize to systematically assess China’s carbon peaking policies and actions.

Methods

The method summary of this study is shown in Fig. 1. The research framework consists of six modules: First, text collection, establishing a carbon peaking policy text library. Second, data processing, including text cleaning, keyword extraction and text content differentiation, extracts intuitive information from policy texts. Third, text classification and manual annotation: Policy instruments are extracted based on keywords and scored, and the remaining policy targets are manually annotated. Fourth, model selection: Use the training set formed by manual annotations for model selection and optimization. Fifth, policy quantification: Use the trained model to quantify the policy intensity, transform the text into intuitive data, and construct the next analysis framework. Sixth, technical verification: Discuss the model performance, comparison with existing databases, and manual checking.

Fig. 1.

The research framework.

Data preparation

The first step is to search and collect the carbon peaking implementation plans publicly issued by governments at all levels from 2021 to 2025 from the official websites of various governments and carbon neutrality network (https://www.ccn.ac.cn/). As of September 30, 2025, a total of 316 carbon peaking implementation plans from the national, provincial, municipal and county levels have been collected and uniformly saved in word format. Among them, there is 1 national-level document, 24 provincial-level documents, 149 municipal-level documents, and 142 county-level documents. The specific distribution is shown in Fig. 2 and Table 1. It should be noted that the dataset only includes publicly released policy documents. Regions not included in this dataset do not mean that they have not released carbon peaking implementation plans. The database provides the enactment dates and URLs of carbon peaking policies across regions over a four-year period, among other information, and serves as a fundamental dataset for exploring the characteristics of carbon peaking policies.

Fig. 2.

The map of policy release region.

Table 1.

Geographic distribution of policies.

Serial number	Region	Province	The number of province	The number of city	The number of county	Total
1	Northeast China	Jilin	1	6	1	12 (3.8%)
		Liaoning	1	3	0
2	North China	Beijing	1	11	0	47 (14.9%)
		Tianjin	1	15	0
		Hebei	1	1	0
		Shanxi	1	7	2
		Inner Mongolia	1	3	3
3	Central China	Henan	1	0	0	68 (21.6%)
		Hunan	1	14	47
		Hubei	0	4	1
4	East China	Shanghai	1	15	0	105 (33.3%)
		Jiangsu	1	8	10
		Zhejiang	0	1	6
		Anhui	1	0	0
		Fujian	0	0	1
		Jiangxi	1	10	34
		Shandong	1	14	1
5	South China	Guangdong	1	11	14	33 (10.5%)
		Guangxi	1	1	1
		Hainan	1	1	2
6	Northwest China	Shaanxi	1	0	2	14 (4.4%)
		Gansu	1	0	0
		Qinghai	1	0	3
		Ningxia	1	2	3
7	Southwest China	Chongqing	0	3	0	36 (11.4%)
		Sichuan	1	14	9
		Guizhou	1	2	1
		Yunnan	1	2	1
		Xizang	0	1	0
Total		—	24	149	142	315

The second step is the clean-up of policy texts. The carbon peaking policy has a highly fixed feature, consisting of structural modules such as overall requirements, main goals, key tasks, international cooperation, policy guarantees, and organization and implementation. Remove the fixed content irrelevant to the policy quantification in the text, such as guiding ideology, working principles, and responsible departments for division of labor, and save the remaining content in txt format.

The third step is policy classification. This article classifies carbon peaking policies into four levels - national, provincial, municipal, and county - based on the issuing institutions, and divides the policy contents into two categories: policy instruments and policy targets. To extract the policy instrument and policy target texts, first, regular expressions are applied to segment the full policy texts into individual sentences based on delimiters such as line breaks and periods. Policy instrument keywords are then used to match and extract relevant sentences, while the remaining unmatched portions are uniformly categorized as policy targets, thereby ensuring non-overlapping classification between policy instruments and policy targets.

