HuCurl: Human-induced Curriculum Discovery

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1862–1877 July 9-14, 2023 ©2023 Association for Computational Linguistics HuCurl: Human-induced Curriculum Discovery Mohamed Elgaar and Hadi Amiri Department of Computer Science University of Massachusetts Lowell {melgaar,hadi}@cs.uml.edu Abstract We introduce the problem of curriculum discov- ery and describe a curriculum learning frame- work capable of discovering effective curricula in a curriculum space based on prior knowl- edge about sample difficulty. Using annota- tion entropy and loss as measures of difficulty, we show that (i): the top-performing discov- ered curricula for a given model and dataset are often non-monotonic as apposed to mono- tonic curricula in existing literature, (ii): the prevailing easy-to-hard or hard-to-easy transi- tion curricula are often at the risk of underper- forming, and (iii): the curricula discovered for smaller datasets and models perform well on larger datasets and models respectively. The proposed framework encompasses some of the existing curriculum learning approaches and can discover curricula that outperform them across several NLP tasks. 1 Introduction Annotation information has been extensively used by previous research in NLP to devise strategies for further data collection (Yang et al., 2019; Dli- gach et al., 2010), model improvement and annota- tion analysis (Zaidan and Eisner, 2008; Paun et al., 2018), pruning and weighting samples for better learning (Yang et al., 2019), or efficient use of mon- etary funds (Dligach et al., 2010). Recent studies show consistent positive correlation between dif- ficulty of samples to the model and their level of human agreement (Nie et al., 2020a; Zaidan and Eisner, 2008; Yang et al., 2019). Building on these findings, we aim to utilize such prior knowledge about sample difficulty to develop a curriculum learning (CL) framework that is capable of discov- ering effective curricula for NLP tasks. A curriculum is a planned sequence of learning materials and an effective one can improve train- ing of NLP systems (Settles and Meeder, 2016; Amiri et al., 2017; Zhang et al., 2019; Lalor and Yu, 2020; Xu et al., 2020; Kreutzer et al., 2021; Agrawal and Carpuat, 2022; Maharana and Bansal, 2022). CL seeks to improve model generalizability by ordering samples for training based on their la- tent difficulty (Bengio et al., 2009). Recent work reported efficiency and effectiveness gains through CL (Jiang et al., 2018; Castells et al., 2020; Zhou et al., 2020), especially in cases of harder tasks and limited or noisy data (Wu et al., 2021). Existing CL approaches are designed to learn a single curriculum that works best for a given model and dataset. However, effective training could be achieved in multiple ways. In addition, existing ap- proaches quantify sample difficulty through model behavior during training. Although efficient and effective, model behavior can be affected by initial- ization and training dynamics (Erhan et al., 2010; Wu et al., 2021), which limits the curriculum space that can be examined for finding effective curricula. This paper advocates a re-imagining of CL paradigms by introducing and formalizing the task …

Content truncated to 500 words