NLP-based Quantification of Low-Carbon Policies in China's Manufacturing Sector

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Feb 20 2024

In an article published in the journal Nature, researchers focused on quantifying low-carbon policies in China's manufacturing industries using natural language processing (NLP) and text-based prompt learning, addressing the deficiency in direct and comprehensive measurement.

The authors constructed a low-carbon policy intensity index using a low-carbon policy inventory covering national, provincial, and prefecture levels from 2007 to 2022. This index considered policy level, objectives, and instruments, enhancing the precision of policy quantification. The resulting dataset, available in two formats, provided valuable information for multidisciplinary researchers interested in merging with macro and micro data related to low-carbon performances.

Background

In the pursuit of carbon neutrality, achieving low-carbon transformation in manufacturing industries is crucial for sustainable development. Previous studies on low-carbon policies often lacked direct and comprehensive quantification, relying on proxy variables or focusing on national-level indices. This paper addressed this gap by constructing a comprehensive low-carbon policy intensity index for China's manufacturing industries from 2007 to 2022. The research innovatively incorporated policy level as a factor, deepening the understanding of low-carbon policy intensity. Utilizing a dataset of 7282 national, provincial, and prefecture-level policies, the index was quantified by multiplying policy objectives, instruments, and levels, following a "policy objective-policy instrument" pattern.

The methodology combined a phrase-oriented NLP algorithm and text-based prompt learning, providing a few-shot learning approach suitable for policy texts with smaller samples. This minimized manual labeling costs and reduced biases, presenting a novel paradigm in the NLP field. The authors further aggregated sub-intensity indices for various policy levels, objectives, and instruments, offering a comprehensive dataset for future research. This research contributed significantly to the literature by enhancing the meaning of low-carbon policy intensity, introducing an innovative NLP-based methodology, and providing a detailed dataset for multidimensional analysis of the impact of low-carbon policies on manufacturing industries.

Method

The researchers focused on constructing a comprehensive low-carbon policy intensity index for China's manufacturing industries from 2007 to 2022. The framework consisted of six modules. The data preparation stage involved building a low-carbon policy inventory with 7282 policies sourced from national, provincial, and prefecture levels. Policies were selected based on keywords related to "carbon reduction," "greenhouse gases," and other factors directly or indirectly linked to low carbon.

Structuring policy texts involved breaking down policy documents into different parts, including titles, backgrounds, objectives, instruments, issuing institutions, and publication years. These structured texts were then disaggregated into separate files for policy objectives and instruments. Policy classification followed a pattern of "policy objective-policy instrument," considering factors such as policy level, objectives (carbon reduction, energy conservation, capacity utilization, technology), and instruments (command-and-control, market-based, composite).

Quantifying low-carbon policy intensity integrated these factors by multiplying policy level, objective, and instrument intensities. Manual labeling was employed for model training, where policies were rated on a scale of 1 to 3 based on their intensity in different dimensions. The research used prompt learning with the ERNIE 3.0 model for predicting policy intensity and addressing data sparsity and class imbalance through data augmentation strategies.

The resulting dataset included a low-carbon policy inventory, intensities for each policy, and aggregated intensities at national, provincial, and prefecture levels. This dataset, provided in Stata and Excel formats, offered valuable insights into the heterogeneous nature of low-carbon policy intensity across policy levels, time, and space. Researchers could access the dataset on Figshare for multidisciplinary analyses and further exploration of the impact of low-carbon policies on China's manufacturing industries. 

Technical Validation and Usage Notes

The authors employed prompt learning to train models and predict the intensity of low-carbon policy objectives and instruments based on a dataset of 7282 policies spanning 2007 to 2022. Of this dataset, 3334 labeled policies were used for model training, with two data augmentation strategies applied to address data sparsity and class imbalance. The model's accuracy in predicting policy intensities for carbon reduction, energy conservation, capacity utilization, technology objectives, and command-and-control and market-based instruments was presented. The study compared its national-level low-carbon policy intensity with existing environmental policy indices, demonstrating a high degree of trend similarity. Manual checking further validated the accuracy of prompt learning results, with accuracies ranging from 0.85 to 0.93.

The dataset, provided in both Stata and Excel formats, included a comprehensive low-carbon policy inventory, policy intensities for individual policies, and aggregated intensities at national, provincial, and prefecture levels. Users were encouraged to interpret the policy intensity index received by policy-takers, considering the multi-layered policy structure in China. The dataset also offered information on the starting and, when available, ending years of each policy. Despite some limitations, such as reliance on publicly available policy texts and focus on release rather than execution, this study contributed valuable insights for future investigations into the relationship between policy intensity and carbon reduction efforts.

Conclusion

In conclusion, the researchers successfully addressed the deficiency in directly and comprehensively measuring low-carbon policies in China's manufacturing industries. The innovative construction of a low-carbon policy intensity index, considering policy level, objectives, and instruments, contributed significantly to understanding policy quantification. The combination of phrase-oriented NLP algorithms and prompt learning introduced a novel paradigm, minimizing manual labeling costs.

The resulting dataset, encompassing 7282 policies, offered valuable insights for multidisciplinary researchers interested in exploring the impact of low-carbon policies on China's manufacturing industries. Despite limitations, this study provided a foundation for future investigations into the nuanced relationship between policy intensity and carbon reduction efforts.