Policy text differences

Use jieba for word segmentation and keyword extraction. Jieba, as a Chinese word segmentation package for python, can split Chinese text into independent words and extract important keywords and their weights from the text based on part of speech, which is conducive to further analyzing the key contents of the carbon peaking policy. The SBert pre-trained model was used for sentence difference detection and text analysis. The paraphrase-multilingual-MiniLM-L12-v2 multilingual sentence embedding model in the Sentence Transformers library was selected³³. Calculate the differences in the content of policy texts between superior and subordinate governments as well as among governments at the same level. Paraphrase-multilingual-MiniLM-L12-v2 is a multilingual sentence embedding model based on the BERT architecture, which can perform sentence difference detection and text analysis in a multilingual environment³⁴. The mathematical formulas for calculating textual dissimilarity are detailed in Formula (1) and Formula (2) below:

$$\mathit{cosine}\mathrm{\_}\mathit{similarity}(A,B)=\frac{A\bullet B}{\parallel A\parallel \parallel B\parallel }=\frac{{\sum }_{i=1}^n{A}_i{B}_i}{\sqrt{{\sum }_{i=1}^n{A}_i^2}\sqrt{{\sum }_{i=1}^n{B}_i^2}}$$ 1

$$\mathit{difference}=1-\mathit{cosine}\mathrm{\_}\mathit{similarity}$$ 2

In these two formulas, cosine_similarity(A, B) denotes the textual similarity between texts A and B, measured by cosine similarity, while difference refers to the textual dissimilarity between the two texts. A and B represent the embedding vectors of texts A and B, respectively, generated by the SBERT model. A⋅B denotes the dot product of vectors A and B. ||A|| and ||B|| refer to the L2 norm (Euclidean norm) of vectors A and B, respectively, calculated as the square root of the sum of the squares of all vector elements.

Types of policy instruments

Existing studies have generally categorized policy instruments into three types. For instance, Dong et al. (2024), in constructing a dataset on the intensity of China’s low-carbon policies, classified policy instruments into command-and-control, market-based, and composite categories²¹. Liao (2018) categorized them as command-and-control, market-based, and information-based instruments³⁵. In the quantification of environmental policies, legal instruments are often regarded as more mandatory and binding, and are thus assigned higher weights³⁶. Under the rigid constraint of achieving a carbon emissions peak by 2030, policy instruments outlined in carbon peaking implementation plans must coordinate and integrate ten major actions—including energy transition, transportation, and urban–rural development—while simultaneously advancing multiple national objectives related to economic restructuring, technological innovation, and social transformation. This unique policy context renders the instrument taxonomy, operational logic, and implementation pathways of such plans distinctly different from those examined in prior research³⁵. Accordingly, this study builds on existing classification frameworks by categorizing policy instruments into three types: legal instruments, market instruments, and supervision instruments. The keywords and scores of each type of policy instrument are 3, 2, and 1 respectively (Table 2). In the same policy instrument sentence, there may be two or more policy instruments simultaneously. To avoid repetition, it is set that each sentence can only belong to one policy instrument, with the priority being legal type > market type > supervision type.

Table 2.

Keywords and scores of policy instrument.

Policy instrument	Score	Keywords
Legal	3	Laws, standards, approvals, violations, and irregularities, etc.
Market	2	Taxation, electricity prices, finance, loans, bonds, credit, etc.
Supervision	1	Assessment, evaluation, commendation, rewards, criticism, etc.

Different policy instruments exhibit variation in effectiveness during actual implementation. To enhance the transparency and flexibility of the dataset, we have provided in the data file (3. A dataset of policy intensity for 316 carbon peaking implementation plans) the raw sentence counts for each category of policy instruments and policy targets. This allows users to adjust weights or recalculate intensity according to their specific research contexts. For instance, if researchers consider supervision instruments to be more effective in certain contexts, they may reassign weights based on the original sentence counts, thereby flexibly achieving their research objectives.

Manual scoring

The manual quantification of policy intensity involves the identification of the precision of policy targets, with the intensity score of each policy target ranging from 1 to 3 points. We randomly selected 99 out of the 316 policy target documents and conducted manual annotation on them. Following sentence segmentation and cleaning, a total of 16,835 individual policy target statements were obtained from these 99 documents. During the model construction phase, we randomly split these 16,835 annotated statements into a training set and a test set at a ratio of 75%:25%. The training set comprised 12,626 statements, and the test set contained 4,209 statements. Referring to the scoring criteria established by Dong et al.²¹, we classified policy targets based on the clarity of their wording: statements containing explicit quantitative indicators were categorized as precise and assigned a value of 3; those with vague content and lacking key delimiting elements such as timeframes or geographic scope were classified as vague and assigned a value of 1; and all other statements falling between these two categories were uniformly classified as general and assigned a value of 2. Examples are shown in Table 3 below. This policy target score will be used to train machine learning models for calculating the policy intensity.

