Abstract
Fueled by recent developments in Neural Network research, the number and diversity of domains employing AI technology has expanded rapidly. This expansion is not limited to the automation of industrial processes and services, but also finds its way to the production of art. The application of AI to the domain of art raises a range of important questions and challenges at the intersection of authorship and authenticity. This study aims to advance and enhance our understanding of the (linguistic) properties that contribute to the perceived authenticity of a specific art form: Hip-Hop lyrics.
The empirical basis of our study is an experiment carried out in the context of a large, mainstream contemporary music festival. Using an experimental setup, reminiscent of the so-called Turing test, we crowdsourced a large dataset of authenticity judgements for both authentic and neurally generated Hip-Hop lyrics. The authenticity judgements enable us to quantitatively assess human biases toward artificially generated text, as well as which linguistic characteristics are perceived as authenticity cues. Additionally, the dataset provides solid ground for thoroughly evaluating different neural language generation systems with respect to their perceived credibility. More specifically, we compare the credibility of character-level, syllable-level and hierarchical language models as well as the effect of conditioning such models on domain-specific features, such as rhyme. Our experiments contribute to research into improving the credibility of generated text, and enhances our understanding of cognitive processes at play in the perception of authentic and artificial art.
The empirical basis of our study is an experiment carried out in the context of a large, mainstream contemporary music festival. Using an experimental setup, reminiscent of the so-called Turing test, we crowdsourced a large dataset of authenticity judgements for both authentic and neurally generated Hip-Hop lyrics. The authenticity judgements enable us to quantitatively assess human biases toward artificially generated text, as well as which linguistic characteristics are perceived as authenticity cues. Additionally, the dataset provides solid ground for thoroughly evaluating different neural language generation systems with respect to their perceived credibility. More specifically, we compare the credibility of character-level, syllable-level and hierarchical language models as well as the effect of conditioning such models on domain-specific features, such as rhyme. Our experiments contribute to research into improving the credibility of generated text, and enhances our understanding of cognitive processes at play in the perception of authentic and artificial art.
| Original language | English |
|---|---|
| Publication status | Published - 31 Jan 2019 |