Semantics-guided domain-specific data augmentation for text-to-graph parsing

1 . A system, comprising:

a memory configured to store computer-executable components; and

a processor that executes at least one of the computer-executable components that:

accesses an annotated training dataset that has been employed to train a deep learning model, wherein the annotated training dataset comprises a set of training abstract meaning representation (AMR) graphs respectively corresponding to a set of training natural language sentences;

generates, using the annotated training dataset and a domain lexicon legend for a domain, an augmented annotated training dataset, based on applying at least one of semantics-guided composition operations or semantics-guided substitution operations to the set of training AMR graphs to generate additional training AMR graphs and respective additional training natural language sentences associated with the additional training AMR graphs, wherein the domain lexicon legend comprises:

words associated with the domain, and

domain-specific graphs that represent the words; and

retrains, using the augmented annotated training dataset, the deep learning model to generate AMR graphs for the domain based on natural language sentences.

2 . The system of claim 1 , wherein the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

in response to a determination that a root node of the domain-specific graph is of a same semantic type as a leaf node of the first training AMR graph, generating a second training AMR graph by composing the domain-specific graph onto the leaf node; and

in response to a determination that the root node of the domain-specific graph is of the same semantic type as a non-leaf node of the first training AMR graph, generating a third training AMR graph by substituting the non-leaf node and its descendant nodes with the domain-specific graph.

3 . The system of claim 1 , wherein the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

in response to a determination that a root node of the domain-specific graph has a same argument structure as a non-leaf node of the first training AMR graph, generating a second training AMR graph by substituting the non-leaf node and its descendant nodes with the domain-specific graph.

4 . The system of claim 1 , wherein the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

in response to a determination that an incoming semantic relation associated with a root node of the domain-specific graph is the same as that between a parent node and a leaf child node of the first training AMR graph, generating a second training AMR graph by composing the domain-specific graph onto the leaf child node; and

in response to a determination that the incoming semantic relation associated with the root node of the domain-specific graph is the same as that between another parent node and a non-leaf child node of the first training AMR graph, generating a third training AMR graph by substituting the non-leaf child node and its descendant nodes with the domain-specific graph.

5 . The system of claim 1 , wherein generating the augmented annotated training dataset comprises applying, commensurately with the semantics-guided composition operations or the semantics-guided substitution operations, node-to-token alignments to the set of training natural language sentences.

6 . The system of claim 1 , wherein the domain-specific graphs respectively do not represent full sentences.

7 . The system of claim 1 , wherein the domain is a first domain, and wherein the annotated training dataset is generated by applying automated term-replacements to another annotated training dataset pertaining to a second domain that is related to the first domain.

8 . A computer-implemented method, comprising:

accessing, by a system operatively coupled to a processor, an annotated training dataset that has been employed to train a deep learning model, wherein the annotated training dataset comprises a set of training abstract meaning representation (AMR) graphs respectively corresponding to a set of training natural language sentences;

generating, by the system, using the annotated training dataset and a domain lexicon legend for a domain, an augmented annotated training dataset, based on applying at least one of semantics-guided composition operations or semantics-guided substitution operations to the set of training AMR graphs to generate additional training AMR graphs and respective additional training natural language sentences associated with the additional training AMR graphs, wherein the domain lexicon legend comprises:

words associated with the domain, and

domain-specific graphs that represent the words; and

retraining, by the system, using the augmented annotated training dataset, the deep learning model to generate AMR graphs for the domain based on natural language sentences.

9 . The computer-implemented method of claim 8 , wherein the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

generating, in response to a determination that a root node of the domain-specific graph is of a same semantic type as a leaf node of the first training AMR graph, a second training AMR graph by composing the domain-specific graph onto the leaf node; and

generating, in response to a determination that the root node of the domain-specific graph is of the same semantic type as a non-leaf node of the first training AMR graph, a third training AMR graph by substituting the non-leaf node and its descendant nodes with the domain-specific graph.

