Generating goal-oriented dialogues from documents

1 . A computer-implemented method for generating a goal-oriented dialogue, the method comprising:

receiving a corpus comprising a grounding document and historic dialogue data pair, the corpus including a grounding document having structured text organized by a set of semantic structures defining text spans of related text within the corpus, the corpus further including historic dialogue data related to the grounding document and generated by human agents and users;

generating, by a natural language understanding system, a document graph from the set of semantic structures of the grounding document;

analyzing, by the natural language understanding system, the received corpus and the generated document graph for contextual data corresponding to the set of semantic structures;

identifying, based on the analyzing, one or more semantic structures for use by a transformer-based sequence-to-sequence model to make dialogue predictions;

correlating, by a dialogue engine, the contextual data and the one or more semantic structures of the corpus of text with a plurality of utterances from the historic dialogue data;

predicting, by the transformer-based sequence-to-sequence model using at least the generated document graph, a plurality of potential utterances from an agent, and a plurality of potential utterances from a user;

generating, by a dialogue generator, a simulated dialogue based on correlated utterances of the plurality of utterances from the historic dialogue data to supplement the historic dialogue data for training a conversational model, wherein the plurality of potential utterances from the agent are correlated to next potential utterances from the user based on analysis of the received corpus;

analyzing, by the dialogue engine, the simulated dialogue in relation to the corpus to determine an accuracy level of content of the simulated dialogue based on content of the corpus; and

responsive to an accuracy threshold being met by the simulated dialogue, supplementing a training of the conversational model to respond to actual user queries, the supplemented training based on the simulated dialogue in addition to the historic dialogue data of the corpus.

2 . The computer-implemented method of claim 1 , further comprising:

generating, by the dialogue engine, a data distribution for evaluating the simulated dialogue, the data distribution being a basis for determining the accuracy level; and

modifying, by the dialogue engine and based on the accuracy level, the simulated dialogue to a second simulated dialogue, the second simulated dialogue including augmented content of the simulated dialogue;

wherein:

responsive to the accuracy threshold being met by the second simulated dialogue, training the conversational model to respond to the actual user queries, the training based on the second simulated dialogue.

3 . The computer-implemented method of claim 2 , wherein the second simulated dialogue includes at least one alternative utterance by the simulated agent and/or the simulated user.

4 . A system comprising:

a processor; and

a computer-readable storage medium communicatively coupled to the processor and storing program instructions which, when executed by the processor, cause the processor to perform a method comprising:

receiving a corpus comprising a grounding document and historic dialogue data pair, the corpus including a grounding document having structured text organized by a set of semantic structures defining text spans of related text within the corpus, the corpus further including historic dialogue data related to the grounding document and generated by human agents and users;

generating, by a natural language understanding system, a document graph from the set of semantic structures of the grounding document;

analyzing, by the natural language understanding system, the received corpus and the generated document graph for contextual data corresponding to the set of semantic structures;

identifying, based on the analyzing, one or more semantic structures for use by a transformer-based sequence-to-sequence model to make dialogue predictions;

correlating, by a dialogue engine, the contextual data and the one or more semantic structures of the corpus of text with a plurality of utterances from the historic dialogue data;

predicting, by the transformer-based sequence-to-sequence model using at least the generated document graph, a plurality of potential utterances from an agent, and a plurality of potential utterances from a user;

generating, by a dialogue generator, a simulated dialogue based on correlated utterances of the plurality of utterances from the historic dialogue data to supplement the historic dialogue data for training a conversational model, wherein the plurality of potential utterances from the agent are correlated to next potential utterances from the user based on analysis of the received corpus;

analyzing, by the dialogue engine, the simulated dialogue in relation to the corpus to determine an accuracy level of content of the simulated dialogue based on content of the corpus; and

responsive to an accuracy threshold being met by the simulated dialogue, supplementing a training of the conversational model to respond to actual user queries, the supplemented training based on the simulated dialogue in addition to the historic dialogue data of the corpus.

