Structural information preserving for graph-to-text generation
A method, computer program, and computer system for training a graph-to-text generation network is provided. Encoded graph information corresponding to a target sentence is received, and the encoded graph information is decoded based on a biaffine attention score. One or more loss values are determined based on the decoded information, whereby the text-to-graph generation network is trained by minimizing the one or more loss values. A first loss value is generated by reconstructing one or more triple relations based on the biaffine attention score, and a second loss value predicts the graph as a linearized sequence.
1. A method of training a graph-to-text generation network, comprising:
receiving graph information corresponding to a target sentence, wherein the graph information includes one or more grounded triples corresponding to the target sentence;
decoding the graph information based on a biaffine attention score;
determining a first loss based on alignments between one or more nodes and target words;
determining a second loss, wherein a loss of graph structural information is associated with the second loss, wherein the loss of graph structural information associated with the second loss is minimized using a depth-first traversal; and
training a graph-to-text model based on the first loss and the second loss.
2. The method of claim 1 , wherein the first loss is generated by reconstructing the one or more grounded triples based on the biaffine attention score.
3. The method of claim 2 , wherein each of the one or more grounded triples contains a pair of nodes and a labeled relation associated with the pair of nodes.
4. The method of claim 3 , wherein each of the one or more grounded triples are mapped to the target sentence based on one or more pre-generated alignments between the one or more nodes and the target words.
5. The method of claim 1 , further comprising:
predicting the graph information as a linearized sequence using the trained graph-to-text model, wherein the second loss is used for predicting the graph information as the linearized sequence, the linearized sequence comprising a sequence of one or more tokens, wherein each token includes one of more of: a graph node, an edge label, and an inserted bracket.
6. The method of claim 1 , wherein the first loss and the second loss are based on a different aspect of the graph information.
7. A computer system for training a graph-to-text generation network, the computer system comprising:
one or more computer-readable non-transitory storage media configured to store program code; and
one or more computer processors configured to access the program code and operate as instructed by the program code, the program code comprising:
receiving code configured to cause the one or more computer processors to receive graph information corresponding to a target sentence, wherein the graph information includes one or more grounded triples corresponding to the target sentence;
decoding code configured to cause the one or more computer processors to decode the graph information based on a biaffine attention score;
first determining code configured to cause the one or more computer processors to determine a first loss based on alignments between one or more nodes and target words;
second determining code configured to cause the one or more computer processors to determine a second loss, wherein a loss of graph structural information is associated with the second loss, wherein the loss of graph structural information associated with the second loss is minimized using a depth-first traversal; and
training code configured to cause the one or more computer processors to train a graph-to-text model based on the first loss and the second loss.
8. The computer system of claim 7 , wherein the first loss is generated by reconstructing the one or more grounded triples based on the biaffine attention score.
9. The computer system of claim 8 , wherein each of the one or more grounded triples includes a pair of nodes and a labeled relation associated with the pair of nodes.
10. The computer system of claim 9 , wherein each of the one or more grounded triples are mapped to the target sentence based on one or more pre-generated alignments between the one or more nodes and the target words.
11. The computer system of claim 7 , the program code further comprising:
predicting code configured to cause the one or more computer processors to predict the graph information as a linearized sequence using the trained graph-to-text model, wherein the second loss is used for predicting the graph information as the linearized sequence, the linearized sequence comprising a sequence of one or more tokens, wherein each token includes one of more of: a graph node, an edge label, and an inserted bracket.
12. The computer system of claim 7 , wherein the first loss and the second loss are based on a different aspect of the graph information.
13. A non-transitory computer readable medium having stored thereon a computer program for training a graph-to-text generation network, the computer program configured to cause one or more computer processors to:
receive graph information corresponding to a target sentence, wherein the graph information includes one or more grounded triples corresponding to the target sentence;
decode the graph information based on a biaffine attention score;
determine a first loss based on alignments between one or more nodes and target words;
determine a second loss, wherein a loss of graph structural information is associated with the second loss, wherein the loss of graph structural information associated with the second loss is minimized using a depth-first traversal; and
train a graph-to-text model based on the first loss and the second loss.
14. The non-transitory computer readable medium of claim 13 , wherein the first loss is generated by reconstructing the one or more grounded triples based on the biaffine attention score.
15. The non-transitory computer readable medium of claim 14 , wherein each of the one or more grounded triples includes a pair of nodes and a labeled relation associated with the pair of nodes.
16. The non-transitory computer readable medium of claim 15 , wherein each of the one or more grounded triples are mapped to the target sentence based on one or more pre-generated alignments between the one or more nodes and the target words.
17. The non-transitory computer readable medium of claim 13 , wherein the computer program is further configured to cause the one or more computer processors to:
predict the graph information as a linearized sequence using the trained graph-to-text model, wherein the second loss is used for predicting the graph information as the linearized sequence, the linearized sequence comprising a sequence of one or more tokens, wherein each token includes one of more of: a graph node, an edge label, and an inserted bracket.