Natural language translation techniques

• US 9,436,681 B1
• Filed: 07/16/2013
• Issued: 09/06/2016
• Est. Priority Date: 07/16/2013
• Status: Active Grant

• Current Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Original Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Inventors: Stacey, Robert Peter, Ashton, Thomas, Wood, Adam John Phillip, Tunstall-Pedoe, William
• Primary Examiner(s): Opsasnick, Michael N
• Assistant Examiner(s): Lee, Kee Young
• Application Number:
  US13/943,637
• Time in Patent Office: 1,148 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06F 40/205 Parsing
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment
• CPC Class Codes: G06F 40/51 Translation evaluation
• CPC Class Codes: G06F 40/58 Use of machine translation, e.g. for multi-lingual retrieval, for server-side translation for client devices or for real-time translation

Abstract

Techniques are described for translating natural language input to a machine-readable form that accurately represents the semantic meaning of the input intended by the user.

Claims

1. A computer program product for translating natural language input to a machine-readable format for processing by a knowledge base, the computer program product comprising one or more non-transitory computer-readable media having computer program instructions stored therein, the computer program instructions being configured such that, when executed, the computer program instructions cause one or more computing devices to:

receive a text string derived from natural language input;

generate an input graph for the text string, the input graph representing a plurality of different subsets of words in the text string;

retrieve a template graph representing a plurality of translation templates associated with the knowledge base, each translation template being for translating matching natural language input to the machine-readable format, and each translation template corresponding to at least one of a plurality of terminal nodes in the template graph, wherein each terminal node of the template graph has an associated template priority that is based on the corresponding translation template, and wherein each internal node of the template graph is labeled with an identifier of a highest priority template that can be reached from that point in the template graph;

generate an intersection graph from the input graph and the template graph, each node in the intersection graph representing a tuple of corresponding nodes in the input and template graphs;

process the intersection graph to identify one or more candidate terminal nodes of the intersection graph, the template priorities associated with the terminal nodes of the template graph causing processing of the intersection graph to proceed preferentially along edges of the intersection graph leading to terminal nodes of the intersection graph that represent terminal nodes of the template graph that correspond to the highest priority templates;

generate a score for each terminal node of the template graph that corresponds to one of the candidate terminal nodes of the intersection graph, each score based at least in part on a probability that the text string corresponds to an intended meaning of the natural language input, and a translation template corresponding to the terminal node of the template graph;

select a first terminal node of the template graph based on the one or more scores, a first translation template corresponding to the first terminal node of the template graph;

translate the text string using the first translation template thereby generating a machine-readable translation that accurately represents a semantic meaning of the natural language input;

process the machine-readable translation using structured data to generate a result;

generate a natural language response using the result; and

present the natural language response to a user on a client device.

2. The computer program product of claim 1, wherein the computer program instructions are further configured such that, when executed, the computer program instructions cause the one or more computing devices to:

receive a second text string derived from second natural language input;

generate a second input graph for the second text string, the second input graph representing a plurality of different subsets of words in the second text string;

compare the second input graph to the template graph;

identify one or more second terminal nodes in the template graph for the second input graph;

generate a second score for each second terminal node, each second score based at least in part on an estimate of one or more probabilities associated with the second text string;

select a second terminal node with reference to the one or more second scores, wherein none of the plurality of translation templates correspond to the selected second terminal node; and

generate a notification that the second natural language input was not understood.

3. A computer-implemented method for translating a text string derived from natural language input to a machine-readable format, comprising:

using one or more computing devices, generating an input graph for the text string, the input graph representing a plurality of different subsets of words in the text string;

using the one or more computing devices, retrieving a template graph representing a plurality of translation templates, each translation template being for translating matching natural language input to the machine-readable format, and each translation template corresponding to at least one of a plurality of terminal nodes in the template graph, wherein each internal node of the template graph is labeled with an identifier of a highest priority template that can be reached from that point in the template graph;

using the one or more computing devices, generating an intersection graph from the input graph and the template graph, each node in the intersection graph representing a tuple of corresponding nodes in the input and template graphs;

using the one or more computing devices, processing the intersection graph according to template priorities associated with the nodes of the template graph thereby identifying one or more candidate terminal nodes of the intersection graph, the template priorities associated with the terminal nodes of the template graph causing processing of the intersection graph to proceed preferentially along edges of the intersection graph leading to terminal nodes of the intersection graph that represent terminal nodes of the template graph that correspond to the highest priority templates;

using the one or more computing devices, selecting a terminal node in the template graph that corresponds to one of the one or more candidate terminal nodes of the intersection graph based at least in part on a probability associated with the selected terminal node of the template graph that the text string corresponds to an intended meaning of the natural language input, wherein a first translation template corresponds to the selected terminal node of the template graph;

using the one or more computing devices, translating the text string using the first translation template thereby generating a machine-readable translation that accurately represents a semantic meaning of the natural language input;

using the one or more computing devices, processing the machine-readable translation using structured data to generate a result;

using the one or more computing devices, generating a natural language response using the result; and

using the one or more computing devices, presenting the natural language response to a user on a client device.