Table 3.

Scores and examples of policy target.

Policy target	Score	Examples
Precise	3	By 2025, fast-charging stations will cover 50% of highway passenger transport hubs.
		By 2030, be capable of responding to over 5% of peak load demand.
General	2	By 2025, initially complete the construction of a low-carbon ecological demonstration zone.
		By 2025, the city’s recycling system for waste materials will be largely established.
Vague	1	Accelerate the green and low-carbon transformation and upgrading of the industrial sector, and continue to promote the phase-out of outdated production capacity.
		Encourage circular transformation of chemical industrial parks to enhance the efficient use of resources.

Machine learning methods

Although humans can understand unstructured information, machines need to learn to understand it³⁷. Therefore, the classification of policy texts first requires text vectorization to achieve the process from sentences to words and then to numerical values, and then uses classifiers in machine learning to classify the texts by score.

Store the manually annotated policy target text in txt in the format of labels, tabs and policy text. TF-IDF (Term Frequency–Inverse Document Frequency), BOW (Bag of Words), and N-gram were selected to vectorize the manually annotated policy texts respectively. Then, MNB (Multinomial Naive Bayes), SVM (Support Vector Machines), and RF (Random Forest) were selected respectively for text classification. TF-IDF considers the importance of words. Specifically, the importance of a word increases proportionally with the number of times it appears in the document, but decreases inversely with the frequency of its appearance in the corpus³⁸. BOW counts the occurrence times of each word in the text based on the word set model³⁹, while N-gram takes into account the order of words⁴⁰. MNB is a Naive Bayes classifier based on the polynomial distribution assumption⁴¹. SVM can map data to a high-dimensional feature space and achieve data partitioning or fitting⁴², RF uses multiple decision trees for classification or regression tasks⁴³. These algorithms have been widely applied due to their excellent performance in the field of text classification.

Policy quantification

The quantification of policy intensity involves two aspects: the diversity of policy instruments and the accuracy of policy targets. Classify according to the keywords of policy instruments and sum up the scores of all policy instrument sentences. The optimized TF-IDF + SVM model combination was used to classify, score and sum up the policy target text. Various approaches exist for the calculation of policy intensity. Schaffrin et al. (2015) proposed the classic “policy density + policy intensity” two-dimensional analytical framework, operationalizing policy intensity across six dimensions of policy instruments: objectives, scope, integration, budget, implementation, and monitoring⁴⁴. This study adopts the bottom-up policy quantification logic introduced by Schmidt and Sewerin (2019), which first decomposes and analyzes individual carbon peaking policy texts, followed by aggregation to construct a multi-level policy intensity dataset⁴⁵. In quantifying agricultural policy texts, Wu et al. define policy intensity as a comprehensive quantitative indicator reflecting the extent of policy support, intervention intensity, and implementation intent embedded within policy texts²⁵. The policy intensity defined in this study is a composite indicator that captures the overall intervention stringency of policy texts along two dimensions: instrument diversity and target specificity. Its calculation follows the method of Zhang et al.²², specifically the product of policy instruments and policy targets, designed to capture the synergistic effect between the two. This indicator enables cross-regional and cross-level policy comparisons by reflecting the stringency and clarity of policy design, as detailed in Formula (3):

$${\mathrm{P}}_{\mathrm{i}}={\mathrm{I}}_{\mathrm{i}}\times {\mathrm{T}}_{\mathrm{i}}$$ 3

P_i is the policy intensity if a policy text, I_i and T_i are respectively the sum of instrument intensity and the sum of target intensity of the same policy text.