10 . The computer-implemented method of claim 8 , wherein the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

generating, in response to a determination that a root node of the domain-specific graph has a same argument structure as a non-leaf node of the first training AMR graph, a second training AMR graph by substituting the non-leaf node and its descendant nodes with the domain-specific graph.

11 . The computer-implemented method of claim 8 , wherein the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

generating, in response to a determination that an incoming semantic relation associated with a root node of the domain-specific graph is the same as that between a parent node and a leaf child node of the first training AMR graph, a second training AMR graph by composing the domain-specific graph onto the leaf child node; and

generating, in response to a determination that the incoming semantic relation associated with the root node of the domain-specific graph is the same as that between another parent node and a non-leaf child node of the first training AMR graph, a third training AMR graph by substituting the non-leaf child node and its descendant nodes with the domain-specific graph.

12 . The computer-implemented method of claim 8 , wherein the generating the augmented annotated training dataset comprises applying, commensurately with the semantics-guided composition operations or the semantics-guided substitution operations, node-to-token alignments to the set of training natural language sentences.

13 . The computer-implemented method of claim 8 , wherein the domain-specific graphs respectively do not represent full sentences.

14 . The computer-implemented method of claim 8 , wherein the domain is a first domain, and wherein the annotated training dataset is generated by applying automated term-replacements to another annotated training dataset pertaining to a second domain that is related to the first domain.

15 . A computer program product for facilitating semantics-guided domain-specific data augmentation for text-to-graph parsing, the computer program product comprising a computer-readable memory having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:

access an annotated training dataset that has been employed to train a deep learning model, wherein the annotated training dataset comprises a set of training abstract meaning representation (AMR) graphs respectively corresponding to a set of training natural language sentences;

generate, using the annotated training dataset and a domain lexicon legend for a domain, an augmented annotated training dataset, based on applying at least one of semantics-guided composition operations or semantics-guided substitution operations to the set of training AMR graphs to generate additional training AMR graphs and respective additional training natural language sentences associated with the additional training AMR graphs, wherein the domain lexicon legend comprises:

words associated with the domain, and

domain-specific graphs that represent the words; and

retrain, using the augmented annotated training dataset, the deep learning model to generate AMR graphs for the domain based on natural language sentences.

16 . The computer program product of claim 15 , the generating the augmented annotated training dataset comprises, for a first training AMR from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

generating, in response to a determination that a root node of the domain-specific graph is of a same semantic type as a leaf node of the first training AMR graph, a second training AMR graph by composing the domain-specific graph onto the leaf node; and

generating, in response to a determination that the root node of the domain-specific graph is of the same semantic type as a non-leaf node of the first training AMR graph, a third training AMR graph by substituting the non-leaf node and its descendant nodes with the domain-specific graph.

17 . The computer program product of claim 15 , the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

generating, in response to a determination that a root node of the domain-specific graph has a same argument structure as a non-leaf node of the first training AMR graph, a second training AMR graph by substituting the non-leaf node and its descendant nodes with the domain-specific graph.

18 . The computer program product of claim 15 , the generating the augmented annotated training dataset comprises, for a first training AMR graph from the annotated training dataset and for a domain-specific graph of the domain-specific graphs:

generating, in response to a determination that an incoming semantic relation associated with a root node of the domain-specific graph is the same as that between a parent node and a leaf child node of the first training AMR graph, a second training AMR graph by composing the domain-specific graph onto the leaf child node; and

generating, in response to a determination that the incoming semantic relation associated with the root node of the domain-specific graph is the same as that between another parent node and a non-leaf child node of the first training AMR graph, a third training AMR graph by substituting the non-leaf child node and its descendant nodes with the domain-specific graph.

19 . The computer program product of claim 15 , wherein generating the augmented annotated training dataset, further based on comprises applying, commensurately with the semantics-guided composition operations or the semantics-guided substitution operations, node-to-token alignments to the set of training natural language sentences.

20 . The computer program product of claim 15 , wherein the domain-specific graphs respectively do not represent full sentences.