A Tensorized Spectral Attention Mechanism for Efficient Natural Language Processing

The attention mechanism is at the core of state-of-the-art Natural Language Processing (NLP) models, owing to its ability to focus on the most contextually relevant part of a sequence. However, current attention models rely on "flat-view" matrix methods to process sequence of tokens embedded in vector spaces, resulting in exceedingly high parameter complexity for practical applications. To this end, we introduce a novel Tensorized Spectral Attention (TSA) mechanism, which leverages on the Graph Tensor Network (GTN) framework to efficiently process tensorized token embeddings via attention based spectral graph filters. By virtue of multi-linear algebra, such tensorized token embeddings are shown to effectively bypass the Curse of Dimensionality, reducing the parameter complexity of the attention mechanism from exponential to linear in the weight matrix dimensions. Furthermore, the graph formulation of the attention domain enables the processing of tensorized embeddings through spectral graph convolution filters, which further increases its expressive power. The benefits of the TSA are demonstrated through five benchmark NLP experiments, where the proposed mechanism is shown to achieve better or comparable results against traditional attention models, while incurring drastically lower parameter complexity.