Structured graph-to-text generation with two step fine-tuning

Patent metadata

Field	Value
Publication number	US-11727210-B2
Application number	US-202117162040-A
Country	US
Kind code	B2
Filing date	Jan 29, 2021
Priority date	Aug 14, 2020
Publication date	Aug 15, 2023
Grant date	Aug 15, 2023

Abstract

Embodiments described herein provide systems and methods for data-to-text generation. The embodiments receive input data that includes a resource description framework (RDF) triples in an RDF graph. A data-to-text generation system generates position aware embeddings, including position embeddings, triple role embeddings, and tree-level embeddings. Using the position aware embeddings and the RDF graph, the data-to-text generation system generates a textual description for the RDF graph.

First claim

What is claimed is:

1. A system comprising:
a non-transitory memory; and
one or more hardware processors coupled to the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform operations comprising:
receiving, at a data-to-text generation system that includes a generative language model, an input data that includes resource description framework (RDF) triples in an RDF graph;
generating, using the data-to-text generation system, embeddings from the RDF graph based on tokens of the input data, wherein the embeddings include a position aware embedding that identifies a position of an RDF triple of the RDF triples in the RDF graph; and
generating, using the data-to-text generation system, a textual description of the input data based on the embeddings and the RDF graph,
wherein the position aware embedding includes a position embedding that identifies a position of a token indicating whether a word in the RDF triple of the RDF triples is a subject, a relation, or an object.

2. The system of claim 1 , wherein the RDF triple includes words that correspond to a subject, a relation, or an object.

3. The system of claim 1 , wherein the position aware embedding includes a position embedding that identifies a position of a token that stores a word in the RDF triple from the RDF triples.

4. The system of claim 1 , wherein the position aware embedding includes a triple role embedding that identifies that a token includes a word or an indication of a role of the word in the RDF triple from the RDF triples that corresponds to a subject, an object, or a relation.

5. The system of claim 1 , wherein the position aware embedding includes a tree-level embedding that identifies a tree distance from a root of a parsing tree to a level in the parsing tree that includes a token, wherein the token stores a word or an indication of a role of the word in the RDF triple from the RDF triples.

6. The system of claim 1 , wherein the generating the embeddings further comprises generating a token embedding that identifies a token that stores a word or an indication of a role of the word in the RDF triple from the RDF triples.

7. A method comprising:
receiving, at a data-to-text generation system that includes a generative language model, the data-to-text generation system configured to execute on a processor, an input data that includes resource description framework (RDF) triples in an RDF graph;
generating, using the data-to-text generation system, embeddings from the RDF graph based on tokens of the input data, wherein the embeddings include a position aware embedding that identifies a position of an RDF triple of the RDF triples in the RDF graph; and
generating, using the data-to-text generation system, a textual description of the input data based on the position aware embedding and the RDF graph,
wherein the position aware embedding includes a position embedding that identifies a position of a token indicating whether a word in the RDF triple of the RDF triples is a subject, a relation, or an object.

8. The method of claim 7 , further comprising:
training the generative language model to generate the position aware embeddings.

9. The method of claim 7 , wherein the position aware embedding includes a position embedding that identifies a position of a token that stores a word in the RDF triple from the RDF triples.

10. The method of claim 7 , wherein the position aware embedding includes a triple role embedding that identifies that a token includes a word or an indication of a role of the word in the RDF triple from the RDF triples that corresponds to a subject, an object, or a relation.

11. The method of claim 7 , wherein the position aware embedding includes a tree-level embedding that identifies a tree distance from a root of a parsing tree to a level in the parsing tree that includes a token, wherein the token stores a word or an indication of a role of the word in the RDF triple from the RDF triples.

12. The method of claim 7 , wherein the generating the embeddings further comprises generating a token embedding that identifies a token that stores a word or an indication of a role of the word in the RDF triple from the RDF triples.

13. A non-transitory machine-readable medium having stored thereon machine-readable instructions executable to cause a machine to perform operations comprising:
receiving, at a data-to-text generator that includes a generative language model, an input data that includes structured data triples in a structured graph;
generating, using the data-to-graph generator, embeddings from the structured graph based on tokens of the input data, wherein the embeddings include position aware embeddings that identify position of a triple of triples in the structured graph; and
generating, using a data-to-text module, a textual description of the input data based on the position aware embeddings and the structured graph,
wherein the position aware embeddings include a position embedding that identifies a position of a token indicating whether a word in the triple of the triples is a subject, a relation, or an object.

14. The non-transitory machine-readable medium of claim 13 , wherein the position aware embeddings include a position embedding that identifies a position of a token that stores a word in the triple from the triples.

15. The non-transitory machine-readable medium of claim 13 , wherein the position aware embeddings include a triple role embedding that identifies that a token includes a word or an indication of a role of the word in the triple from the triples that corresponds to a subject, an object, or a relation.

16. The non-transitory machine-readable medium of claim 13 , wherein the position aware embeddings include a tree-level embedding that identifies a tree distance from a root of a parsing tree to a level in the parsing tree that stores a token, wherein the token includes a word or an indication of a role of the word in the triple from the triples.

17. The non-transitory machine-readable medium of claim 13 , wherein the generating the embeddings further comprises generating token embeddings that identify a token that stores a word or an indication of a role of the word in the triple from the triples.

Assignees

• Salesforce Com Inc

Inventors

• Wang Qingyun
• Rajani Nazneen
• Yavuz Semih
• Lin Xi

Classifications

• G06F40/284 Primary

  Lexical analysis, e.g. tokenisation or collocates · CPC title

• G06F40/10

  Text processing (natural language analysis G06F40/20; semantic analysis G06F40/30; processing or translation of natural language G06F40/40) · CPC title

• G06F40/205

  Parsing · CPC title

• G06F40/56 Primary

  Natural language generation · CPC title

• G06F40/14

  Tree-structured documents (parsing G06F40/205; validation G06F40/226) · CPC title