Abstract: Cloud-based Automatic Speech Recognition (ASR) services generate vast quantities of user-generated data. This paper proposes a framework to transform this data into personalized knowledge services. The authors introduce the Quadruple Transfer Learning (QTL) algorithm, which mitigates "negative transfer" by distinguishing between shared and non-shared concepts across domains. Coupled with a high-performance distributed architecture featuring a custom caching layer (HDCache), the system enables real-time personalized model loading. Evaluation on real-world datasets shows that QTL significantly outperforms existing transfer learning methods in sentiment and expertise prediction.
Introduction
The proliferation of cloud-based ASR services, such as Siri and iFlyTek, has resulted in a massive influx of user-generated voice and text data. While this data represents a significant opportunity for personalization, effectively utilizing it across different application domains remains challenging. This paper presents a robust framework for turning this data "exhaust" into a valuable knowledge service, addressing the cold start problem and the risk of negative transfer.
The Challenge of Negative Transfer
Traditional transfer learning often suffers from "negative transfer," where knowledge from a source domain degrades performance in a target domain. This occurs because concepts can change meaning across contexts. For instance, the term "imaginative" might be positive in a movie review but negative in a scientific critique. The authors address this by explicitly modeling the relationships between shared and distinct concepts.
Quadruple Transfer Learning (QTL)
The proposed QTL algorithm utilizes non-negative matrix tri-factorization to decompose data into four specific concept types. By isolating these categories, the model can selectively transfer only relevant knowledge, thereby avoiding the pitfalls of ambiguity.
System Architecture and Data Collection
The system is designed to handle approximately 20 terabytes of incremental daily data. It consists of two primary subsystems:
- Model Production (MP) Subsystem: An offline pipeline using Spark and Spark MLlib for heavy-duty training. It incorporates crowdsourcing with specialized consensus algorithms to label massive datasets efficiently.
- Real-Time Service (RS) Subsystem: This subsystem ensures low-latency performance using a custom distributed cache called HDCache. This allows the system to retrieve and load user-specific models in roughly 30 milliseconds, ensuring the total round-trip time for ASR remains under 2 seconds.
Experimental Results
The QTL method was evaluated using real-world data from Smart TV and E-Book domains. The experiments focused on transferring knowledge from the established Smart TV domain to the newer E-Book domain to predict user sentiment and expertise.
- Sentiment Classification: QTL achieved 97.52% accuracy, a 3.77% improvement over the Tri-TL baseline.
- User Experience Prediction: QTL reached 75.29% accuracy, outperforming the runner-up by 3.26%.
- System Performance: The HDCache implementation successfully maintained model loading times at an average of 30ms, proving the feasibility of real-time personalized cloud services.
Conclusion
The framework demonstrates that user-generated data can be effectively repurposed for cross-domain personalization. By categorizing concepts and optimizing infrastructure for low latency, the authors provide a scalable path toward "Knowledge as a Service" (KaaS). Future work will focus on integrating a wider variety of algorithms and addressing the complexities of diverse dialects.
Summary updated