Effective Token Graph Modeling using a Novel Labeling Strategy for Structured Sentiment Analysis

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4232 - 4241 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Effective Token Graph Modeling using a Novel Labeling Strategy for Structured Sentiment Analysis Wenxuan Shi, Fei Li, Jingye Li, Hao Fei, Donghong Ji† Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, Wuhan, China {shiwenxuan,lifei_csnlp,theodorelee,hao.fei,dhji}@whu.edu.cn Abstract The state-of-the-art model for structured sen- timent analysis casts the task as a dependency parsing problem, which has some limitations: (1) The label proportions for span prediction and span relation prediction are imbalanced. (2) The span lengths of sentiment tuple compo- nents may be very large in this task, which will further exacerbates the imbalance problem. (3) Two nodes in a dependency graph cannot have multiple arcs, therefore some overlapped senti- ment tuples cannot be recognized. In this work, we propose nichetargeting solutions for these issues. First, we introduce a novel labeling strategy, which contains two sets of token pair labels, namely essential label set and whole label set. The essential label set consists of the basic labels for this task, which are rela- tively balanced and applied in the prediction layer. The whole label set includes rich labels to help our model capture various token rela- tions, which are applied in the hidden layer to softly influence our model. Moreover, we also propose an effective model to well col- laborate with our labeling strategy, which is equipped with the graph attention networks to iteratively refine token representations, and the adaptive multi-label classifier to dynamically predict multiple relations between token pairs. We perform extensive experiments on 5 bench- mark datasets in four languages. Experimental results show that our model outperforms previ- ous SOTA models by a large margin.1 1 Introduction Structured Sentiment Analysis (SSA), which aims to predict a structured sentiment graph as shown in Figure 1(a), can be formulated into the problem of tuple extraction, where a tuple (h, e, t, p) denotes a holder h who expressed an expression e towards a target t with a polarity p. SSA is a more chal- lenging task, because other related tasks only focus †Corresponding author 1Our code is available at https://github.com/ Xgswlg/TGLS Moscow Government has expressed the wish to import the Mongolian meat Neutral Holder Expression Target Moscow Government has expressed the wish to import the Mongolian meat Exp : Neutral Holder Holder Exp : Neutral Exp : Neutral Target Target Target Target (a) (b) Moscow Government has expressed the wish to import the Mongolian meat Exp Head to Holder Head Exp Tail to Holder Tail Exp Tail to Target Tail Exp : Neutral Holder Target Exp Head to Target Head (c) Figure 1: (a) An example of structured sentiment analy- sis. (b) The head-first parsing graph proposed by Barnes et al. (2021), where the arcs related to holder(target)- expression linking relations are bold. (c) Our proposed essential label set, which has more balanced label …

Content truncated to 500 words