5 . The system of claim 4 , wherein the method performed by the processor further comprises:

generating, by the dialogue engine, a data distribution for evaluating the simulated dialogue, the data distribution being a basis for determining the accuracy level; and

modifying, by the dialogue engine and based on the accuracy level, the simulated dialogue to a second simulated dialogue, the second simulated dialogue including augmented content of the simulated dialogue;

wherein:

responsive to the accuracy threshold being met by the second simulated dialogue, training the conversational model to respond to the actual user queries, the training based on the second simulated dialogue.

6 . A computer program product comprising a computer-readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform a method comprising:

receiving a corpus comprising a grounding document and historic dialogue data pair, the corpus including a grounding document having structured text organized by a set of semantic structures defining text spans of related text within the corpus, the corpus further including historic dialogue data related to the grounding document and generated by human agents and users;

generating, by a natural language understanding system, a document graph from the set of semantic structures of the grounding document;

analyzing, by the natural language understanding system, the received corpus and the generated document graph for contextual data corresponding to the set of semantic structures;

identifying, based on the analyzing, one or more semantic structures for use by a transformer-based sequence-to-sequence model to make dialogue predictions;

correlating, by a dialogue engine, the contextual data and the one or more semantic structures of the corpus of text with a plurality of utterances from the historic dialogue data;

predicting, by the transformer-based sequence-to-sequence model using at least the generated document graph, a plurality of potential utterances from an agent, and a plurality of potential utterances from a user;

generating, by a dialogue generator, a simulated dialogue based on correlated utterances of the plurality of utterances from the historic dialogue data to supplement the historic dialogue data for training a conversational model, wherein the plurality of potential utterances from the agent are correlated to next potential utterances from the user based on analysis of the received corpus;

analyzing, by the dialogue engine, the simulated dialogue in relation to the corpus to determine an accuracy level of content of the simulated dialogue based on content of the corpus; and

responsive to an accuracy threshold being met by the simulated dialogue, supplementing a training of the conversational model to respond to actual user queries, the supplemented training based on the simulated dialogue in addition to the historic dialogue data of the corpus.

7 . The computer program product of claim 6 , wherein the method performed by the processor further comprises:

generating, by the dialogue engine, a data distribution for evaluating the simulated dialogue, the data distribution being a basis for determining the accuracy level; and

modifying, by the dialogue engine and based on the accuracy level, the simulated dialogue to a second simulated dialogue, the second simulated dialogue including augmented content of the simulated dialogue;

wherein:

responsive to the accuracy threshold being met by the second simulated dialogue, training the conversational model to respond to the actual user queries, the training based on the second simulated dialogue.

8 . The computer-implemented method of claim 1 , wherein generating the document graph includes:

determining relevant context of the grounding document; and

extracting semantic relationships from the set of semantic structures defining text spans of related text within the corpus.

9 . The computer-implemented method of claim 1 , wherein the accuracy threshold is based on a scoring evaluation technique comparing the content of the corpus with the content of the simulated dialogue.

10 . The system of claim 4 , wherein the second simulated dialogue includes at least one alternative utterance by the simulated agent and/or the simulated user.

11 . The system of claim 4 , wherein generating the document graph includes:

determining relevant context of the grounding document; and

extracting semantic relationships from the set of semantic structures defining text spans of related text within the corpus.

12 . The system of claim 4 , wherein the accuracy threshold is based on a scoring evaluation technique comparing the content of the corpus with the content of the simulated dialogue.

13 . The computer program product of claim 6 , wherein the second simulated dialogue includes at least one alternative utterance by the simulated agent and/or the simulated user.

14 . The computer program product of claim 6 , wherein generating the document graph includes:

determining relevant context of the grounding document; and

extracting semantic relationships from the set of semantic structures defining text spans of related text within the corpus.

15 . The computer program product of claim 6 , wherein the accuracy threshold is based on a scoring evaluation technique comparing the content of the corpus with the content of the simulated dialogue.