Kurzweil also wrote a program that read poems, analyzed them, and then used what it observed to create its own original poetry following a similar style. The Ray Kurzweil Cybernetic Poet (RKCP) uses algorithms similar to neural networks to review poetry and generate poems that are similar. It can analyze one style or author to create poetry of a specific form or read several different styles and poets, merge them into one style, and create a seemingly original poem. New poems are based on words, word structures, rhythm structures, and the overall poem structure it perceives.
To do this, the program picks a word to start with. It sets up "goals" for this word, determining the type, size, and the sound of the next word. If multiple words are generated that meet those parameters, they are given "scores". Words with higher scores are more likely to be chosen. However, in order for the RKCP to write different original poems on an analysis of the same author, the same words will not always be chosen. In either case, if the program cannot find a new word that follows a piece of the poem, the program backs out of that piece of the poem, picks another word and tries again. It will back up to the beginning of the poem if it needs to.
If the RKCP cannot write a poem within a given set of constraints, it will loosen the constraints and try again until it can. It is also aware of syllable counts or specific sounds needed at the end of a line, enabling it to write haiku or other basic forms of poetry. In order to avoid plagiarism, the poetry program cannot place more than three words in a row that appear in the original text.
The Turing Test states that a computer is intelligent if it can convince a human that it is another human. Domain-specific Turing Tests use this design to test a specific area, such as chess playing. Ray Kurzweil tested a group of 16 adults and children with 28 poems, half of which were written by humans and the other half generated by his RKCP program. The adults guessed correctly 63% of the time while the children were correct 52% of the time. There was no visible correlation between scores and poetry or computer experience. Because both age groups scored approximately 50%, it is assumed that Kurzweil has, to some extent, succeeded in this domain-specific Turing Test. However, the real, long-range challenge will be getting the computer to create complete poems that incorporate genuine rhyme and meter over several stanzas.
See also: The Act of Creation (Music and Drawing)