Basic poetry generation

• US 7,184,949 B2
• Filed: 05/10/2002
• Issued: 02/27/2007
• Est. Priority Date: 11/01/1999
• Status: Active Grant

• Current Assignee: Kurzweil Cyberart Technologies Incorporated
• Original Assignee: Kurzweil Cyberart Technologies Incorporated
• Inventors: Kurzweil, Raymond, Keklak, John A.
• Primary Examiner(s): SPOONER, LAMONT M
• Application Number:
  US10/143,274
• Publication Number:
  US 20030036040A1
• Time in Patent Office: 1,754 Days
• Field of Search: 704/9, 704/1
• US Class Current: 704/9
• CPC Class Codes: G06F 40/20 Natural language analysis semantic analysis of natural language G06F40/30
• CPC Class Codes: G06F 40/211 Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06F 40/284 Lexical analysis, e.g. tokenisation or collocates
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: G06F 40/56 Natural language generation
• CPC Class Codes: Y10S 707/99943 Generating database or data structure, e.g. via user interface

Abstract

A method of analyzing an author'"'"'s work, including reading a text file, generating an analysis model from the text file and storing the analysis model. The text file may be a poem, the poem containing ASCII text. The analysis model is a linked data structure. The linked data structure includes n-gram data structures, the n-gram structures including combinations of 1-gram data structures, bigram data structures, trigram data structures and quadrigram data structures.

Claims

1. A computer-implemented method for generating poetry, comprising:

analyzing in a computer system at least one pre-existing poem to generate an author analysis model from the pre-existing poem, with analyzing further comprising;

producing linked n-gram data structures, the n-gram data structures comprising 1-grams, bigrams, trigrams, and quadrigrams in both a forward and a backward direction;

generating a new poem from the author analysis model; and

outputting the new poem to an output device, rendering the new poem to a user.

2. The computer-implemented method of claim 1 wherein, analyzing further comprising:

generating the new poem from the author analysis model including for any subsequent user input word, using the linked n-gram data structures to locate the user input word in the linked n-gram data structures and determine words that follow it in the linked n-gram structures.

3. The computer-implemented method of claim 2 wherein the n-gram data structures include 1-grams, bigrams, trigrams, and quadrigrams.

4. The computer-implemented method of claim 2 wherein the n-gram data structures are selected from the group consisting of 1-grams, bigrams, trigrams, and quadrigrams.

5. The computer-implemented method of claim 1 wherein the output device is a display device.

6. The computer-implemented method of claim 1 wherein the output device is a storage device.

7. The computer-implemented method of claim 1 with the word ngram frequency counts comprising quadrigrams that have a negative weight to avoid sequences of four words that appeared in text that is analyzed to produce the ngrams.

8. An automatic composition system comprising:

a computer readable medium for storage of computer programs;

a central processing unit that executes instructions from programs stored on the computer readable medium;

a random access memory to provide memory execution space to execute programs stored on the computer readable medium;

a display unit; and

a computer program product residing on the computer readable medium for automatically composing text on the display unit by accessing word ngram frequency counts in a determination of a next word to compose, with selected word ngram frequency counts having a negative weight to avoid sequences of words that appear in text that is analyzed to produce the ngrams.

9. The system of claim 8 wherein the ngrams comprise bigrams.

10. The system of claim 8 wherein the ngrams comprise trigrams.

11. The system of claim 8 wherein the ngrams comprise two quadrigrams.

12. The system of claim 11 wherein the quadrigrams have a negative weight to avoid sequences of four words that appear in text that is analyzed to produce the ngrams.

13. The system of claim 8 wherein accessing word ngram frequency counts, accesses n-gram data structures selected from the group consisting of 1-grams, bigrams, trigrams, and quadrigrams.

14. The system of claim 8 wherein accessing word ngram frequency counts, accesses n-gram data structures selected from the group consisting of 1-grams, bigrams, trigrams, and quadrigrams in both a forward and backward direction.

15. The system of claim 8 wherein words of text include an end of line marker.

16. The system of claim 15 wherein the end of line marker designates an end of line in the text.

17. The system of claim 8 wherein words of text includes an end of poem marker.

18. The system of claim 17 wherein the end of poem marker designates an end of text.

19. The system of claim 8 wherein the text is generated recursively to meet selected criteria.

20. The system of claim 19 wherein the selected criteria include a line length.

21. The system of claim 19 wherein the selected criteria include a number of syllables.

22. The system of claim 19 wherein the selected criteria include ending a line with a first word of text that bears a relationship to a second word of text.

23. The system of claim 22 wherein the relationship is that the first word of text rhymes with the second word of text.

24. The system of claim 8 wherein the computer program product for automatically composing comprises instructions to:

determine a next word to composed using word ngram frequency counts.

25. An automatic composition system comprising:

a computer readable medium for storage of computer programs;

a central processing unit that executes instructions from programs stored on the computer readable medium;

a random access memory to provide memory execution space to execute programs stored on the computer readable medium;

an output device; and

a computer program product residing on the computer readable medium for automatically composing text on the output device with an author analysis model based on pre-existing compositions, including instructions to;

examine weights on the author analysis model to avoid counts that would tend to repeat same words from the pre-existing compositions given a start word in an analyzed one of the pre-existing compositions to avoid plagiarism.

26. The system of claim 25 wherein instructions to avoid plagiarism further comprises instructions to:

avoid a first sequence of words matching a second sequence of words.

27. The system of claim 25 wherein word ngram frequency counts are accessed to determine a next word to add to a sequence of words.