It is worth noting that this study focuses on the intensity of textual content rather than policy implementation effectiveness. Policy intensity is explicitly defined here as the diversity of instruments and the specificity of targets reflected in policy texts—an attribute measured at the textual level. Under the condition that policy text types and objectives are highly consistent, text length itself serves as a meaningful proxy for the comprehensiveness and detail of policy content. To identify the best-fitting relationship between policy intensity and both word count and sentence count, seven statistical models were tested: linear, quadratic polynomial, cubic polynomial, quartic polynomial, power law, logarithmic, and exponential. The evaluation metrics used were the coefficient of determination (R²) and mean squared error (MSE), with model robustness ensured through five-fold cross-validation. The results show that, among the seven models, the relationship between policy intensity and sentence count is best explained by a quartic polynomial (R² = 0.9028), while the cubic polynomial provides the optimal fit for the relationship between policy intensity and word count (R² = 0.8413). Both optimal models passed five-fold cross-validation, with average R² values of 0.8410 (sentence count) and 0.7630 (word count), and outlier proportions of 3.2% and 5.4%, respectively. Spearman correlation coefficients between policy intensity and word count, and between policy intensity and sentence count, are 0.9028 and 0.9547, respectively, indicating a highly consistent monotonic increasing trend. We acknowledge that text length is not the sole determinant of policy intensity; however, we consider its influence acceptable when analyzing comparable policies. All policies included in the analysis are carbon peaking implementation plans, which share similar policy orientations, structural frameworks, and discursive conventions. This consistency helps to mitigate potential confounding effects arising from textual heterogeneity in the measurement of intensity.

Data Records

By searching for policy documents publicly released on different government official websites and carbon neutrality websites, we have established a dataset of carbon peaking implementation plan policies officially released by China from the national level to provincial, municipal and county levels from 2021 to 2025⁴⁶. This dataset provides basic information, textual dissimilarity values, and policy intensity scores for 316 carbon peaking policies, stored in Excel format.

The dataset has been deposited in Figshare (10.6084/m9.figshare.31329616.v3), all content in the dataset is publicly accessible. The data files include:

1. A dataset of basic information on 316 carbon peaking implementation plans (.xlsx): Contains fundamental information on 316 policies (e.g., issuing region, issuance date, URL, etc.).

2. A dataset of text difference values of 316 carbon peaking implementation plans (.xlsx): Contains numerical values of policy textual dissimilarity.

3. A dataset of policy intensity for 316 carbon peaking implementation plans (.xlsx): Contains the counts of three types of policy instruments and three types of policy targets, assigned weights, intensity scores, and overall policy intensity values.

4. Policy Text Analysis Code (.ipynb): Contains Python codes used in the research.

Text content differences

As shown in Fig. 3, the number of carbon peaking implementation plan issued reached its peak in the fourth quarter of 2023. With the completion of the 14th Five-Year Plan, the number of carbon peaking policies issued has gradually decreased.

Fig. 3.

The release time and quantity of carbon peaking implementation plans.

A total of 316 policies were merged into 4 policy texts according to their release levels. Keywords were extracted from them using jieba word segmentation, as shown in Table 4. There are 14 keywords in total, including green, carbon peak, carbon emission, carbon reduction, energy, construction, carbon neutrality, development, utilization, system, ecology, focus, building, and project. Among them, green, carbon peak and construction are the key terms of the carbon peaking policy. The word cloud maps of governments at all levels are shown in Fig. 4.

Table 4.

Keywords comparison across different government levels.

Fig. 4.

(a) Key words of the national carbon peaking implementation plan. (b) Key words of provincial carbon peaking implementation plans. (c) Key words of the municipal carbon peaking implementation plans. (d) Key words of county-level carbon peaking implementation plans.

Figure 5 reflects the differences in the content of carbon peaking implementation plan policy texts at the provincial, municipal and county levels. From the average value of each policy level, the difference values among policy texts within provinces, cities and counties have gradually increased (0.084-0.108-0.132). From the perspective of the difference values in provincial policy texts (Fig. 5a), the average difference between Hebei Province and other provinces is the smallest (0.054), while Hainan Province has the largest (0.164). From the perspective of the difference values in municipal policy texts (Fig. 5b), the average difference between Weifang City, Shandong Province and other cities is the smallest (0.073), while Wuxi City, Jiangsu Province, has the largest (0.311). From the perspective of the difference values in county-level policy texts (Fig. 5c), the average difference between Shangyou County in Ganzhou City, Jiangxi Province and other counties is the smallest (0.082), while that between Zhungeer Banner in Ordos City, Inner Mongolia Autonomous Region is the largest (0.290).

