In some embodiments, constructing the at least one edge comprises: constructing a first directed edge from the first node to the second node in the sentence graph; and constructing a second directed edge from the second node to the first node in the sentence graph, the first and second directed edges both indicating the sequential relationship.

In some embodiments, the at least one syntactic relationship comprises at least one of a dependency relationship from a third word to a fourth word of the plurality of words and an opposite dependency relationship from the fourth word to the third word, the opposite dependency relationship being opposite from the dependency relationship. In some embodiments, generating the sentence graph comprises constructing at least one of the following edges: constructing a third directed edge from a third node representing the third word to a fourth node representing the fourth word to indicate the dependency relationship; and constructing a fourth directed edge from the fourth node to the third node to indicate the opposite dependency relationship.

In some embodiments, generating the sentence graph further comprises: constructing, by one or more processors, a further edge in the sentence graph to connect one of the nodes itself, the further edge indicating a self-relationship.

In some embodiments, applying the graph convolution operation comprises: applying a graph convolution operation on the set of neighbor nodes based on weights specific for the set of neighbor nodes, the weights indicating contributions of the set of neighbor nodes to the word representation.