Methods and apparatuses for training service model and determining text classification category

A method for determining a text classification category, implemented by using a service model including an encoding network, a fusion layer, and a fully connected layer, the method comprising:

• obtaining a second text for classification category determination;
• processing the second text by using the encoding network, to obtain a second semantic vector corresponding to the second text;
• obtaining category expression vectors respectively corresponding to a plurality of candidate categories and obtained by processing a category hierarchical relationship network by using a graph convolutional network, wherein the category hierarchical relationship network includes nodes having a one-to-one correspondence with individual candidate categories of the plurality of candidate categories, at least two of the nodes that correspond to candidate categories having a direct superior-inferior relationship are connected by a connection edge, each node corresponds to a respective node expression vector having a corresponding category feature vector as its initial value, the graph convolutional network is configured to perform node expression vector aggregation of neighboring nodes for individual nodes to update corresponding node expression vectors, and the category expression vectors are node expression vectors corresponding to the nodes obtained through updating using the last layer of the graph convolutional network;
• fusing the category expression vectors together based on degrees of association between the category expression vectors and the second semantic vector by using the fusion layer, to obtain a second fused vector for the second text;
• obtaining a second prediction result for the second text based on the second fused vector and the second semantic vector by using the fully connected layer;
• and
• determining a classification category of the second text based on the second prediction result.

A non-transitory computer readable medium storing contents that, when executed by one or more processors, cause the one or more processors to perform actions for training a service model, wherein the service model includes an encoding network, a fusion layer, and a fully connected layer and the actions comprise:

• obtaining a first text from a training sample set, the first text corresponding to a text label, and the text label being determined from a plurality of candidate categories;
• processing the first text by using the encoding network, to obtain a first semantic vector;
• processing a category hierarchical relationship network by using a graph convolutional network, to obtain category expression vectors respectively corresponding to a plurality of candidate categories, wherein the category hierarchical relationship network includes nodes having a one-to-one correspondence with individual candidate categories of the plurality of candidate categories, at least two of the nodes that correspond to candidate categories having a direct superior-inferior relationship are connected by a connection edge, each node corresponds to a respective node expression vector having a corresponding category feature vector as its initial value, the graph convolutional network is configured to perform node expression vector aggregation of neighboring nodes for individual nodes to update corresponding node expression vectors, and the category expression vectors are node expression vectors corresponding to the nodes obtained through updating using the last layer of the graph convolutional network;
• fusing the category expression vectors together based on degrees of association between the category expression vectors and the first semantic vector by using the fusion layer, to obtain a first fused vector for the first text;
• obtaining a first prediction result for a classification category of the first text based on the first fused vector and the first semantic vector by using the fully connected layer;
• and
• adjusting model parameters in the encoding network, the graph convolutional network, the fusion layer, and the fully connected layer with an objective of consistency between the first prediction result and the text label, to train the service model.

A non-transitory computer readable medium storing contents that, when executed by one or more processors, cause the one or more processors to perform actions for determining a text classification category by using a service model including an encoding network, a fusion layer, and a fully connected layer, the actions comprising:

• obtaining a second text for classification category determination;
• processing the second text by using the encoding network, to obtain a second semantic vector corresponding to the second text;
• obtaining category expression vectors respectively corresponding to a plurality of candidate categories and obtained by processing a category hierarchical relationship network by using a graph convolutional network, wherein the category hierarchical relationship network includes nodes having a one-to-one correspondence with individual candidate categories of the plurality of candidate categories, at least two of the nodes that correspond to candidate categories having a direct superior-inferior relationship are connected by a connection edge, each node corresponds to a respective node expression vector having a corresponding category feature vector as its initial value, the graph convolutional network is configured to perform node expression vector aggregation of neighboring nodes for individual nodes to update corresponding node expression vectors, and the category expression vectors are node expression vectors corresponding to the nodes obtained through updating using the last layer of the graph convolutional network;
• fusing the category expression vectors together based on degrees of association between the category expression vectors and the second semantic vector by using the fusion layer, to obtain a second fused vector for the second text;
• obtaining a second prediction result for the second text based on the second fused vector and the second semantic vector by using the fully connected layer;
• and
• determining a classification category of the second text based on the second prediction result.