Fig. 5.

(a) Province Text Difference Matrix. (Note: The labels on the horizontal and vertical coordinate axes represent the abbreviations of each province. For example, “IM” refers to Inner Mongolia. The corresponding relationship between each province and the abbreviations can be found in The dataset: 1. A dataset of basic information on 316 carbon peaking implementation plans (.xlsx)). (b) City Text Difference Matrix. (Note: The labels on the horizontal and vertical coordinate axes represent the abbreviations of each city. For example, “JS-WX” refers to Wuxi City, Jiangsu Province. The corresponding relationship between each city and the abbreviations can be found in The dataset: 1. A dataset of basic information on 316 carbon peaking implementation plans (.xlsx)). (c) County Text Difference Matrix. (Note: The labels on the horizontal and vertical coordinate axes represent the abbreviations of each county. For example, “JX-GZ-SY” refers to Shangyou County, Ganzhou City, Jiangxi Province. The corresponding relationship between each county and the abbreviations can be found in The dataset: 1. A dataset of basic information on 316 carbon peaking implementation plans (.xlsx)).

Policy intensity

From the perspective of the average proportion of various policy instruments, the proportion of legal policy instruments at the national, provincial, municipal and county levels decreases in sequence (Fig. 6). National-level policies employ the most legal instruments, while provincial, municipal and county policies more often use market instruments for regulation. The proportion of supervision policy instruments is the lowest among all government levels.

Fig. 6.

The average proportions of policy instrument.

From the perspective of the average proportion of various policy targets, the precise policy targets at the national, provincial, municipal and county levels are all less than 10% (Fig. 7), and the average proportion of precise and general policy targets at the provincial, municipal and county levels is higher than that of national policy. Overall, municipal governments have the highest precision in setting policy targets.

Fig. 7.

The average proportions of policy target.

As shown in Table 5, on average, the intensity of policy instruments and the intensity of policy at the national, provincial, municipal and county levels all show a decreasing trend in sequence, while the policy targets are different. The intensity of policy targets at the provincial level exceeds that at the national level. From the perspective of the average policy intensity within the province (Fig. 8), among the 29 provinces that have publicly released provincial, municipal or county-level carbon peaking implementation plans, only 5 provinces have an average that exceeds the national policy, namely Liaoning, Inner Mongolia, Hainan, Guizhou and Chongqing.

Table 5.

Average policy intensity under different policy levels.

Policy level	Instrument intensity	Target intensity	Policy intensity
Country	129.00	351.00	45279.00
Province	109.71	371.17	42053.63
City	88.86	342.48	33071.13
County	88.51	331.65	31997.06

Fig. 8.

Policy intensity is derived from the product of policy instrument intensity and policy target intensity, and is dimensionless. The average policy intensity within the province. (Note: The labels on the horizontal and vertical coordinate axes represent the abbreviations of each province. For example, “IM” refers to Inner Mongolia. The corresponding relationship between each province and the abbreviations can be found in The dataset: 1. A dataset of basic information on 316 carbon peaking implementation plans (.xlsx)).

Technical Validation

Model performance

In this study, we combined three text vectorization models and three classifiers respectively, and used 75% of the manually labeled policy target text for training and the remaining 25% for validation. The effect of policy text classification is achieved by averaging the weights of four indicators: Accuracy, Precision, Recall and F1 Score. Among these, Accuracy refers to the proportion of correctly classified samples out of the total number of samples, as detailed in Formula (4). Precision is defined as the proportion of samples predicted as positive that are actually positive; a high precision indicates a low false positive rate, with an emphasis on minimizing misclassification, as shown in Formula (5). Recall denotes the proportion of actual positive samples that are correctly predicted as positive; a high recall reflects a low false negative rate, with an emphasis on minimizing omissions, as specified in Formula (6). The F1 Score is the harmonic mean of precision and recall, jointly considering the reliability and coverage of model predictions, and serves as a comprehensive metric for evaluating model performance, as presented in Formula (7). The performance of the nine model combinations is reported in Table 6.

$$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}+\mathrm{TN}}$$ 4

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$ 5

$$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$ 6

$$\mathrm{F}1\mathrm{Score}=2\times \frac{\mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$ 7

Table 6.