28. The system of claim 25 wherein the computer program further comprises instructions to:

analyze at least one pre-existing poem to generate an author analysis model from the pre-existing poem, with the author analysis model further comprising linked n-gram data structures.

29. The system of claim 28 wherein the instructions to automatically compose further comprises instructions to:

generate a new composition from the author analysis model including for any subsequent user input word, using the linked n-gram data structures to locate the user input word in the linked n-gram data structures and determine words that follow it in the linked n-gram data structures.

30. The system of claim 25 wherein the instructions to automatically compose further comprises instructions to:

generate a new composition from the author analysis model including linked data structures of comprising 1 grams, bigrams, trigrams and quadrigrams.

31. The system of claim 25 wherein the computer program further comprises instructions to:

assign high negative weights on quadrigrams compared to weights assigned on 1 grams, bigrams, and/or trigrams, such that generated compositions do not plagiarize original compositions.

32. The system of claim 25 wherein ngrams can be characterized as having an ngram count, which is the ngram count for a word for which a score is computed, an ngram weight, which is a weight for the ngram within an author model, and an ngram exponent, which is an exponent for the ngram count within the author model.

33. The system of claim 32 further comprising instructions to produce a score for a new word including instructions to:

compute for each new word input by a user the score based on the ngram count raised to the ngram exponent, and multiplied by the ngram weight.

34. The system of claim 33 wherein a high exponent will tend to select the highest count and thus be likely to repeat a poem given the same start word, compared to a low exponent and a lower count.

35. The system of claim 33 wherein lower valued exponents will produce a relatively large number of different poems relative to a higher valued exponent, but will follow the original author less closely than the higher valued exponent.

36. The system of claim 33 wherein an exponent of 1 will weight probabilities as in the at least one pre-existing poem.

37. The system of claim 33 wherein exponents between 0 and 1 will emphasize higher counts for scores computed for words compared to higher exponent values.

38. The system of claim 33 wherein an exponent of 0 will ignore all counts and will give possible bigram and trigram words an equal weight regardless of their frequency in the author analysis model.

39. The system of claim 33 wherein a negative exponent will give preference to lower counts for scores computed for words.

40. The system of claim 33 wherein bigram and trigram of original author weights provide a relative influence of two original authors, a trigram original author and a quadrigram original author in a generated composition, where the trigram original author will tend to produce word sequences that more closely follow the trigram original author, while plagiarism is avoided by assigning a strong negative weight on the quadrigram original author.

41. A method for composition generation comprises:

automatically composing text on an output device with an author analysis model based on pre-existing compositions; and

examining weights on the author analysis model to avoid counts defined in the author analysis model that would tend to repeat same word sequences from the pre-existing compositions given a start word in an analyzed one of the pre-existing compositions in order to avoid plagiarism.

42. The method of claim 41 further comprising:

avoiding a first sequence of words that matches a second sequence of words.

43. The method of claim 41 wherein word ngram frequency counts are accessed to determine a next word to add to a sequence of words.

44. The method of claim 41 further comprising:

analyzing at least one pre-existing composition to generate an author analysis model from the pre-existing composition, with the author analysis model further comprising linked n-gram data structures; and

generating a new composition from the author analysis model including for any subsequent user input word, using the linked n-gram data structures to locate the user input word in the linked data structures and determine words that follow it in the linked n-gram data structures.

45. The method of claim 41 wherein composing further comprises:

generating a new composition from the author analysis model including linked data structures of comprising 1 grams, bigrams, trigrams and quadrigrams.

46. The method of claim 45 wherein generating a new composition further comprises:

assigning a relatively high negative weights on quadrigrams, compared to 1 grams, bigrams, trigrams weights, such that generated compositions do not plagiarize original compositions.

47. The method of claim 46 wherein ngrams can be characterized as having an ngram count, which is the ngram count for a word for which a score is computed, an ngram weight, which is a weight for the ngram for an author model, and an ngram exponent, which is an exponent for the ngram count for author model.

48. The method of claim 47 further comprising:

producing a score for a new word by computing for each new word input by a user the score based on the ngram count raised to the ngram exponent, and multiplied by the ngram weight.

49. A computer readable medium storing a computer program product for producing poetry, the computer program product comprises instructions to cause a computer to:

automatically composing text on a display unit by accessing word ngram frequency counts in a determination of a next word to compose with the word ngram frequency counts comprising quadrigrams that have a negative weight to avoid sequences of four words that appeared in text that is analyzed to produce the ngrams.

50. The computer program product of claim 49 wherein the ngram frequency counts comprise bigrams.

51. The computer program product of claim 49 wherein the ngram frequency counts comprise trigrams.

52. The computer program product of claim 49 wherein the ngram frequency counts comprise n-gram data structures comprising 1-grams, bigrams, trigrams, and quadrigrams.

53. The computer program product of claim 49 wherein the ngram frequency counts comprise n-gram data structures comprising 1-grams, bigrams, trigrams, and quadrigrams in both a forward and a backward direction.

54. The computer program product of claim 49 wherein the text is generated recursively to meet selected criteria.

55. The computer program product of claim 54 wherein the selected criteria include a line length, a number of syllables, and/or ending a line with a first word of text that bears a rhyme relationship to a second word of text.

56. The computer program product of claim 49 wherein the computer program product further comprises instructions to:

determine a next word to composed using word ngram frequency counts.

57. The computer program product of claim 49 further comprising instructions to:

examine weights on an author analysis model to avoid counts that would tend to repeat same words from pre-existing compositions given a start word in an analyzed one of the preexisting compositions to avoid plagiarism.