Empowering Active Learning to Jointly Optimize System and User Demands

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4233–4247 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 4233 Empowering Active Learning to Jointly Optimize System and User Demands Ji-Ung Lee Christian M. Meyer {lastname}@ukp.informatik.tu-darmstadt.de Ubiquitous Knowledge Processing (UKP) Lab Computer Science Department Technische Universit¨at Darmstadt, Germany https://www.ukp.tu-darmstadt.de Iryna Gurevych Abstract Existing approaches to active learning maxi- mize the system performance by sampling un- labeled instances for annotation that yield the most efficient training. However, when active learning is integrated with an end-user appli- cation, this can lead to frustration for partic- ipating users, as they spend time labeling in- stances that they would not otherwise be inter- ested in reading. In this paper, we propose a new active learning approach that jointly optimizes the seemingly counteracting objec- tives of the active learning system (training efficiently) and the user (receiving useful in- stances). We study our approach in an edu- cational application, which particularly bene- fits from this technique as the system needs to rapidly learn to predict the appropriateness of an exercise to a particular user, while the users should receive only exercises that match their skills. We evaluate multiple learning strate- gies and user types with data from real users and find that our joint approach better satisfies both objectives when alternative methods lead to many unsuitable exercises for end users.1 1 Introduction State-of-the-art machine learning approaches re- quire huge amounts of training data. But for many NLP applications, there is little to no training data available. Interactive NLP systems are a viable solution to alleviate the cost of creating large train- ing datasets before a new application can be used. Such systems start with no or few labeled instances and acquire additional training data based on user feedback for their predictions. Active learning (Set- tles, 2012) is a frequently used technique to quickly maximize the prediction performance, as the sys- tem acquires user feedback in each iteration for 1Our code and simulated learner models are avail- able on Github: https://github.com/UKPLab/ acl2020-empowering-active-learning those instances that likely yield the highest per- formance improvement (e.g., because the system is yet uncertain about them). Active learning has been shown to reduce the amount of user feed- back required while improving system performance for interactive NLP systems (P.V.S and Meyer, 2017; Gao et al., 2018) and to reduce the anno- tation costs in crowdsourcing scenarios (Fang et al., 2014). However, outside the typical annotation setup, it can be boring or frustrating for users to pro- vide feedback on ill-predicted instances that hardly solve their needs. Consider a newly launched web application for learning a foreign language, which aims at suggesting exercises that match the user’s proficiency according to Vygotsky’s Zone of prox- imal development (Vygotsky, 1978). The under- lying machine learning system starts without any data, but employs active learning to select an exer- cise the system cannot confidently predict. Then, it adjusts its model interactively based on the user’s feedback. While the system is still …

Content truncated to 500 words