Training results of policy target texts.

Vectorizer	Classifier	Accuracy	Precision	Recall	F1
TF-IDF	MNB	0.78	0.80	0.78	0.75
	RF	0.83	0.84	0.83	0.81
	SVM	0.85	0.85	0.85	0.84
BOW	MNB	0.84	0.83	0.84	0.84
	RF	0.83	0.84	0.83	0.82
	SVM	0.87	0.87	0.87	0.86
N-gram	MNB	0.84	0.84	0.84	0.83
	RF	0.82	0.83	0.82	0.79
	SVM	0.86	0.86	0.86	0.86

TP (True Positives)

The number of samples that are actually positive and are correctly predicted as positive by the model.

TN (True Negatives)

The number of samples that are actually negative and are correctly predicted as negative by the model.

FP (False Positives)

The number of samples that are actually negative but are incorrectly predicted as positive by the model.

FN (False Negatives)

The number of samples that are actually positive but are incorrectly predicted as negative by the model.

By comparing the experimental results of the nine model combinations, it can be seen that SVM performs exceptionally well in this text classification. After optimizing the grid search parameters for the three combinations of TF-IDF + SVM, BOW + SVM, and N-gram + SVM respectively, overfitting occurred in BOW + SVM, and N-gram + SVM, and their accuracy rate, precision rate and other indicators all decreased (Table 7), indicating that their generalization ability on the test data is poor. Therefore, in this paper, the TF-IDF vectorization method and the SVM classifier are selected as the final model combination. Grid Search CV is used for grid search, and param_grid is specified as the parameter grid to be searched. 5-fold cross-validation is used to evaluate the performance of each parameter combination, and the test set is evaluated. The optimal combination of hyperparameters of the SVM classifier is shown in Table 7 below, including the regularization parameter C, the kernel function coefficient gamma, and the kernel function type is selected as the Radial Basis Function Kernel. This model combination is used to evaluate the precision of policy targets.

Table 7.

Model optimization performance.

Model combination	Report				Best parameters found
Parameter	Accuracy	Precision	Recall	F1	C	Gamma	Kernel
TF-IDF + SVM	0.87	0.87	0.87	0.87	100	0.1	RBF
BOW + SVM	0.87	0.86	0.87	0.86	10	0.01	RBF
N-gram + SVM	0.86	0.86	0.86	0.86	100	0.001	RBF

The final optimized model combination (TF-IDF + SVM) was evaluated using five-fold cross-validation. All 99 annotated policy target documents used for model training and testing were treated as a consolidated dataset and randomly partitioned into five mutually exclusive subsets. In each iteration, four subsets were used for training and the remaining subset for testing. This process was repeated five times to ensure that every subset served as the test set exactly once. The detailed results are presented in Table 8 below. The five-fold cross-validation results demonstrate that the TF-IDF + SVM model performs robustly and consistently in the classification of carbon peaking policy targets, achieving an average accuracy of 83.5%. The performance variation across folds is minimal, with a standard deviation of only 0.0048 and a maximum–minimum accuracy gap of just 1.4%.

Table 8.

Five-fold cross-validation results.

Fold	Accuracy	Precision	Recall	F1
1	0.83665	0.83621	0.83665	0.83592
2	0.83754	0.83978	0.83754	0.83835
3	0.84051	0.84098	0.84051	0.84054
4	0.83190	0.83082	0.83190	0.83102
5	0.82685	0.82804	0.82685	0.82738
Mean	0.83469	0.83517	0.83469	0.83464
Std	0.00480	0.00502	0.00480	0.00482
Min	0.82685	0.82804	0.82685	0.82738
Max	0.84051	0.84098	0.84051	0.84054

Comparison with existing studies

The databases of Dong et al. in the field of low-carbon policies²¹ and Zhang et al. in the field of environmental policies²² were selected as the comparison objects. In terms of time, this study fills the gap of special policies for carbon peaking from 2021 to 2025. Spatially, 142 county-level policy documents have been included, expanding the policy coverage of the existing database during the carbon peaking stage. The detailed results are presented in Table 9.

Table 9.

Comparison with the existing database.

