Browsing by Subject "ai"

Now showing 1 - 1 of 1

Towards Computer Generation Of Text
(2019-05-01) Ceverha, Jack; Pennebaker, James
This project explores the challenges of generating theatrical text in the form of a script using automated and semi-automated techniques. Two different systems, a generative system built on neural language models and an original system called Mosaic, were trained on subsets of a corpus of 127 full-length English-language plays. These systems generate short plays using the outline of a handwritten narrative as a source of structure.The language model was used to generate a large amount of plays that were curated and analyzed qualitatively to demonstrate the play’s cohesiveness and the extent to which the generated text matches the handwritten narrative. The system produced suboptimal results in this task -the language model was unable to generate cohesive plays that matched supplied narrativesPlays generated by Mosaic were analyzed qualitatively and through a large experiment that used numerical survey feedback from 300 human respondents on the crowdsourcing platform Amazon Mechanical Turk. Four sets of sixteen-line plays with two characters were compiled using four different protocols. The first set consisted of fifty direct excerpts from the play corpus, while the second set contained fifty plays with lines from the corpus arranged in a random order. These two sets represent the control data. The third set contains fifty plays generated by Mosaic. With the help of ten volunteers, the fourth set contained fifty plays that were written interactively. A volunteer would write a play with Mosaic -Mosaic wrote lines for the first character, and the volunteer wrote lines for the second character.Both experimental sets of plays consistently overperformed the random control to statistically significant degree, with the interactively-writtenplays scoring within five percent of the corpus