Natural language vocabulary generation and usage

• US 9,588,964 B2
• Filed: 11/21/2012
• Issued: 03/07/2017
• Est. Priority Date: 09/18/2012
• Status: Active Grant

• Current Assignee: Adobe Inc.
• Original Assignee: Adobe Systems Incorporated (Adobe Inc.)
• Inventors: Chang, Walter W., Wilensky, Gregg D., Dontcheva, Lubomira A.
• Primary Examiner(s): COLUCCI, MICHAEL C
• Application Number:
  US13/683,509
• Publication Number:
  US 20140081626A1
• Time in Patent Office: 1,567 Days
• Field of Search: 704/9, 704/4, 704/277, 704/257, 704/2, 709/206, 707/797, 707/755, 707/738, 707/737, 706/59, 345/589
• US Class Current: 1/1
• CPC Class Codes: G06F 40/242 Dictionaries
• CPC Class Codes: G06F 40/247 Thesauruses; Synonyms
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30

Abstract

Natural language vocabulary generation and usage techniques are described. In one or more implementations, one or more search results are mined for a domain to determine a frequency at which words occur in the one or more search results, respectively. A set of the words is selected based on the determined frequency. A sense is assigned to each of the selected set of the words that identifies a part-of-speech for a respective word. A vocabulary is then generated that includes the selected set of the words and a respective said sense, the vocabulary configured for use in natural language processing associated with the domain.

Claims

1. A method implemented at least partially in hardware of a computing device, the method comprising:

mining, by the computing device, one or more search results returned from a search engine for a particular one of a plurality of domains to determine a frequency at which words occur in the one or more search results relating to the particular domain, respectively;

selecting, by the computing device, a set of the words from the search results of the particular domain based on the determined frequency;

assigning, by the computing device, a sense to each of the selected set of the words that identifies a part-of-speech for a respective said word, the sense based in part on the particular domain; and

generating, by the computing device, a vocabulary for the particular domain that includes the selected set of the words and a respective said sense, the vocabulary describing a term sense bias as applied to a term semantic distance, the vocabulary configured for use in natural language processing to disambiguate natural language input according to the term sense bias applied to the term semantic distance of the particular domain.

2. A method as described in claim 1, wherein the one or more search results are generated such that the particular domain is used as part of a search query that is used to generate the one or more search results.

3. A method as described in claim 1, wherein the particular domain involves image editing operations.

4. A method as described in claim 1, wherein the mining of the one or more search results to determine the frequency is performed by building an N-gram tree in which a first said word that has a greater frequency of occurrence in the one or more search results than a second said word is disposed closer to a root of the N-gram tree than the second said word.

5. A method as described in claim 1, wherein the assigning of the sense is performed by examining one or more lexicon ontologies.

6. A method as described in claim 1, wherein the assigning of the sense is performed such that each of the selected set of the words is assigned to sense that is a noun, verb, adjective, or adverb.

7. A method as described in claim 1, wherein the vocabulary is generated such that the assigned sense to each of the selected set of the words is used to bias the vocabulary for the domain.

8. A method as described in claim 1, wherein the one or more search results include webpages involving a tutorial associated with the particular domain.

9. A method as described in claim 1, wherein the one or more search results are generated using a plurality of searches in which at least one of the searches is performed using word identified from a previous said search as part of a search query.

10. A method implemented at least partially in hardware of one or more computing devices, the method comprising:

using a spreading activation algorithm, by the one or more computing devices, to generate a set of words relating to a particular one of a plurality of domains to determine which words are most frequently used in the particular domain;

biasing, by the one or more computing devices, the set of words by assigning a sense based on one or more lexicon ontologies to each of the words that identifies a part-of-speech of a respective said word for the particular domain, the biasing configured to reflect a sense corresponding to how each of the words are used in the particular domain; and

configuring, by the one or more computing devices, the set of words and the assigned sense for each of the words to form a vocabulary that is configured for use in natural language processing for the particular domain, the vocabulary describing the assigned sense bias for the particular domain as applied to a term semantic distance.

11. A method as described in claim 10, wherein the part-of-speech is a noun, verb, adjective, or adverb.

12. A method as described in claim 10, wherein the using of the spreading activation algorithm includes building an N-gram tree to represent frequency of occurrence of words in search results.

13. A method as described in claim 10, wherein the using of the spreading activation algorithm is performed such that the particular domain is used as part of a search query that is used to generate one or more search results that are used to derive the set of words.

14. A method as described in claim 13, wherein the one or more search results include web pages.

15. One or more computer-readable hardware storage media comprising instructions stored thereon that, responsive to execution by a computing device, causes the computing device to perform operations comprising:

generating, at least partially in hardware of the computing device, a vocabulary for use in natural language processing for a particular one of a plurality of domains by;

building an N-gram tree to represent frequency of occurrence of words in search results involving the particular domain, the search results obtained using a spreading activation algorithm constrained to the particular domain;

selecting a set of the words based on the represented frequency of occurrence in the N-gram tree; and

assigning a sense to each of the words in the set that identifies a part-of-speech of a respective said word, the sense reflecting, for each of the words, how each of the words are used in the particular domain of the plurality of domains.

