Article Preview
TopIntroduction
In emotion-cause extraction (ECE), the underlying causes that lead to the emotional polarity of texts are extracted. Since ECE was proposed, it has received widespread attention in natural language processing (Lee et al., 2010). ECE has great significance for customer evaluation analysis and public opinion monitoring. For example, in terms of business, it can combine the work of Alharbi et al. (2020) to better analyze users' satisfaction with suppliers and help facilitate the decision-making process of the consumer (Abayomi-Alli et al., 2021; Guebli & Belkhir, 2021). However, emotion labels are required in ECE, and emotion annotation is quite labor intensive, thus limiting the applicability of ECE in practice (Gui et al., 2016; Gui et al., 2017). To address this limitation, the emotion-cause pair extraction (ECPE) task was proposed (Xia & Ding, 2019). ECPE is a new task in which emotion clauses and the corresponding cause clauses in unannotated emotion texts are identified. This involves two main subtasks. The first subtask is extracting the emotion and cause clauses from the unannotated emotion text. The second subtask is matching the emotion clauses with the corresponding cause clauses and removing the nonexistent causal relationships. An example that demonstrates the difference between the ECE task and the ECPE task is presented in Figure 1. The objective of ECE is to track two corresponding cause clauses: “C2 A policeman visited the old man with the lost money” and “C3 and told him that the thief was caught”. In the ECPE task, the objective is to extract all pairs of emotion clauses and cause clauses (“C4 The old man was very happy” and “C2 A policeman visited the old man with the lost money” and “C4 The old man was very happy” and “C3 and told him that the thief was caught”). From the above comparison, we find ECPE to be much more challenging. For ECPE, the authors’ model must be able to identify the structure and content of the text and then accurately extract the corresponding emotion-cause pairs from the text.
Figure 1. Example that demonstrates the differences between ECE and ECPE
Xia and Ding (2019) proposed a two-step framework for the ECPE task. In the first step, a multitask long short-term memory (LSTM) is used to extract emotion clauses and cause clauses. Then, a classifier is used to filter out negative candidate clause pairs. However, in a two-step approach, any misclassification from the first step is magnified in the second step. To overcome this drawback related to error propagation, the framework of emotional cause clause extraction should be considered as an integral framework. Hence, the authors propose a one-step framework, which is referred to as the semantic decision internal-attention graph convolutional network (SIGCN). This end-to-end approach can extract both the emotion-clause and the cause clause, thus avoiding the propagation of errors across stages.
Fan et al. (2020) transformed the task into a process that is similar to analyzing the construction of a directed graph. However, LSTM still has the disadvantage of long-range dependence when capturing the hidden state of the clause (Shi et al., 2015; Geng et al., 2020). Graph convolutional networks (GCNs) have made breakthrough progress in addressing long-range dependencies (C. Zhang et al., 2019; M. Zhang et al., 2018). A GCN treats each clause as a node in a graph and identifies the hidden state that contains the semantic structure of the text through the information transfer between each pair of nodes.