Neural Architecture Search Based on Bipartite Graphs for Text Classification.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Neural Architecture Search (NAS) is vital for text representation in NLP but less explored than in computer vision.
• Existing NAS methods for text classification often use Directed Acyclic Graphs (DAGs), limiting generalization and increasing computational cost.
• DAGs struggle to accurately represent the essential topological order of NAS operators for text classification.

Purpose of the Study:

• To propose a novel Bipartite Graph-based NAS (BGNAS) framework for text classification.
• To address the limitations of DAG-based search spaces in capturing topological order and enhancing generalization.
• To develop a computationally efficient NAS method for text classification.

Main Methods:

• Transformed Directed Acyclic Graphs (DAGs) into bipartite graphs via dual graphs.
• Utilized multi-bigraph matching to accurately capture topological order.
• Formulated NAS as a submodular function lower bound identification problem.
• Implemented a pruning strategy to reduce the search space by eliminating ineffective matching rules.

Main Results:

• BGNAS demonstrated superior performance compared to state-of-the-art NAS algorithms on public benchmarks.
• The proposed method achieved higher computational efficiency.
• The bipartite graph search space effectively captured contextual semantics, leading to enhanced generalization capabilities.

Conclusions:

• BGNAS offers a more effective and efficient approach to Neural Architecture Search for text classification.
• The bipartite graph representation accurately models topological order, improving generalization.
• This framework reduces search space complexity and computational overhead in NLP tasks.