Contrastive Learning enhanced Author-Style Headline Generation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 505 - 525 December 7-11, 2022 ©2022 Association for Computational Linguistics InstructDial: Improving Zero and Few-shot Generalization in Dialogue through Instruction Tuning Prakhar Gupta♣ Cathy Jiao♣ Yi-Ting Yeh♣ Shikib Mehri♣ Maxine Eskenazi♣ Jeffrey P. Bigham♣,♡ ♣Language Technologies Institute, Carnegie Mellon University ♡Human-Computer Interaction Institute, Carnegie Mellon University {prakharg,cljiao,yitingye,amehri,max,jbigham}@cs.cmu.edu Abstract Instruction tuning is an emergent paradigm in NLP wherein natural language instructions are leveraged with language models to induce zero- shot performance on unseen tasks. Dialogue is an especially interesting area in which to ex- plore instruction tuning because dialogue sys- tems perform multiple tasks related to language (e.g., natural language understanding and gen- eration, domain-specific interaction), yet in- struction tuning has not been systematically explored for dialogue-related tasks. We intro- duce INSTRUCTDIAL, an instruction tuning framework for dialogue, which consists of a repository of 48 diverse dialogue tasks in a uni- fied text-to-text format created from 59 openly available dialogue datasets. We explore cross- task generalization ability on models tuned on INSTRUCTDIAL across diverse dialogue tasks. Our analysis reveals that INSTRUCTDIAL en- ables good zero-shot performance on unseen datasets and tasks such as dialogue evaluation and intent detection, and even better perfor- mance in a few-shot setting. To ensure that models adhere to instructions, we introduce novel meta-tasks. We establish benchmark zero-shot and few-shot performance of mod- els trained using the proposed framework on multiple dialogue tasks1. 1 Introduction Pretrained large language models (LLMs) (Devlin et al., 2019; Radford et al., 2019; Brown et al., 2020) are not only few-shot learners, but can also perform numerous language tasks without the need for fine-tuning. However, LLMs are expensive to train and test. Instruction tuning has emerged as a tool for directly inducing zero-shot generaliza- tion on unseen tasks in language models by using natural language instructions (Mishra et al., 2021; Sanh et al., 2022; Wei et al., 2022; Ouyang et al., 1Code available at https://github.com/ prakharguptaz/Instructdial Figure 1: We investigate instruction tuning on dialogue tasks. Instruction tuning involves training a model on a mixture of tasks defined through natural language instructions. Instruction tuned models exhibit zero-shot or few-shot generalization to new tasks. 2022). Natural language instructions can contain components such as task definitions, examples, and prompts which allows them to be customized for multitask learning. Instruction tuning enables de- velopers, practitioners, and even non-expert users to leverage language models for novel tasks by specifying them through natural language, without the need for large training datasets. Furthermore, instruction tuning can work for models that are sig- nificantly smaller than LLMs (Mishra et al., 2021; Sanh et al., 2022), making them more practical and affordable. Most recent work (Mishra et al., 2021; Sanh et al., 2022; Wei et al., 2022) on instruction tuning has focused on general NLP tasks such as para- phrase detection and reading comprehension, but not specifically on dialogue. While some work 505 such as (Wang et al., 2022a) include a few dialogue tasks, those tasks are …

Content truncated to 500 words