Quantum Probability-inspired Graph Attention Network for Modeling Complex Text Interaction

Inspired by quantum-like phenomena in human language understanding, recent studies propose quan-tum probability-inspired neural networks to model natural language by treating words as superposition states and a sentence as a mixed state. However, many complex natural language processing tasks (e.g., emotion-cause pair extraction or joint dialog act recognition and sentiment classification) require modeling the complex and graphical interaction of multiple text pieces (e.g., multiple clauses in a document or multiple utterances in a dialog). The existing quantum probability-inspired neural networks only encode sequential interaction of a sequence of words, but cannot model the complex interaction of text pieces. To generalize the quantum framework from modeling word sequence to modeling complex and graphical text interaction, we propose a Quantum Probability-inspired Graph Attention NeTwork (QPGAT) by combining quantum probability and graph attention mechanism in a unified framework. Specifically, a text interaction graph is firstly constructed to describe the complex interaction of text pieces. Then QPGAT models each text node as a particle in a superposition state and each node's neighborhood in the graph as a mixed system in a mixed state to learn interaction-aware text node representations. We apply QPGAT to the two important and complex NLP tasks, emotion- cause pair extraction and joint dialog act recognition and sentiment classification. Experiment results show that QPGAT is competitive compared with the state-of-the-art methods on the two complex NLP tasks, demonstrating the effectiveness of QPGAT. Moreover, QPGAT can also provide a reasonable post-hoc explanation about the model decision process for emotion-cause pair extraction. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

Inspired by quantum phenomena in human language understanding, a Quantum Probability-inspired Graph Attention Network (QPGAT) is proposed to model complex and graphical text interaction, showing competitive performance in emotion-cause pair extraction and joint dialog act recognition tasks.