Short Text Mining Technology for Plastic Industry Integrating Text Clustering and Sentiment Analysis

The transition to the circular economy aims to replace the traditional linear "take-make-dispose" model with a regenerative system focused on reuse, recycling, and sustainable resource management. Achieving this transition is challenging and complex, requiring both systemic change and innovative technological solutions. One of the most visible and problematic materials in the circular economy is plastic—an essential but raising significant sustainability concerns for the environment. Public opinion plays a key role in shaping policies and practices in the circular economy. The analysis of public opinion is essential because it can offer a valuable and accurate reflection of the attitude of the public, which will support the decision-making of the related stakeholders.This thesis proposes a text-mining framework rooted in the analysis of public opinion for the plastic industry, by integrating text clustering and sentiment analysis. First, we developed a web-crawler system to collect user comments related to plastic recycling. The system targets discussions under news articles and extracts a rich dataset of short texts that reflects a diverse range of public opinions. Subsequently, a method combining text classification and sentiment analysis is proposed to automatically identify and extract user opinions. This framework analyses sentiment changes across different topics and time frames, which helps to understand the changes in public attitudes towards different aspects of plastic use and policy. While analysing the process, we found that following certain paths can help improve the accuracy and efficiency of this analysis. To break through the limitations of traditional classification models on short text classification, we proposed an innovative method that uses Latent Dirichlet Allocation (LDA) for topic modelling and Bidirectional Long Short-Term Memory (Bi-LSTM) networks for sequence learning. This hybrid model fully exploits the advantages of unsupervised topic discovery and deep contextual embedding. To improve the accuracy of sentiment analysis, this paper adopts a dynamic sentiment lexicon approach. A new sentiment analysis method is implemented using a dynamic, domain-specific lexicon.This research highlights the potential of text mining with advanced machine learning models for domain-specific knowledge with a case study in the plastic industry. The proposed framework can be applied to any domain or industry to offer accurate, dynamic and timely insights for public opinion.