When is multitask learning effective? Semantic sequence prediction under varying data conditions
This work addresses the problem of determining effective conditions for MTL in semantic sequence prediction, which is incremental as it extends prior morphosyntactic findings to semantic tasks.
The paper investigated when multitask learning (MTL) is effective for semantic sequence labeling tasks, finding that it significantly improves performance for only 1 out of 5 tasks, with success linked to auxiliary tasks having compact and uniform label distributions.
Multitask learning has been applied successfully to a range of tasks, mostly morphosyntactic. However, little is known on when MTL works and whether there are data characteristics that help to determine its success. In this paper we evaluate a range of semantic sequence labeling tasks in a MTL setup. We examine different auxiliary tasks, amongst which a novel setup, and correlate their impact to data-dependent conditions. Our results show that MTL is not always effective, significant improvements are obtained only for 1 out of 5 tasks. When successful, auxiliary tasks with compact and more uniform label distributions are preferable.