Design and Analysis of a Collaborative Story Generation Game for Social Robots

Abstract

Storytelling plays a central role in human socializing and entertainment, and research on conducting storytelling with robots is gaining interest. However, much of this research assumes that story content is curated. In this paper, we introduce the task of collaborative story generation, where an artificial intelligence agent, or a robot, and a person collaborate to create a unique story by taking turns adding to it. We present a collaborative story generation system which works with a human storyteller to create a story by generating new utterances based on the story so far. Our collaborative story generation system consists of a publicly-available large scale language model that was tuned on a dataset of writing prompts and short stories, and a ranker that samples from the language model and chooses the best possible output. We improve storytelling quality by optimizing the ranker’s sample size to strike a balance between quality and computational cost. Since latency can be detrimental to human-robot interaction, we examine the performance-latency trade-offs of our approach and find the optimal ranker sample size that strikes the best balance between quality and computational cost. We evaluate our system by having human participants play the collaborative story generation game and comparing the stories they create with our system to a naive baseline. Next, we conduct a detailed elicitation survey that sheds light on issues to consider when adapting our collaborative story generation system to a social robot. Finally, in a first step towards allowing human players to control the genre or mood of stories generated, we present preliminary work on steering story generation sentiment polarity with a sentiment analysis model. We find that our proposed method achieves a good balance of steering capability and text coherence. Our evaluation shows that participants have a positive view of collaborative story generation with a social robot and consider rich, emotive capabilities to be key to an enjoyable experience.

Would you be impressed: applying principles of magic to chatbot conversations

“…Following power analysis, we collected survey responses with the goal of exceeding participants. A university with an established history of evaluating conversations with social robots ( Nichols et al, 2021a ; Nichols et al 2022a ; Nichols et al 2023 ) was used to recruit survey participants and conduct the survey.…”

“…Haru’s range of communicative strategies positions the robot as a potent embodied communication agent capable of long-term interaction with people. Haru has been deployed in a variety of scenarios, including classroom companion, interactive storyteller ( Nichols et al, 2021a ), and, of course, conversational partner ( Nichols et al, 2022a ).…”

Developing Autonomous Robot-Mediated Behavior Coaching Sessions with Haru

“…The robot is using Google Speech-to-Text API for automatic speech recognition (ASR), and Intent Classification and Entity Recognition Models developed by Honda Research Institute. The dialogue structure is then loaded into behavior trees, with all the components interacting with the robot hardware using ROS [27]. This system, which combines off-the-shelf components with custom robot elements, was previously tested in other applications [28], such as small talk scenarios and hospital dialogues.…”

CreativeBot: a Creative Storyteller Agent Developed by Leveraging Pre-trained Language Models

“…build an autonomous agent capable of collaborative storytelling with adult human users is presented in [28,29]. The implemented agent exhibited creativity by introducing plot twists to the story.…”

“…Another attempt for story generation utilised the "yes-and" principle that proposes accepting a suggestion provided by another participant in a collaborative storytelling setting and building on it [5]. A recent attempt of generating a story collaboratively between a human and an agent was presented in [28,29], where the authors have used GPT-2 tuned on a large provided data set.…”