Fig.1 shows an overview of our methods. The research framework includes four modules: Manual quantification and Text data preparation, Modeling, Validation, and Modeling.
Research framework. (1) Manual Quantification: Each policy is first read and scored manually. We then categorize policies according to their attributes. (2) Text Data Preparation: We analyze the text content using a variety of text-mining operations, such as cleaning, tokenization and stop-word removal, word frequency analysis and feature extraction from text. The specific lexicon for environmental policy is screened, and then constructed. (3) Modeling: To quantify the environmental policy’s intensity, we use a number of prediction models. (4) ValidationWe evaluate and compare models based their performance and key features to validate the credibility and validity of the results.
Data collection
We first collect environmental policy from the Global Legal and Regulatory Network.http://policy.mofcom.gov.cn/), China Legal Resources Database (http://www.lawyee.org), Wanfang Database (http://c.g.wanfangdata.com.cn), China National Knowledge Infrastructure (CNKI, https://www.cnki.net/), PKULaw.com database (https://pkulaw.com/), the official websites of the China and its ministries, etc. Keywords like energy savings, emission reduction, energy conservation and reducing pollutant omissions were used to search and gather environmental policies jointly or separately promulgated from 1978 to 2019.
The policy text was then carefully read, taking into consideration the policy background, release date and issuing institution, as well as the policy type, policy objectives, screening criteria, and policy measures. After much discussion, classification, discussion, screening, and long periods of collation, a dataset of China’s environmental policies was finally created. The dataset includes 1912 environmental regulation policies from more than 40 agencies. Table 1 shows some of the departments that promulgated policies.
Manual quantification
Combinations with pre-set dimensions like policy objectives and measures were the main components of manual environmental policy intensity quantification. Each policy is rated by its intensity on a scale from 1 to 5, based on how enforceable and detailed the different policy objectives and measures are.25,26. Policy measures include administrative measures, personnel measures and fiscal and tax measures. Financial measures, guiding and other economic measures are also included. These policy objectives include preventing pollution, increasing the effectiveness of emission reduction and energy conservation, raising awareness about energy conservation, optimizing energy consumption, optimizing energy use, and promoting technological transformations in energy conservation.
A group of staff was trained to manually read each policy and rate it for its intensity and objectives. Each policy was independently rated and verified for interter reliability by multiple raters. Not only do the ratings reflect the extent to which the policy emphasizes specific measures or objectives, but they also help to solve the problem associated with weight selection when building indicators. This is how you calculate the intensity of an environmental policy:
$$begin{array}{c}ERI={M}_{tik}{O}_{tin}end{array}$$
(1)
Where tRepresents the year. iThis is a policy. MtikIt is the sum of all the intensity of kPolicies in a specific year OTinThe sum of all the intensity of nPolicy objectives for a specific year.
Our personnel have rated the ERI scores as an important and reliable resource for future research on China’s environmental policy. These ERI scores, together with the policy text can be used to train an algorithm that can estimate future environmental policies’ intensity.
Types of environmental policy
The policy instrument is the carrier for policy. It allows policy scientists to study the main content, the policy formulation process, and the policy tools. It also serves as an objective, traceable, and accessible written record of the policy system.30,31. The majority of research on environmental policies focuses on policy tools, types and game behaviours among central bodies of government at all levels of implementation32,33. The World Bank, for example, divides environmental policy tools in four types: market application, market creation and environmental regulations34. Based on the driving mechanism, environmental policy can be classified into command-control policy or market-based policy35,36,37,38,39,40,41.
We used related research to divide environmental policy into three types. command-control environmental policy (CCEP), Market-based environmental policy(MBEP) public participation environmental policy (PPEP). Literature has thoroughly investigated MBEP as well as CCEP. CCEP has one main characteristic: it is mandatory. Therefore, it relies on administrative instruments like certain types and standards of governance. MBEP, however, focuses primarily on market-based instruments to ensure environmental governance. This includes fiscal policies related environmental governance, emissions fees, emission trading, product promotion catalogs, etc.
PPEP refers both to the public and private sector that has the right to participate environmental protection. 2015’s Environmental Protection Act established the principle of public participation under the General Regulations section. All individuals and organizations have an obligation to protect the environment. They also have the right to report on and accuse those who pollute or destroy it.42. Therefore, we place an environmental policy within the category of public-participationIf it contains terms like Participation by the public, citizens, Opinions, Get advice, Supervision, Hearing, Argument, subject declaration.
Pre-processing the text data
Structured data is used to enrich and supplement the encoded text. Many studies have been conducted on text in recent years. These include financial news, social media, speech by politicians and corporate documents.43,44,45,46. The most popular method for text analysis is the bag of words model, which is based on the document word matrix. Researchers have discussed the practical applications of different word lists and specific words in this field.47. But, Loughran et al.44Many studies rely on classifications dictionaries that are derived from other disciplines. These applications can produce false results. The problem of choosing or building a suitable lexicon for a specific field is one that must be addressed before research can begin.
This was avoided by the creation of a lexicon specifically applicable to environmental policy during the processing textual data. This paper describes several steps in the text preprocessing process, including text cleaning, tokenization, word frequency statistics, and text cleaning. The term-document matrix includes all words in policy documents. Many words are not important to policy intensity. It is therefore necessary to screen the words before creating a specific vocabulary.
-
Words that appear in more then 99% of documents and less than 1% of them are screened and removed.
-
The correlation coefficients between word frequency, policy intensity (manual quantification intensity) are calculated.
-
Based on the coefficients results, words with a high correlation (>0.1) to the intensity are chosen.
-
The word selected is divided into two parts, policy objectives or policy measures. The words that represent objectives and measures are not repeated.
-
Words that aren’t substantive, such names of people, places, and departments, are deleted.
Machine learning methods
In the literature, the use text language to calculate indicators is steadily increasing48,49,50. This combination of machine learning tools and this method allows for verification and testing of predicted scores based on specific structures. Harrison, for example. et al.51A new language-based method was developed to measure five personality traits of CEOs. First, we use the text analysis method in order to create a lexicon that is suitable for environmental policy. Second, we create a model that explores the relationship between the policy intensity and the lexicon. Due to the nature of this task, both traditional measurement methods as well as machine learning algorithms can be used. This paper presents the traditional linear regression model. Eight groups of algorithm models are also selected: Ridge (Ridge), Lasso(Lasso), robust line model (RLM), partial most squares (PLS), generalized line model (GLM), support Vector Machine (SVM), eXtreme Graduation Boosting (XgbLinear) and random forest (RF).
Ridge and Lasso shrinkage methods for prediction models are particularly useful when there are many explanatory variables.52. GLM and RLM loosen the conditions of least-squares regression, making them more able to deal with outliers and external variances. PLS projects prediction and observation variables to a new area to find a linear equation. This method is ideal for high-dimensional data. SVM is also suitable to deal with high-dimensional information and find fitted curves. XgbLinear is primarily tree-based, while RF is primarily tree-based. These are both popular and effective methods for estimating regression functions flexiblely when out-of-sample predictions are essential.53. Many of the algorithm representative have also performed well in other fields and can therefore be used to support this research.
