Text sentiment analysis method based on multi-level graph pooling

A text sentiment analysis method based on multi-level graph pooling, comprising steps of:

• preprocessing a target text, wherein the preprocessing comprises:
• removing noisy information, performing word segmentation, removing stop words, and training text representation of numerical vectorization;
• building a graph from preprocessed text, comprising:
• setting a sliding window with a fixed length to slide on text word sequences, calculating collocate point mutual information between word nodes, taking the collocate point mutual information between the word nodes as an edge weight between the word nodes, and performing edge connection between words, to thereby build a text graph for each text individually;
• constructing a multi-level graph pooling model, wherein the multi-level graph pooling model comprises:
• a gated graph neural network layer, a first graph self-attention pooling layer, a graph convolutional layer, and a second graph self-attention pooling layer;
• the gated graph neural network layer is connected to the first graph self-attention pooling layer, the first graph self-attention pooling layer is connected to the graph convolutional layer, and the graph convolutional layer is connected to the second graph self-attention pooling layer;
• wherein the gated graph neural network layer is configured to transfer node information, which specifically comprises:
• the gated graph neural network layer is set with a rest gate and an update gate, calculation formulas for the reset gate in one time of information transfer are as follows:
• r
• v
• =
• σ
• ⁡
• (
• W
• (
• r
• )
• ·
• x
• v
• +
• U
• (
• r
• )
• ⁢
• h
• )
• ⁢
• h
• =
• 1
• K
• ⁢
• ∑
• i
• K
• h
• i
• ⁢
• h
• ′
• =
• tanh
• ⁡
• (
• W
• ·
• x
• v
• +
• r
• v
• ⊙
• Uh
• )
• where, xv represents a feature vector of current central node, hi represents a feature vector of any one neighboring node of the current central node, h′ represents feature information after passing through the reset gate, K represents a total number of neighboring nodes of the current central node, W(r), u(r), W and U represent trainable weight matrices, σ represent a sigmoid activation function, and ⊙ represents a Hadamard product operation;
• calculation formulas for the update gate in one time of information transfer are as follows:
• z
• v
• =
• σ
• ⁡
• (
• W
• (
• z
• )
• ·
• x
• v
• +
• U
• (
• z
• )
• ⁢
• h
• )
• ⁢
• h
• =
• 1
• K
• ⁢
• ∑
• i
• K
• h
• i
• ⁢
• x
• v
• ′
• =
• z
• v
• ⊙
• h
• +
• (
• 1
• -
• z
• v
• )
• ⊙
• h
• ′
• )
• where, xv′ represent a feature vector obtained after the current central node passing through node information transfer and aggregation of the reset gate and the update gate, and W(z) and U(z) represent trainable weight matrices;
• and
• during each time of parameter update, each node receiving information of its neighboring nodes and sending information to its the neighboring nodes, realizing transfer of information in the text graph based on information transfer of gated recurrent unit (GRU) in text sequences, and outputting updated text graph vector representation;
• and
• wherein the first graph self-attention pooling layer is configured to perform an initial graph pooling operation and use a Readout function to extract low-level features for obtaining a low-level text graph vector representation;
• the graph convolution layer is configured to simultaneously employ a node feature vector matrix of the text graph and a normalized Laplacian matrix of the text graph to participate in calculation;
• and the second graph self-attention pooling layer is configured to perform a graph pooling operation again, perform a pruning update on a structure of the text graph through calculating attention scores of nodes in the text graph, and use a Readout function to extract high-level features for obtaining a high-level text graph vector representation;
• feature fusion, comprising:
• using a feature fusion function to fuse obtained text graph vector representations of different levels to thereby obtain a multi-level resultant vector representation;
• and
• sentiment category output, comprising:
• taking the multi-level resultant vector representation as input, obtaining a sentiment category probability distribution through a linear fully concatenated layer and a SoftMax classification layer, and selecting a sentiment category corresponding to a maximum probability value as a resultant sentiment category output of text.