When Creative AI Meets Conversational AI

Abstract

Following the rapid developing of artificial intelligence (AI) boosted by deep neural networks, we are wondering how far is it going for AI to be really creative. We know that human beings are intelligent in terms of creating novel things from scratch. Art can be selected as one type of Turing-Test : given two paintings, can we distinguish which one is painted by AI and which one is by a real-world person? This same question stands for music creating, poem writing, singing and so on. We further ask ourselves, can we be involved in the whole AI creation process? Say, we communicate with AI to exchange ideas of what’s the next painting topic or genre, or what’s the next movement of the music. Through these interactive AI creation processes, we hope to inspire AI and be inspired by AI, just alike we learn from Alpha-Go after it learnt from us first and then by itself through self-supervised learning. For Go gaming, we know the rules of win or lose. However, for AI creation, evaluation metrics are subjectively defined in creative ways. In this year’s (27th) Natural Language Processing (NLP 2021) conference, we proposed and organized the first workshop named “when creative AI meets conversational AI”, or, briefly “CAI+CAI=CAI”. In this one-day workshop, there are two technical papers and eight invited talks. These two technical papers respectively and interestingly cover one direction of text-toimage leveraging DM-GAN (Zhu et al. 2019) and ManiGAN (Li et al. 2020) for bird image generation (Azuaje et al. 2021) and its reversed direction of image-to-text by visual-text integrated Transformer (森 他 2021) for story generation. The eight invited talks cover most major directions of creative AI and conversational AI, such as voice conversion (Zhao et al. 2020), image generation (Jiang et al. 2021), contents retrieving (Yu et al. 2019), fine-art paintings (Huckle et al. 2020), image-to-image translation (Guo et al. 2021), AI painting, music generating and singing (Wu et al. 2020), poem-writing (Wu et al. 2017), and multi-modal learning for medical and healthcare applications (Ma et al. 2021; Obinata et al. 2020).