Quantum-Inspired Neural Language Representation, Matching and Understanding.

Peng Zhang,Hui Gao,Jing Zhang,Dawei Song

DOI: https://doi.org/10.1561/1500000091

2023-01-01

Abstract:The introduction of Quantum Theory (QT) provides a unified mathematical framework for Information Retrieval (IR). Compared with the classical IR framework, the quantum-inspired IR framework is based on user-centered modeling methods to model non-classical cognitive phenomena in human relevance judgment in the IR process. With the increase of data and computing resources, neural IR methods have been applied to the text matching and understanding task of IR. Neural networks have a strong learning ability of effective representation and generalization of matching patterns from raw data. However, these methods show some unavoidable defects, such as the inability to model user cognitive phenomena, large number of model parameters and the "black box" characteristics of network structure. These problems greatly limit the development of neural IR and related fields. Although the quantum-inspired retrieval framework can theoretically solve the above problems, it is faced with problems such as poor model efficiency and difficulty in integrating with neural network, which lead to a huge gap between QT and neural network modeling. This review gives a systematic introduction to quantum-inspired neural IR, including quantum-inspired neural language representation, matching and understanding. This is not only helpful to non-classical phenomena modeling in IR but also to break the theoretical bottleneck of neural networks and design more transparent neural IR models. We introduce the language representation method based on QT and the quantum-inspired text matching and decision making model under neural network, which shows its theoretical advantages in document ranking, relevance matching, multimodal IR, and can be integrated with neural networks to jointly promote the development of IR. The latest progress of quantum language understanding is introduced and further topics on QT and language modeling provide readers with more materials for thinking.