C. Zhu et al.
Keywords Social emotions · Topic models · Sentiment analysis
1 Introduction
With the increasing prosperity of Web 2.0, people are encouraged to have various social
interactions on the Web sites. A recent development trend of online news Web sites, such
as Yahoo! and Sina, is to allow readers to specify different types of emotions (e.g., angry
and shocked) after reading news. Compared with traditional users’ feedback (e.g., reviews,
tags), such specific emotion annotations are more accurate for expressing users’ personal
emotions. For example, Fig. 1a shows an example of users’ aggregated emotions at Sina
News,
1
which has six different kinds of emotions. Each user can choose one emotion, which
most accurately reflects his impression after reading, to annotate a piece of news. If we
collectively look at and analyze all user emotions from the news Web site, we may be able
to get a good picture about the overall emotions of online social media users, namely social
emotions. The social emotions often vary with respect to the topics of news and time and thus
have intrinsic dynamics. For example, Fig. 1b shows the distribution of aggregated social
emotions with respect to different time spans in our data set c ollected from Sina News. We
can observe different distributions of emotions along the time, which indicates that the social
emotions evolve over time. In fact, such evolution of social emotions is inherently driven by
the dynamic topics of news at different time. Capturing such dynamic characteristics of social
emotions is critically important for the successful development of various social services,
such as social opinion monitoring and social event detections. In the literature, there are recent
studies about social emotion-related problems. For example, some works focus on sentiment
analysis [24,35], social emotion analysis [6,13,18] of online documents and user emotion
modeling [3]. However, few of them have paid attention to the dynamic characteristics of
social emotions.
To this end, in this paper, we propose to exploit the users’ emotion annotations from
online news to track the evolution of social emotions. A critical challenge is how to model
emotions with respect to time spans. Along this line, we propose a time-aware topic modeling
perspective for solving this problem. Specifically, we first develop a model named emotion-
Topic over Time (eToT), where we represent each topic of news with a beta distribution over
time and a multinomial distribution over emotions. In this model, the process of modeling
topics of news is affected not only b y the word co-occurrences but also the emotions and time.
Particularly, with some experimental observations, we find that not all the topics have distinct
relationships with time, such as the common topics or the topics about news terminology.
Therefore, we further propose an extension of eToT model, named mixed emotion-Topic over
Time (meToT), in which a word can be generated from two different topic sets, i.e., one is time-
independent and the other is time-dependent. Indeed, similar to the popular Latent Dirichlet
Allocation (LDA) [5], news topics in both eToT and meToT are static which do not change
with respect to different time spans. However, while some researchers have revealed that the
latent topics of documents may evolve as time unfolds [4], eToT and meToT cannot capture the
dynamics of topics. Therefore, we propose another model, namely emotion-based Dynamic
Topic Model (eDTM), to capture the dynamics of news topics. In eDTM, we first divide all
news into different segmentations with respect to their timestamps and then implement topic
modeling for each segmentation. Models learned from different segmentations are linked
together by the Markov state space model. In fact, all eToT, meToT, and eDTM can model
1
http://news.sina.com.cn/.
123