Aspect-level sentiment analysis method and system based on dual knowledge perception graph convolution network

Abstract

The application discloses an aspect-level sentiment analysis method and system based on a double-knowledge-aware graph convolutional network. A multi-dimensional attention mechanism is constructed to synchronously calculate sentence self-attention weights and attention distribution for specific aspects, thereby strengthening the mutual dependence between words in the text and the relevance between the sentence and the aspect word. In addition, a dependency tree and a split word dependency tree are constructed, and the matching attention score improves the effect by using syntactic information while ensuring that the aspect word focuses on the most relevant word. A double-graph structure based on different context strategies is also constructed, the graph structure is enhanced by calculating a proportional position weight for each element in the adjacency matrix of a specific aspect, the nodes in the aspect theme graph represent the words in the sentence and cover the aspect words and context words. The specific aspect is also identified and an aspect inference adjacency matrix is established to deduce the context sentiment dependence relationship of the identified aspect, thereby effectively improving the accuracy of aspect-level sentiment analysis.

Images