Sentiment Tagging with Partial Labels using Modular Architectures

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 579–590 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 579 Sentiment Tagging with Partial Labels using Modular Architectures Xiao Zhang Purdue University zhang923@purdue.edu Dan Goldwasser Purdue University dgoldwas@purdue.edu Abstract Many NLP learning tasks can be decomposed into several distinct sub-tasks, each associated with a partial label. In this paper we focus on a popular class of learning problems, sequence prediction applied to several sentiment analy- sis tasks, and suggest a modular learning ap- proach in which different sub-tasks are learned using separate functional modules, combined to perform the final task while sharing infor- mation. Our experiments show this approach helps constrain the learning process and can alleviate some of the supervision efforts. 1 Introduction Many natural language processing tasks attempt to replicate complex human-level judgments, which often rely on a composition of several sub-tasks into a unified judgment. For example, consider the Targeted-Sentiment task (Mitchell et al., 2013), assigning a sentiment polarity score to entities de- pending on the context that they appear in. Given the sentence “according to a CNN poll, Green Book will win the best movie award”, the sys- tem has to identify both entities, and associate the relevant sentiment value with each one (neutral with CNN, and positive with Green Book). This task can be viewed as a combination of two tasks, entity identification, locating contiguous spans of words corresponding to relevant entities, and sen- timent prediction, specific to each entity based on the context it appears in. Despite the fact that this form of functional task decomposition is natural for many learning tasks, it is typically ignored and learning is defined as a monolithic process, com- bining the tasks into a single learning problem. Our goal in this paper is to take a step to- wards modular learning architectures that exploit the learning tasks’ inner structure, and as a re- sult simplify the learning process and reduce the annotation effort. We introduce a novel task de- composition approach, learning with partial la- bels, in which the task output labels decompose hierarchically, into partial labels capturing differ- ent aspects, or sub-tasks, of the final task. We show that learning with partial labels can help sup- port weakly-supervised learning when only some of the partial labels are available. Given the popularity of sequence labeling tasks in NLP, we demonstrate the strength of this approach over several sentiment analysis tasks, adapted for sequence prediction. These include target-sentiment prediction (Mitchell et al., 2013), aspect-sentiment prediction (Pontiki et al., 2016) and subjective text span identification and polar- ity prediction (Wilson et al., 2013). To ensure the broad applicability of our approach to other prob- lems, we extend the popular LSTM-CRF (Lample et al., 2016) model that was applied to many se- quence labeling tasks1. The modular learning process corresponds to a task decomposition, in which the prediction la- bel, y, is deconstructed into a set of partial la- bels {y0, .., yk}, …

Content truncated to 500 words