Did You Read the Instructions? Rethinking the Effectiveness of Task Definitions in Instruction Learning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3063–3079 July 9-14, 2023 ©2023 Association for Computational Linguistics Did You Read the Instructions? Rethinking the Effectiveness of Task Definitions in Instruction Learning Fan Yin∗§, Jesse Vig♢†, Philippe Laban♢†, Shafiq Joty†, Caiming Xiong†, Chien-Sheng Jason Wu† §UCLA †Salesforce AI Research fanyin20@cs.ucla.edu {jvig, plaban, sjoty, wu.jason, cxiong}@salesforce.com Abstract Large language models (LLMs) have shown impressive performance in following natural language instructions to solve unseen tasks. However, it remains unclear whether models truly understand task definitions and whether the human-written definitions are optimal. In this paper, we systematically study the role of task definitions in instruction learning. We first conduct an ablation analysis informed by human annotations to understand which parts of a task definition are most important, and find that model performance only drops sub- stantially when removing contents describing the task output, in particular label information. Next, we propose an automatic algorithm to compress task definitions to a minimal support- ing set of tokens, and find that 60% of tokens can be removed while maintaining or even im- proving model performance. Based on these results, we propose two strategies to help mod- els better leverage task instructions: (1) provid- ing only key information for tasks in a com- mon structured format, and (2) adding a meta- tuning stage to help the model better understand the definitions. With these two strategies, we achieve a 4.2 Rouge-L improvement over 119 unseen test tasks. 1 Introduction Large language models or LLMs (Devlin et al., 2019; Raffel et al., 2020; Brown et al., 2020) demonstrate the ability to perform zero-shot cross- task generalization through learning from instruc- tions of tasks (Sanh et al., 2022; Wei et al., 2022a; Mishra et al., 2022; Wang et al., 2022b; Ouyang et al., 2022; OpenAI, 2023). By fine-tuning an LLM with task definitions and a few demonstration examples on upstream training tasks, the model acquires the power to perform new tasks with un- seen definitions and example. This is known as instruction learning. ∗Work done when Fan Yin was an intern at Salesforce. ♢Jesse and Philippe contributed equally; order is random. However, a natural question is: to what ex- tent does the zero-shot generalization ability de- rive from the model’s understanding of task defi- nitions? Recent work in prompt-based learning suggests models might not interpret even short prompts as people expect (Webson and Pavlick, 2022; Shin et al., 2020; Deng et al., 2022; Prasad et al., 2022). Task definitions are special prompts that are usually long and encode rich information. We imagine models’ understanding of definitions also departs from human expectation. To investi- gate this question, we conduct a systematic analysis using both human annotation and computational ap- proaches. Our study is based on the English portion of the large-scale SUPER-NATURALINSTRUCTION (NIv2) dataset (Wang et al., 2022b), which comprises 757 training tasks and 119 unseen test tasks. First, we explore which type of information in task definitions is necessary for maintaining …

Content truncated to 500 words