Natural language system question classifier, semantic representations, and logical form templates

• US 10,262,062 B2
• Filed: 12/21/2015
• Issued: 04/16/2019
• Est. Priority Date: 12/21/2015
• Status: Active Grant

• Current Assignee: Adobe Inc.
• Original Assignee: Adobe Inc.
• Inventors: Chang, Walter Wei-Tuh, Girlea, Codruta L.
• Primary Examiner(s): Zhu, Richard Z
• Application Number:
  US14/977,334
• Publication Number:
  US 20170177715A1
• Time in Patent Office: 1,212 Days
• Field of Search: None
• CPC Class Codes: G06F 16/3329 Natural language query formulation or dialogue systems
• CPC Class Codes: G06F 16/3344 using natural language analysis
• CPC Class Codes: G06F 16/35 Clustering; Classification
• CPC Class Codes: G06F 40/186 Templates
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/04 Inference or reasoning models
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering
• CPC Class Codes: G10L 2015/0635 updating or merging of old and new templates; Mean values; Weighting

Abstract

Natural language system question classifier, semantic representations, and logical form template techniques and systems are described. In one or more implementations, a natural language input is classified as corresponding to respective ones of a plurality of classes of questions. A semantic intent of the natural language input is extracted as a semantic entity and a semantic representation. Question classification labels that classify the question included in the natural language input is then used to select at least one of a plurality of logical form templates. The semantic intent that is extracted from the natural language input is then used to fill in the selected logical form templates, such as to fill in entity, subject, predicate, and object slots using the semantic entity and semantic representation. The filled-in logical form template is then mapped to form a database query that is then executed to query a database to answer the question.

Claims

1. In a digital medium environment to answer a question included in a natural language input, a system implemented by at least one computing device, the system comprising:

a question classifier module implemented at least partially in hardware to classify the natural language input as corresponding to respective ones of a plurality of classes of questions using at least one question classification label;

a semantic extractor module implemented at least partially in hardware to extract a semantic intent of the natural language input as a semantic entity and a semantic representation of the natural language input;

a logical form selector module implemented at least partially in hardware to;

select a plurality of logical form templates, the selection based at least in part on the at least one question classification label;

rank the plurality of logical form templates based on an amount of completion of logical form slot-filling and alignment of named entity types in the natural language input to expected entity types in the logical form templates;

select one of the logical form templates based at least in part on the rank; and

at least partially fill in the selected one of the logical form templates using the semantic entity and the semantic representation;

a query mapper module implemented at least partially in hardware to produce a database query by mapping the at least partially filled in logical form to the database query, the at least partially filled in logical form including;

a pre-condition slot containing a first attribute that controls grouping of results of the database query;

a post-condition slot containing a second attribute that constrains an order of the results of the database query; and

a slot populated with an indicator of a statistical function selected based on the natural language input from a plurality of available statistical functions; and

a database query module implemented at least partially in hardware that generates an answer to the question included in the natural language input by executing the database query against a database structured in accordance with a structure of the database query, the answer providing user assistance by responding to the question.

2. The system as described in claim 1, wherein the plurality of classes of questions include who, what, when, where, which, why, or how many.

3. The system as described in claim 1, wherein the question classifier module is configured to perform the classification using a multi-level question type taxonomy that is trained from example questions across a plurality of domains.

4. The system as described in claim 1, wherein the semantic extractor module is configured to extract the semantic intent automatically and without user intervention through use of one or more models that are configured to address language ambiguities contained in the natural language input.

5. The system as described in claim 1, wherein the semantic entities include person, place, time entities, time units, time period, role entities, or thing entities.

6. The system as described in claim 1, wherein the semantic representation is a subject-predicate-object tuple that functions as a statement linking a subject of the subject-predicate-object tuple to an object of the subject-predicate-object tuple via a predicate of the subject-predicate-object tuple.

7. The system as described in claim 1, wherein the selected one of the logical form templates includes attribute/value slots that are filled in using the semantic entity or semantic representation.

8. The system as described in claim 1, wherein the query mapper module is configured to map the at least partially filled in logical form to the database query in which a structure of the at least partially filled in logical form is used to form the database query as a structured database query.

