Article Preview
TopIntroduction
Sarcasm presents a negative meaning using positive expression. The surface sentiment of words in sarcastic text is not always the same as the intended sentiment.
In Example 1, even though the writer uses the positive word “good,” the intention is to criticize the boss. A basic task in sentiment analysis is to classify sentences as positive or negative, and the surface sentiment of words is an important feature for such classification. However, often, sarcastic sentences are misclassified; thus, sarcasm detection is a significant challenge in sentiment analysis. To identify sarcastic sentences, we must consider both surface and non-surface features.
In this study, we focus on relation information as a non-surface feature, and we propose a method to detect sarcastic sentences based on the relations between the target of sarcasm and the person who is dissatisfied with the target. Note that sarcasm typically involves a specific target (Campbell & Katz, 2012). Kumon-Nakamura et al. (1995) claimed that ironic remarks have effect by alluding to a failed expectation. Sarcasm is considered a type of irony. In Example 1, the boss is the target, and the writer criticizes the boss from the perspective of staff who are dissatisfied with the boss’ behavior. The implied expectation that the boss should be a good manager affects the interpretation of sarcasm. To recognize sarcasm in a sentence, the differences in the relation inferred by the role expressions are important features. Consider the following example.
In Example 2, we replace “boss” and “staff” with “teacher” and “students,” respectively. Compared to Example 1, Example 2 is a non-sarcastic sentence because the relationship between teachers and students differs from that of a boss and staff. In this study, we focus on the relationship between two roles, e.g., “boss and staff,” “student and teacher,” and “parent and child.” The proposed method uses this relationship to detect sarcasm. In addition, the proposed method uses a vector to express the relations and uses a Recurrent Neural Network (RNN) to detect sarcastic sentences. The RNN is trained using two types of features, i.e., words in sentences and a vector that expresses the relations between role pairs, e.g., “boss and staff.” We extract role pairs from a corpus to obtain a list of role pairs. We construct the relation vector generated from a role and evaluate the effectiveness of the vector to express the relation of role pairs in a sarcasm detection task.
TopIn previous studies, sarcasm detection has been treated as a task that classifies text as sarcastic or non-sarcastic. Many researchers have used machine learning approaches to classify text. Words or phrases are commonly used as features in machine learning (Davidov et al., 2010; Reyes & Rosso., 2011; Joshi et al., 2015, 2016). However, for sarcasm detection, we must apply both surface and non-surface features. Often, perceiving sarcasm requires background knowledge. For example, to understand that Example 1 is a sarcastic sentence, we must know the relationship between the “boss” and “staff.” In contrast, Example 2 is understood as a non-sarcastic sentence because we have a general understanding of the relationship between the “teacher” and “students.”
Consider the following example, where “boss” and “staff” are replaced by “manager” and “clerks,” respectively.
Here, the relationship between “manager” and “clerks” is similar to the relationship between “boss” and “staffs;” thus, we understand that Example 3 is a sarcastic sentence. The differences and similarities of the relations in a role pair are important to detect sarcasm. Therefore, we use role pair relations to detect sarcastic sentences.
Figure 1. Overview of the proposed method