4. The method of claim 3, wherein the selected terminal node of the template graph is included in one or more terminal nodes of the template graph that correspond to the one or more candidate terminal nodes of the intersection graph, the method further comprising generating a score for each of the one or more terminal nodes of the template graph, each score representing an estimate of one or more probabilities associated with the text string and a translation template corresponding to the terminal node of the template graph.

5. The method of claim 4, wherein the estimate of the one or more probabilities is generated using a statistical estimate for smoothing of data representing previously received and translated natural language input.

6. The method of claim 4, wherein the score for each terminal node of the template graph is based at least in part on an estimate that the user intended input matching the translation template corresponding to the terminal node of the template graph.

7. The method of claim 3, wherein the input graph represents alternatives for at least some words of the text string, and wherein the template graph represents alternatives for at least some patterns represented by the translation templates.

8. The method of claim 3, wherein the input graph includes probabilities associated with the text string, and the template graph includes probabilities associated with the translation templates.

9. The method of claim 8, wherein processing the intersection graph includes accumulating an intermediate score for the text string with reference to the probabilities represented by the input graph and the template graph.

10. The method of claim 3, wherein each translation template has a corresponding priority score that specifies a priority ordering relative to others of the translation templates and is determined by one or more characteristics of that translation template.

11. The method of claim 3, wherein the text string is one of multiple text strings derived from the natural language input, the method further comprising receiving the text strings from speech recognition logic, each text string representing a particular interpretation of the natural language input, and generating the input graph to include the text strings.

12. A computing system for translating a text string derived from natural language input to a machine-readable format, comprising one or more computing devices configured to:

generate an input graph for the text string, the input graph representing a plurality of different subsets of words in the text string;

retrieve a template graph representing a plurality of translation templates, each translation template being for translating matching natural language input to the machine-readable format, and each translation template corresponding to at least one of a plurality of terminal nodes in the template graph, wherein each internal node of the template graph is labeled with an identifier of a highest priority template that can be reached from that point in the template graph;

generate an intersection graph from the input graph and the template graph, each node in the intersection graph representing a tuple of corresponding nodes in the input and template graphs;

process the intersection graph according to template priorities associated with the nodes of the template graph thereby identifying one or more candidate terminal nodes of the intersection graph, the template priorities associated with the terminal nodes of the template graph causing processing of the intersection graph to proceed preferentially along edges of the intersection graph leading to terminal nodes of the intersection graph that represent terminal nodes of the template graph that correspond to the highest priority templates;

select a terminal node in the template graph that corresponds to one of the one or more candidate terminal nodes of the intersection graph based at least in part on a probability associated with the selected terminal node of the template graph that the text string corresponds to an intended meaning of the natural language input, wherein a first translation template corresponds to the selected terminal node of the template graph;

translate the text string using the first translation template thereby generating a machine-readable translation that accurately represents a semantic meaning of the natural language input;

process the machine-readable translation using structured data to generate a result;

generate a natural language response using the result; and

present the natural language response to a user on a client device.

13. The computing system of claim 12, wherein the selected terminal node of the template graph is included in one or more terminal nodes of the template graph that correspond to the one or more candidate terminal nodes of the intersection graph, the one or more computing devices being further configured to generate a score for each of the one or more terminal nodes of the template graph, each score representing an estimate of one or more probabilities associated with the text string and a translation template corresponding to the terminal node of the template graph.

14. The computing system of claim 13, wherein the estimate of the one or more probabilities is generated using a statistical estimate for smoothing of data representing previously received and translated natural language input.

15. The computing system of claim 13, wherein the score for each terminal node of the template graph is based at least in part on an estimate that the user intended input matching the translation template corresponding to the terminal node of the template graph.

16. The computing system of claim 12, wherein the input graph represents alternatives for at least some words of the text string, and wherein the template graph represents alternatives for at least some patterns represented by the translation templates.

17. The computing system of claim 12, wherein the input graph includes probabilities associated with the text string, and the template graph includes probabilities associated with the translation templates.

18. The computing system of claim 17, wherein the one or more computing devices are configured to process the intersection graph in a manner that includes accumulating an intermediate score for the text string with reference to the probabilities represented by the input graph and the template graph.

19. The computing system of claim 12, wherein each translation template has a corresponding priority score that specifies a priority ordering relative to others of the translation templates and is determined by one or more characteristics of that translation template.

20. The computing system of claim 12, wherein the text string is one of multiple text strings derived from the natural language input, and wherein the one or more computing devices are further configured to receive the text strings from speech recognition logic, each text string representing a particular interpretation of the natural language input, and generating the input graph to include the text strings.