Incorporating Task-Specific Concept Knowledge into Script Learning

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3026–3040 May 2-6, 2023 ©2023 Association for Computational Linguistics Incorporating Task-Specific Concept Knowledge into Script Learning Chenkai Sun, Tie Xu, ChengXiang Zhai, Heng ji University of Illinois at Urbana-Champaign, IL, USA Alibaba, Hangzhou, China {chenkai5, czhai, hengji}@illinois.edu xutie.xt@alibaba-inc.com Abstract In this paper, we present TETRIS, a new task of Goal-Oriented Script Completion. Unlike previous work, it considers a more realistic and general setting, where the input includes not only the goal but also additional user context, including preferences and history. To address this problem, we propose a novel approach, which uses two techniques to improve perfor- mance: (1) concept prompting, and (2) script- oriented contrastive learning that addresses step repetition and hallucination problems. On our WikiHow-based dataset, we find that both meth- ods improve performance. 1 Introduction A Goal-Oriented Script refers to a sequence of events that describe some stereotypical activities for achieving a specified goal (Feigenbaum et al., 1981; Lyu et al., 2021). It is important to study how to automatically construct instructional scripts because it enables many high-impact applications such as robot action planning (Conti et al., 2020; Mohanan and Salgoankar, 2018), causal reason- ing (Guo et al., 2020), and task-oriented dialogue generation (Chen et al., 2017). While previous work has shown that neural models are capable of constructing the entire script given a goal (Lyu et al., 2021; Sakaguchi et al., 2021), their proposed tasks are highly restrictive and based on overly sim- plified assumptions about the application because it ignored both the usage context (e.g., what is a pre- ferred way of solving the problem, and what steps have been executed already) and the variable per- sonal preferences that a user may have (e.g., a goal might need be achieved in different ways). Take “Make Eggless Cupcakes” as an example, the user might want to make it with different equipment (e.g., by using a rice cooker instead of an oven) or styles (e.g., to have some banana flavor), or the https://github.com/chenkaisun/Tetris Goal: Wash Linen Shirts Added detergent I �lled a sink with cool water Wash the shirt by swishing it around in the sudsy water. Soak the shirt in the tub for 10-20 minutes. Rinse the shirt with water until all the suds are gone. Washing by Hand Hang up your shirt and let it air-dry for the best results. History Preference Figure 1: An example illustrating the task. The input consists of a specified goal (e.g., wash line shirts), an op- tional preference (e.g., washing by hand), and a history of steps. The model is asked to generate the remaining steps to achieve the goal. user has already completed some steps but would like to seek information on what to proceed next; a machine learning model trained on goal-to-script tasks as done in the existing work cannot adaptively provide solutions under such situations. We address the limitations of the previous task formulation and propose TETRIS, a more general task …

Content truncated to 500 words