A Survey of Active Learning for Natural Language Processing

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6166 - 6190 December 7-11, 2022 ©2022 Association for Computational Linguistics A Survey of Active Learning for Natural Language Processing Zhisong Zhang, Emma Strubell, Eduard Hovy Language Technologies Institute, Carnegie Mellon University zhisongz@cs.cmu.edu, strubell@cmu.edu, hovy@cmu.edu Abstract In this work, we provide a literature review of active learning (AL) for its applications in natu- ral language processing (NLP). In addition to a fine-grained categorization of query strategies, we also investigate several other important as- pects of applying AL to NLP problems. These include AL for structured prediction tasks, an- notation cost, model learning (especially with deep neural models), and starting and stopping AL. Finally, we conclude with a discussion of related topics and future directions. 1 Introduction The majority of modern natural language process- ing (NLP) systems are based on data-driven ma- chine learning models. The success of these models depends on the quality and quantity of the avail- able target training data. While these models can obtain impressive performance if given enough su- pervision, it is usually expensive to collect large amounts of annotations, especially considering that the labeling process can be laborious and challeng- ing for NLP tasks (§3.2). Active learning (AL), an approach that aims to achieve high accuracy with fewer training labels by allowing a model to choose the data to be annotated and used for learning, is a widely-studied approach to tackle this labeling bottleneck (Settles, 2009). Active learning has been studied for more than twenty years (Lewis and Gale, 1994; Lewis and Catlett, 1994; Cohn et al., 1994, 1996) and there have been several literature surveys on this topic (Settles, 2009; Olsson, 2009; Fu et al., 2013; Aggar- wal et al., 2014; Hino, 2020; Schröder and Niekler, 2020; Ren et al., 2021; Zhan et al., 2022). Nev- ertheless, there is still a lack of an AL survey for NLP that includes recent advances. Settles (2009) and Olsson (2009) provide great surveys covering AL for NLP, but these surveys are now more than a decade old. In the meantime, the field of NLP has been transformed by deep learning. While 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 0 5 10 15 20 Counts (AL) AL-Conference AL-Workshop 0 100 200 300 400 500 600 Counts (Neural) Neural Figure 1: Counts of AL (left) and “neural” (right) pa- pers in the ACL Anthology over the past twenty years. other more recent surveys cover deep active learn- ing, they are either too specific, focused only on text classification (Schröder and Niekler, 2020), or too general, covering AI applications more broadly (Ren et al., 2021; Zhan et al., 2022). Moreover, applying AL to NLP tasks requires specific consid- erations, e.g. handling complex output structures and trade-offs in text annotation cost (§3), which have not been thoroughly discussed. In order to provide an NLP-specific AL survey,1 we start by searching the ACL Anthology for AL- …

Content truncated to 500 words