TL;DR
This paper proposes a novel system that combines blockchain technology with an improved Long Short-Term Memory (LSTM) neural network to enhance the accuracy and trustworthiness of sentiment analysis on e-commerce reviews. By leveraging blockchain for secure and transparent data storage and an enhanced LSTM for nuanced sentiment detection (including idioms), the system aims to provide regulators with reliable data for more effective market oversight and sustainable development.
What Problem Does This Paper Address?
The rapid growth of e-commerce has led to an explosion of user-generated reviews. Analyzing these reviews offers valuable insights for consumers, merchants, and regulators alike. However, traditional sentiment analysis methods face significant challenges: the unreliability and susceptibility to tampering of review data, and the inherent ambiguity and complexity of natural language, especially when idioms are involved. These issues can lead to inaccurate analyses, potentially resulting in flawed regulatory decisions. The paper aims to address this by creating a system that ensures data integrity and improves the sophistication of sentiment analysis, thereby enabling more effective regulation for a sustainable market.
Methods
The core of the proposed method lies in the synergistic integration of two key technologies:
- Blockchain for Data Integrity: The system utilizes a blockchain platform (specifically Ethereum) to store transaction and review data. This ensures that all data is transparent, immutable, and accessible to all nodes, thereby establishing a trustworthy source of information. Smart contracts are employed to manage transactions, user interactions, and data recording.
- Improved LSTM for Sentiment Analysis: A Long Short-Term Memory (LSTM) neural network is employed for sentiment analysis due to its effectiveness in processing sequential data like text. The authors introduce an enhancement to the standard LSTM by incorporating an "external memory component." This component is designed to specifically handle idioms, which often pose challenges for traditional sentiment analysis models. By memorizing idiom semantics, the model can better interpret nuanced language and improve sentiment classification accuracy. The architecture includes an embedding layer, the enhanced LSTM layers, and a dense output layer.
Datasets
The study utilizes both English and Chinese review datasets for training and evaluation. The datasets were collected from existing research and filtered. Specifically, the authors mention a total of 179,396 collected reviews, with the English and Chinese datasets balanced to 89,698 each. Positive reviews were labeled as '1' and negative reviews as '0'.
Key Insights and Ideas
- Blockchain as a Trust Anchor: The paper highlights blockchain's crucial role in establishing a reliable foundation for sentiment analysis by guaranteeing data authenticity and immutability, which is essential for regulatory purposes.
- Idiom Handling is Key: The authors demonstrate that explicitly addressing idioms through an external memory component significantly improves sentiment analysis accuracy, moving beyond simple word-based analysis.
- Integrated System for Regulation: The proposed framework seamlessly integrates data collection, sentiment analysis, and regulatory action, creating a closed-loop system for market oversight.
- Dual-Channel Structure for Enhanced Learning: The LSTM's external memory component creates a dual-channel structure that not only improves representation capability but also enhances extensibility for future knowledge integration.
Key Results
The experimental results demonstrate the effectiveness of the proposed method:
- Performance: The improved LSTM model achieved high accuracy, with results approaching 90% on both English and Chinese datasets when the epoch was set to 3.
- Comparison: The proposed method outperformed traditional models like Support Vector Machine (SVM) and Naive Bayes in terms of Accuracy and F1-score, especially on larger datasets. Statistical significance tests confirmed these differences.
- Scalability: The model's performance generally improved with increasing dataset sizes, indicating good scalability for large-scale review data.
- Case Study: A case study involving 38 merchants showed that the system could effectively identify merchants with a high proportion of negative reviews, providing a basis for targeted regulation. For instance, Merchant B had no warning messages, indicating good reputation, while Merchant A received six, signaling dissatisfaction.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Blockchain Throughput: The current research is constrained by the inherent limitations of blockchain throughput and block generation speed.
- Cross-Lingual Performance: Further attention is needed to investigate the performance of the proposed method when processing review data in different languages.
- Sampling Errors: While statistical tests were performed, the authors note the need to avoid sampling errors that could lead to inaccuracies in analysis.
- Future Research: The paper suggests that addressing the aforementioned limitations will be the focus of future research.
Summary updated