16. One or more computer-readable hardware storage media as described in claim 15, wherein the assigning is performed using one or more lexicon ontologies.

17. One or more computer-readable hardware storage media as described in claim 15, wherein the spreading activation algorithm is constrained to the particular domain by including the particular domain in a search query used to perform a search that is used to generate the search results.

18. One or more computer-readable hardware storage media as described in claim 15, wherein the part-of-speech is a noun, verb, adjective, or adverb.

19. A method implemented at least partially in hardware of a computing device comprising:

receiving a natural language input by the computing device;

processing the natural language input by the computing device to identify an image editing operation, the processing performed using a vocabulary that includes a set of words and a sense that identifies a part-of-speech for each of the words that is biased for a domain that includes image editing operations, the natural language input not being part of the set of words of the vocabulary, the processing comprising disambiguating the natural language input according to a term sense bias applied to a term semantic distance, the term sense bias relating to the sense that identifies the part-of-speech for each of the words based on the domain that includes image editing operations; and

performing the image editing operation by the computing device.

20. A method as described in claim 19, wherein the part-of-speech is a noun, verb, adjective, or adverb.

21. A method as described in claim 19, wherein the sense is determined based on one or more lexicon ontologies.

22. A method as described in claim 19, wherein the set of words included in the vocabulary is generated using a spreading activation algorithm.

23. A method as described in claim 22, wherein the spreading activation algorithm includes building an N-gram tree to represent frequency of occurrence of words in search results.

24. A method as described in claim 19, wherein the processing includes parsing the natural language input into words and finding a sense for each word that identifies a respective said part-of-speech for the natural language input.

25. A method as described in claim 24, wherein the term semantic distance is determined by computing semantic distances between the words in the natural language input and words included in the vocabulary, respectively.

26. A method as described in claim 24, wherein the processing is based at least in part on similarity of lexical functional groups of the words in the natural language input and the plurality of words included in the vocabulary, respectively.

27. A method as described in claim 19, wherein the natural language input is text processed from audio data using a speech-to-text engine.

28. A method as described in claim 19, wherein the processing includes filtering pairs of the parsed word and particular ones of the plurality of words in which a path that is usable to compute the semantic distance includes an ontology root.

29. A method implemented at least partially in hardware of one or more computing devices, the method comprising:

identifying, by the one or more computing devices, a word parsed from a natural language input that is not included in a vocabulary associated with an image editing operation domain; and

disambiguating, by the one or more computing devices, the parsed word using the vocabulary to identify non-matching words to accomplish image editing operations, the disambiguating based on;

semantic distance between the parsed word and a plurality of words included in the vocabulary, the semantic distance based at least in part on a distance matrix computed between the parsed word and one or more of the plurality of words included in the vocabulary, the distance matrix comprising semantic distances between terms for respective parts of speech; and

similarity of lexical functional groups of the parsed word and the plurality of words.

30. A method as described in claim 29, wherein the lexical functional groups correspond to respective parts-of-speech.

31. A method as described in claim 30, wherein the parts-of-speech include nouns, verbs, adjectives, or adverbs.

32. A method as described in claim 29, wherein the similarity of the lexical functional groups is utilized in an instance in which the semantic distance between the parsed word and a first said word included in the vocabulary matches the semantic distance between the parsed word and a second said word included in the vocabulary.

33. A method as described in claim 29, wherein the word from the natural language input is not included in the vocabulary.

34. A method as described in claim 29, wherein the disambiguating includes filtering pairs of the parsed word and particular ones of the plurality of words in which a path that is usable to compute a semantic distance includes an ontology root.

35. One or more computer-readable hardware storage media comprising instructions stored thereon that, responsive to execution at least partially in hardware by a computing device, causes the computing device to perform operations comprising:

receiving a word via a natural language input, the received word not being included in an image editing operation vocabulary;

determining which of a plurality of words included in the image editing operation vocabulary is considered to be more similar to the received word than other said words in the image editing operation vocabulary, the determining based at least in part on;

semantic distance between the received word and a plurality of words included in the vocabulary, the semantic distance based at least in part on a distance matrix computed between the received word and one or more of the plurality of words included in the image editing operation vocabulary, the distance matrix comprising semantic distances between terms for respective parts of speech; and

similarity of lexical functional groups of the received word and the plurality of words; and

identifying an image editing operation based on the determining.

36. One or more computer-readable hardware storage media as described in claim 35, wherein the lexical functional groups correspond to respective parts-of-speech.

37. One or more computer-readable hardware storage media as described in claim 36, wherein the parts-of-speech include nouns, verbs, adjectives, or adverbs.

38. One or more computer-readable hardware storage media as described in claim 35, wherein the determining includes filtering pairs of the received word and particular ones of the plurality of words included in the image editing operation vocabulary in which a path that is usable to compute a semantic distance includes an ontology root.