Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 646–651
Brussels, Belgium, October 31 - November 4, 2018.
c
2018 Association for Computational Linguistics
646
Abstract
We present a neural network-based joint
approach for emotion classification and
emotion cause detection, which attempts
to capture mutual benefits across the two
sub-tasks of emotion analysis. Consider-
ing that emotion classification and emo-
tion cause detection need different kinds
of features (affective and event-based sep-
arately), we propose a joint encoder which
uses a unified framework to extract fea-
tures for both sub-tasks and a joint model
trainer which simultaneously learns two
models for the two sub-tasks separately.
Our experiments on Chinese microblogs
show that the joint approach is very prom-
ising.
1 Introduction
The analysis of emotions in texts is an important
task in NLP. Traditional studies treat this task as
a pipeline of two separated sub-tasks: emotion
classification and emotion cause detection. The
former identifies the category of an emotion and
the latter detects the cause of an emotion. This
separated framework makes each sub-task more
flexible to deal with, but it neglects the relevance
between the two sub-tasks. In this paper, we ex-
plore joint approaches which can capture mutual
benefits across the relevant two sub-tasks. To the
best of our knowledge, this work is the first at-
tempt to incorporate both emotion classification
and emotion cause detection into a unified
framework.
Although emotion classification relies on af-
fective features and emotion cause detection
needs event-based features, we propose a joint
encoder which uses a unified framework to ex-
tract features for both emotion classification in-
stances and emotion cause detection instances.
Then, we propose a joint model trainer which
simultaneously learns two models for the two
sub-tasks separately. The experiments on Chinese
microblogs show that our joint approach can ef-
fectively learn models for both sub-tasks.
2 Our Approach
2.1 Corpus
In this paper, we use the human-labeled emotion
corpus provided by Cheng et al. (2017) as our
experimental data (namely Cheng emotion cor-
pus). To better explain our work, we adopt twit-
ter’s terminology used in Cheng et al. (2017).
Cheng emotion corpus can be considered as a
collection of subtweets. For each emotion in a
subtweet, all emotion keywords expressing the
emotion are selected, and then the class and the
cause of the emotion are annotated. The emotion
categorization used in Huang et al. (2016) is
adopted, which includes four basic emotions (i.e.,
joy, angry, sad and fearful) and three complex
emotions (i.e., positive, neutral and negative).
E.g. in the following example, the class of the
emotion keyword (“ ”) is sad, and the cause of
the emotion is “only I was at home again”.
Figure 1: An example of a subtweet
Joint Learning for Emotion Classification and Emotion Cause Detection
Ying Chen
1
, Wenjun Hou
1
, Xiyao Cheng
1
, Shoushan Li
2
1
College of Information and Electrical Engineering, China Agricultural University, China
2
Natural Language Processing Lab, Soochow University, China
{chenying,houwenjun,chengxiyao}@cau.edu.cn
lishoushan@suda.edu.cn
Chinese :
兴冲冲勒跑回家~~发现又是我一个人再
家。。 早知道就去蹭饭了
English Translation: I was very excited to run back
home. I found that only I was at home again. If I
knew it earlier
I would have a meal for free.