Text-based emotion classification using emotion cause extraction

https://doi.org/10.1016/j.eswa.2013.08.073Get rights and content

Highlights

  • Our work is the first emotion classification method that uses the cause extraction technique.

  • Our system is effective in handling short informal microblog posts.

  • The basic idea and the framework of our approach are language-independent.

Abstract

In recent years, increasing impact of social networks on people’s opinions and decision making has attracted lots of attention. Microblogging, one of the most popular social network applications that allows people to share ideas and discuss over various topics, is taken as a rich resource of opinion and emotion data. In this paper, we propose and implement a novel method for identifying emotions in microblog posts. Unlike traditional approaches which are mostly based on statistical methods, we try to infer and extract the reasons of emotions by importing knowledge and theories from other fields such as Sociology. Based on the theory that a triggering cause event is an integral part of emotion, the technique of emotion cause extraction is used as a crucial step to improve the quality of selected features. First, after thorough analysis on sample data we constructed an automatic rule-based system to detect and extract the cause event of each emotional post. We build an emotion corpus with Chinese microblog posts labeled by human annotators. Then a classifier is trained to classify emotions in microblog posts based on extracted cause events. The overall performance of our system is very promising. The experiment results show that our approach is effective in selecting informative features. Our system outperformed the baseline noticeably in most cases, suggesting its great potential. This exploration should provide a new way to look at the emotion classification task and lay the ground for future research on textual emotion processing.

Introduction

Microblogging is a form of blogging that allows users to publish brief text posts (usually with a strict length limit of not more than 200 characters) (Kaplan & Haenlein, 2011). Microbloggers post about a wide range of topics including daily life, comments on movies or books and opinions on social events. Because of the simplicity and casualness, the number of microbloggers has been growing rapidly in recent years. Since users are able to update their content quickly, microblogging services also act as a hub of real-time news. Organizations such as companies, charity groups and departments of government use microblogging as a tool for marketing and public relations as well. Microblogging services are gradually becoming a platform where information, ideas and opinions converge. Nowadays, many people make their decisions under the influence of the microbloggers they follow. Microblog posts are considered as rich sources of emotion and opinion data (Pak & Paroubek, 2010). It is of great interest to mining user emotions in a microblogging community for the purpose of public opinion tracking, content filtering and customer relationship management (Zhang, Zeng, Li, Wang, & Zuo, 2009).

Emotion processing in text is currently a hot and active area in the field of Natural Language Processing (NLP). Textual emotion detection or classification is one task that many scholars and researchers concentrate on. Though the details may vary, the general goal is the same – to detect and recognize the type of emotion, for example, happiness, anger and surprise, conveyed by the target document (Mihalcea & Liu, 2006). Traditionally, due to the statistical classification nature, the most common practices adopted by researchers are mainly statistics-based models. Feature selection methods like InfoGain and χ2 Test and classifier algorithms like Support Vector Machine (SVM) and k-nearest neighbor (k-NN) are some of the traditional ways for text classification tasks. However, those approaches are very limited in two ways. First, complicated sentences with negation or rhetorical questions cannot be handled well. Second, information of deeper levels, such as the why and how this specific emotion rises are neglected. However, they are very interesting information and sometimes can better reflect the emotion. If we think about how a typical person feels and understands the emotion within a piece of article, we will notice that it is often factors which are not statistically significant, such as the events, the reaction of people, that truly define our perception of emotion. It is intuitive to take such information into consideration when classifying emotions in posts people put on the web.

Among many elements regarding how emotions rise, expressed and perceived by others, the triggering event are often considered one of the most crucial ones. Many scholars in various disciplines have been studying the relation and interplay between emotions and cause events. From a Psychological point of view, there are theories (James, 1884) believe that the cause event itself should be an integral part of emotional experience. In the field of Sociology, Kleres (2011) proposed “narrative analysis” as a methodological approach to systematically analyze emotions by finding out what happens. Lee, Chen, and Huang (2010a) designed a rule-based system to detect emotion causes. Even though those researches do not concentrate on the task of automatic emotion classification, they lay the ground for us.

In this study, we take a fresh approach on classifying textual emotions. We propose a method to classify emotions using the emotion cause extraction technique, based on the combination of cross-disciplinary knowledge and careful investigation on microblog data. We consider it as a novel method because are unaware of any previous emotion classification works using this technique. We focus on identifying emotions based on posts extracted from the website Weibo, the most popular and influential Chinese microblogging community. We take emotion cause events as our entrance point to overcome some of the drawbacks of traditional approaches. Emotions and reactions triggered by the same event are assumed to be similar, so the errors caused by rhetoric can be reduced. Also, deep-level information is taken into account. The experiment results show that the our system can extract emotion cause events from microblog posts with a good accuracy. Based on an efficient cause event extraction, the emotion classification results of our system improves noticeably.

The rest of this paper is structured as follows. Section 2 discussed the related work on emotion analysis including traditional methods and new explorations. Section 3 gives a brief introduction of the Chinese microblogging platform Weibo, and describes the proposed method using emotion cause extraction technique. In Section 4, experiment results of the two stages of our approach are reported and discussed. Section 5 presents the conclusions and our future work.

Section snippets

Related work

We present and briefly introduce related work on the task of emotion processing in this section.

Emotion classification using the emotion cause detection technique

The basic idea behind our approach is looking for features that are “meaningful” to emotions instead of simply choosing words with high co-occurrence degree. Fig. 1 depicts the general framework of our emotion classification method. In the following subsections, we expand on the important parts of our system to explain how we classify emotions in microblog posts with emotion causes.

Experiments and discussion

Our experiment is conducted in a two-stage manner:

  • 1.

    Extracting emotion cause events;

  • 2.

    Training and testing the classifier.

In this section, we first introduce our dataset based on Weibo. Then we describe the two steps and corresponding evaluation measures separately.

Conclusion and future work

Emotion classification has a broad range of applications. An accurate and efficient classification system is of great interest. In this study, we imported the knowledge regarding the relation between emotions and narratives from Sociology and explored the task of textual emotion classification with the technique of emotion cause extraction. We conducted the experiment on a Chinese microblogging platform. First, as many as 16485 potentially emotion-provoking posts were automatically collected.

Acknowledgments

This work is supported by National Natural Science Foundation of China (Grant No.: 61175110) and National Basic Research Program of China (973 Program, Grant No.: 2012CB316305).

References (25)

  • W. James

    Ii. What is an emotion?

    Mind

    (1884)
  • J. Kleres

    Emotions and narrative analysis: A methodological approach

    Journal for the Theory of Social Behaviour

    (2011)
  • Cited by (0)

    View full text