9. The system as described in claim 1, further comprising a clarification generator module implemented at least partially in hardware to clarify at least one ambiguity contained in the natural language input by soliciting information from a user or infer appropriate values to use from a user'"'"'s prior history or history of similar users automatically and without user intervention.

10. In a digital medium environment to classify a question included in a natural language input as part of answering a question included in the natural language input, a method implemented by a computing device, the method comprising:

receiving a natural language input by the computing device;

classifying the natural language input, by the computing device, as corresponding to respective ones of a plurality of classes by;

extracting features from the natural language input including, for each word in the natural language input, building a feature representation by concatenating semantic and syntactic information; and

processing the extracted features using a classifier model that is trained across a plurality of domains to obtain a question classification label based on each feature representation;

generating a database query based at least in part on the obtained question classification label by the computing device, the generating including using the obtained question classification label to select at least one of a plurality of logical form templates that define a structure of the database query, the at least one of a plurality of logical form templates including;

a pre-condition slot containing a first attribute that controls grouping of results of the database query;

a post-condition slot containing a second attribute that constrains an order of the results of the database query; and

a slot populated with an indicator of a statistical function selected based on the natural language input from a plurality of available statistical functions; and

producing an answer to the question included in the natural language input by executing the database query against a database structured in accordance with the structure of the database query, the answer providing user assistance by responding to the question.

11. The method as described in claim 10, wherein the features are configured as part of a type taxonomy that is usable to support multilevel question classification that is definable using the question classification label.

12. The method as described in claim 10, wherein the generating further includes filling in the selected logical form template using a semantic entity and a semantic representation extracted from the natural language input.

13. The method as described in claim 12, wherein the semantic representation is a subject-predicate-object tuple that functions as a statement linking a subject of the subject-predicate-object tuple to an object of the subject-predicate-object tuple via a predicate of the subject-predicate-object tuple.

14. The method as described in claim 10, wherein the natural language input is received by the computing device via a typed input or spoken utterance.

15. The method as described in claim 10, wherein the syntactic information includes, for an adjective and an adverb, an actual adjective and an actual adverb in the natural language input, respectively, in place of part-of-speech tags, and the feature representation indicates whether said each word is constant or variable.

16. In a digital medium environment to extract a semantic intent of a question included in a natural language input as part of answering a question included in the natural language input, a method implemented by a computing device, the method comprising:

receiving a natural language input by the computing device;

generating part-of-speech tags for each word in the natural language input;

extracting the semantic intent of the natural language input automatically and without user intervention by the computing device as a semantic entity and a semantic representation of the natural language input based on the part-of-speech tags, the semantic representation configured as a subject, predicate, object tuple that functions as a statement linking the subject to the object via the predicate, the extracting including correcting a language error in the natural language input and a part-of-speech tag error resulting from the language error;

generating a database query based at least in part on the extracted semantic intent by the computing device, the generating including using the semantic entity or semantic representation to at least partially fill in a logical form template that defines a structure of the database query, the at least partially filled in logical form including;

a pre-condition slot containing a first attribute that controls grouping of results of the database query;

a post-condition slot containing a second attribute that constrains an order of the results of the database query; and

a slot populated with an indicator of a statistical function selected based on the natural language input from a plurality of available statistical functions; and

producing an answer to the question included in the natural language input by executing the database query against a database structured in accordance with the structure of the database query, the answer providing user assistance by responding to the question.

17. The system as described in claim 16, wherein the extracting of the semantic intent is performed automatically and without user intervention through use of one or more models that are configured to address language ambiguities contained in the natural language input.

18. The method as described in claim 16, wherein the generating further includes selecting the logical form template based on a classification of the question included in the natural language input.

19. The method as described in claim 16, wherein the natural language input is received by the computing device via a typed input or spoken utterance.

20. The system as described in claim 16, wherein the extracting the semantic intent includes extracting the subject, predicate, object tuple according to:

when extraction rules are available, the extraction rules; and

when the extraction rules are not available, a greedy algorithm that uses a first verb expression encountered in the natural language input as a predicate of the subject, predicate, object tuple.