Attention recurrent cross-graph neural network for selecting premises

• Published in: International journal of machine learning and cybernetics Vol. 13; no. 5; pp. 1301 - 1315
• Main Authors: Liu, Qinghua, Xu, Yang, He, Xingxing
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.05.2022; Springer Nature B.V
• Subjects: Artificial Intelligence; Automation; Complex Systems; Computational Intelligence; Control; Datasets; Embedding; Engineering; Exchanging; Experiments; Graph neural networks; Graph representations; Graphical representations; Mechatronics; Messages; Neural networks; Nodes; Original Article; Pattern Recognition; Robotics; Syntax; Systems Biology; Theorem proving; Premise selection; Automated theorem proving; Graph neural network; Node types; Cross-graph information
• ISSN: 1868-8071; 1868-808X
• DOI: 10.1007/s13042-021-01448-9

Premise selection is a task that selects likely useful premises from a large-scale premise set, which is one of the challenges for automated theorem proving. Nowadays, using graph neural networks to encode logical formulae becomes more and more appealing, as graph representations demonstrate the ability to preserve the semantic and syntactic information of logical formulae. However, graph neural networks in the prior works iteratively update each node’s embedding via information aggregated solely from its neighbors and always ignore the types of nodes. Besides, evidence shows that sharing and exchanging some information between formula pairs can ensure neural-based models more robust in measuring graph relevance. Unluckily, previous graph neural networks generate final graph embeddings independently in the premise selection task. To overcome these shortages, we propose a novel graph neural network, called ARCG-NN, for embedding the first-order logical formulae. The embedding model firstly takes the node types into full consideration and utilizes an attention mechanism based on newly proposed node types to compute weights in the message aggregation. Except for the message from the local neighborhood, ARCG-NN also dynamically exchange the cross-graph information of particular nodes along with the propagation rounds. Besides, an extra gate function on node types is used to filter out irrelevant information in the graph aggregation phase. To train, validate, and test our approach, we build balanced and unbalanced datasets based on the MPTP2078 benchmark. The experimental results demonstrate that the proposed graph neural architecture achieves state-of-the-art classification accuracy on the first-order premise selection task and helps the automated theorem prover E prove more problems in the test dataset.