Exemplar-based natural language processing

Patent metadata

Field	Value
Publication number	US-10169329-B2
Application number	US-201615220276-A
Country	US
Kind code	B2
Filing date	Jul 26, 2016
Priority date	May 30, 2014
Publication date	Jan 1, 2019
Grant date	Jan 1, 2019

Abstract

Systems and processes for exemplar-based natural language processing are provided. In one example process, a first text phrase can be received. It can be determined whether editing the first text phrase to match a second text phrase requires one or more of inserting, deleting, and substituting a word of the first text phrase. In response to determining that editing the first text phrase to match the second text phrase requires one or more of inserting, deleting, and substituting a word of the first text phrase, one or more of an insertion cost, a deletion cost, and a substitution cost can be determined. A semantic edit distance between the first text phrase and the second text phrase in a semantic space can be determined based on one or more of the insertion cost, the deletion cost, and the substitution cost.

First claim

What is claimed is:

1. A non-transitory computer-readable storage medium for natural language processing comprising computer-executable instructions for causing a processor to:
receive a speech input representing a user request;
generate a first text phrase corresponding to the speech input;
determine, with respect to a semantic space, a plurality of semantic edit distances between the first text phrase and a plurality of exemplar text phrases, wherein each exemplar text phrase of the plurality of exemplar text phrases is associated with a respective predetermined intent of a plurality of predetermined intents;
determine a plurality of centroid distances between a centroid position of the first text phrase in the semantic space and a plurality of centroid positions of the plurality of exemplar text phrases in the semantic space;
determine a plurality of degrees of semantic similarity between the first text phrase and the plurality of exemplar text phrases, wherein each degree of semantic similarity of the plurality of degrees of semantic similarity is determined based on a linear combination of a respective semantic edit distance of the plurality of semantic edit distances and a respective centroid distance of the plurality of centroid distances;
identify, based on the plurality of degrees of semantic similarity, a first exemplar text phrase of the plurality of exemplar text phrases, wherein the first exemplar text phrase is most semantically similar to the first text phrase among the plurality of exemplar text phrases, and wherein the first exemplar text phrase is associated with a first predetermined intent of the plurality of predetermined intents;
determine, based on the identified first exemplar text phrase, a user intent corresponding to the first text phrase, wherein the determined user intent corresponds to the first predetermined intent; and
in accordance with the determined user intent, perform one or more tasks responsive to the user request.

2. The computer-readable storage medium of claim 1 , wherein the centroid position of the first text phrase is determined based on a semantic position of one or more words of the first text phrase in the semantic space, and wherein a centroid position of the first exemplar text phrase is determined based on a semantic position of one or more words of the first exemplar text phrase in the semantic space.

3. The computer-readable storage medium of claim 1 , wherein the centroid position of the first text phrase is determined based on a salience of one or more words of the first text phrase, and wherein a centroid position of the first exemplar text phrase is determined based on a salience of one or more words of the first exemplar text phrase.

4. The computer-readable storage medium of claim 1 , wherein a first degree of semantic similarity of the plurality of degrees of semantic similarly is based on whether the first text phrase includes a first word that the first exemplar text phrase does not include and whether a predetermined list of keywords includes the first word.

5. The computer-readable storage medium of claim 1 , wherein determining a first semantic edit distance of the plurality of semantic edit distances comprises determining one or more cost values associated with editing the first text phrase to match the first exemplar text phrase.

6. The computer-readable storage medium of claim 1 , wherein the centroid position of the first text phrase comprises a weighted centroid of a plurality of points in the semantic space, and wherein each point of the plurality of points is a semantic representation of a respective word in the first text phrase.

7. A non-transitory computer-readable storage medium for natural language processing comprising computer-executable instructions for causing a processor to:
receive a speech input representing a user request;
generate a first text phrase corresponding to the speech input;
determine a plurality of sets of word-level differences between the first text phrase and a plurality of exemplar text phrases, wherein each exemplar text phrase of the plurality of exemplar text phrases is associated with a respective predetermined intent of a plurality of predetermined intents;
determine a plurality of total semantic costs representing the plurality of sets of word-level differences between the first text phrase and the plurality of exemplar text phrases;
determine a plurality of centroid distances between a centroid position of the first text phrase in a semantic space and a plurality of centroid positions of the plurality of exemplar text phrases in the semantic space;
determine a plurality of degrees of semantic similarity between the first text phrase and the plurality of exemplar text phrases, wherein each degree of semantic similarity of the plurality of degrees of semantic similarity is determined based on a linear combination of a respective total semantic cost of the plurality of total semantic costs and a respective centroid distance of the plurality of centroid distances;
identify, based on the plurality of degrees of semantic similarity, a first exemplar text phrase of the plurality of exemplar text phrases, wherein the first exemplar text phrase is most semantically similar to the first text phrase among the plurality of exemplar text phrases, and wherein the first exemplar text phrase is associated with a first predetermined intent of the plurality of predetermined intents;
determine, based on the identified first exemplar text phrase, a user intent corresponding to the first text phrase, wherein the determined user intent corresponds to the first predetermined intent; and
in accordance with the determined user intent, perform one or more tasks responsive to the user request.

8. The computer-readable storage medium of claim 7 , wherein the centroid position of the first text phrase is determined based on a semantic position of one or more words of the first text phrase in the semantic space, and wherein a centroid position of the first exemplar text phrase is determined based on a semantic position of one or more words of the first exemplar text phrase in the semantic space.

9. The computer-readable storage medium of claim 7 , wherein the centroid position of the first text phrase is determined based on a salience of one or more words of the first text phrase, and wherein a centroid position of the first exemplar text phrase is determined based on a salience of one or more words of the first exemplar text phrase.

10. The computer-readable storage medium of claim 7 , wherein a first degree of semantic similarity of the plurality of degrees of semantic similarity is based on whether the first text phrase includes a first word that the first exemplar text phrase does not include and whether a predetermined list of keywords includes the first word.

11. The computer-readable storage medium of claim 7 , wherein a first total semantic cost of the plurality of total semantic costs comprises a linear combination of a plurality of costs representing a first set of word-level differences of the plurality of sets of word-level differences, and wherein the first set of word-level differences is between the first text phrase and the first exemplar text phrase.

12. The computer-readable storage medium of claim 7 , wherein the centroid position of the first text phrase comprises a weighted centroid of a plurality of points in the semantic space, and wherein each point of the plurality of points is a semantic representation of a respective word in the first text phrase.

13. A method for processing natural language comprising:
at an electronic device having a processor and

Assignees

• Apple Inc

Inventors

• Futrell Richard L
• Gruber Thomas R

Classifications

• G06F40/194 Primary

  Calculation of difference between files · CPC title

• G06F40/30 Primary

  Semantic analysis · CPC title

• G06F17/2211

  Physics · mapped topic

• G06F17/2785 Primary

  Physics · mapped topic