Comparative dimension	This study	Dong et al.21	Zhang et al.22
Time range	2021–2025	2007–2022	1978–2019
Policy levels	National/provincial/municipal/county	National/provincial/municipal	Mainly at the central level, with a small number at the provincial level
Policy number	316: for the carbon peaking special project	7282: for the manufacturing special project	1912: for the comprehensive environmental policy
Quantified dimension	Total intensity + instrument/target sub-intensity	Total intensity + level/instrument/target sub-intensity	Total intensity (measure × target)
Model combination	TF-IDF + SVM	ERNIE3.0 (Prompt Learning)	Random Forest

Manual checking

To verify the accuracy of the quantification results generated by the TF-IDF + SVM model, while mitigating potential biases arising from single-level sampling and subjective interpretation, a stratified random sampling approach was adopted. Fifty samples were randomly selected from the 316 policy target documents. Three scholars with extensive experience in environmental policy research were invited to independently score the 50 samples in accordance with the predefined scoring criteria, following a standardized training session. The intraclass correlation coefficient ICC(3,1) was computed to assess inter-rater consistency. The formula for ICC(3,1) is as follows:

$$\mathrm{ICC}\left(3,1\right)=\frac{{\mathit{MS}}_R-{\mathit{MS}}_E}{{\mathit{MS}}_R+{\left(k-1\right)\mathit{MS}}_E+\frac{k}{n}{(\mathit{MS}}_C-{\mathit{MS}}_E)}$$ 8

In the formula, MS_R denotes the mean square for raters, MS_C denotes the mean square for samples, MS_E denotes the mean square for error, k represents the number of raters (k = 3), and n represents the number of samples (n = 50).

The consistency test results show that ICC = 0.90 (>0.85), indicating good consistency of expert scores. The average score of the three individuals can be taken as the final manual quantification result and used as the benchmark for comparison with the model results. Table 10 presents the prediction results of the TF-IDF + SVM model and the manual annotation results. The average relative error of the 50 samples is 7.84%.

Table 10.

Manual verification result. MAE refers to the mean absolute error, and MRE refers to the mean relative error.

Policy level	Sample size	Precise target		General target		Vague target
		MAE	MRE	MAE	MRE	MAE	MRE
Country	1	2.00	9.52%	5.00	15.15%	3.00	1.36%
Province	5	0.40	2.79%	3.20	7.53%	3.60	1.92%
City	22	1.86	7.27%	7.41	13.36%	7.82	4.49%
County	22	1.62	7.55%	5.52	12.39%	6.52	3.96%
Total	50	1.66	7.19%	6.12	12.39%	6.60	3.94%

Limitations and future work

This study calculates the policy intensity based on the publicly released carbon peaking implementation plans and quantifies it starting from the policy texts. The policy data collected in this study exhibit limited coverage and spatial representativeness, with county-level policies accounting for less than 5% of the total. Moreover, the policy documents are predominantly concentrated in East, Central, and North China, resulting in an imbalanced geographical distribution. As such, the dataset may not fully capture the actual policy formulation landscape of all regions—particularly those that are less developed or characterized by lower policy transparency. Accordingly, caution is warranted when using this dataset for cross-regional comparisons or for making inferences about national-level conditions, due to the potential influence of sample selection bias.

This dataset can serve as a foundation for various dimensions of quantitative research in the future. For instance, in assessing policy implementation effectiveness, policy intensity can be employed as a core explanatory variable and combined with outcome indicators such as carbon emissions, energy structure, and industrial composition to empirically examine the emission reduction performance of carbon peaking policies. In the study of policy diffusion and intergovernmental interaction, the policy textual dissimilarity matrix can be leveraged to explore the diffusion pathways and imitation mechanisms of carbon peaking policies across provincial, municipal, and county levels.

Author contributions

H.L. proposed the idea for this research, and M.Z. designed and generated the dataset. M.Z. and H.L. have stipulated the classification and scoring criteria for policy instruments and policy targets. H.L. and Y.W. provided technical guidance and data verification. All authors made significant contributions to the editing and revision of the manuscript and determined the final version of the manuscript.

Code availability

The relevant code data used for calculation and analysis in this paper can be obtained from the dataset (10.6084/m9.figshare.31329616.v3).

Competing interests

The authors declare no competing interests.