Improving Social Emotion Prediction with Reader Comments Integration

Abdullah Alsaedi

, Phillip Brooker

, Floriana Grasso

and Stuart Thomason

Department of Computer Science, University of Liverpool, U.K.

Department of Sociology, Social Policy and Criminology, University of Liverpool, U.K.

Keywords:

Social Emotion Prediction, Emotion Analysis.

Abstract:

Social emotion prediction is concerned with the prediction of the reader’s emotion when exposed to a text.

In this paper, we propose a comment integration method for social emotion prediction. The basic intuition

is that enriching social media posts with related comments can enhance the models’ ability to capture the

conversation context, and hence improve the performance of social emotion prediction. We developed three

models that use the comment integration method with different approaches: word-based, topic-based, and

deep learning-based. Results show that our proposed models outperform popular models in terms of accuracy

and F1-score.

1 INTRODUCTION

In recent years, social media platforms have pro-

vided an open environment for data scientists to ex-

plore and analyze users behavior (Ruths and Pfeffer,

2014). Many studies have been conducted on social

media platforms such as Facebook, Twitter, Reddit,

etc. that aim to identify the sentiment of users toward

a certain event, company, or product (Fan and Gor-

don, 2014). Research into sentiment analysis has pro-

gressed rapidly in recent years, with some remarkable

results. However, traditionally, most of the proposals

made in this ﬁeld use a too coarse-grained approach in

which emotions are classiﬁed as positive, negative, or

neutral. Some recent research has emerged attempt-

ing to integrate this with emotion analysis, to obtain a

deeper understanding of the text based on predeﬁned

emotion models (Lerner and Keltner, 2000).

Unlike sentiment analysis, the analysis of emotion

can be conducted for both subjective and objective

text (Rao et al., 2012). Emotion analysis also deals

with text from different perspectives: the writer’s per-

spective, which represents the expressed emotion in

the text; and the reader’s perspective, which shows

the emotions the text provokes in its readers (Lin

et al., 2007). However, most existing work focuses

on identifying the emotion from the writer’s perspec-

tive (Guan et al., 2019).

Social emotion prediction is the consensus term to

designate work on reader’s emotion prediction, that is

research concerned with the problem of predicting the

emotion provoked to the reader after being exposed to

the text. Predicting the social emotion is a challeng-

ing task as the reader’s emotion is not declared in the

text, but it is triggered by reading the text, and it is

likely inﬂuenced by many factors, such as the read-

ers’ personal background and experiences.

One common way to approach the problem is to

understand whether there is a relationship between

reader’s and writer’s emotion, for instance by estab-

lishing if the reader experiences the same emotion

that the writer is portraying in the text. This can be

done by considering, as a single unit, a piece of text

and any comments to that text. For example, in social

media, one would consider posts followed by their re-

lated comments. The assumption is that the readers

who wrote comments can be assumed to have been

affected by the content of the post. Consequently, the

written emotions in the comments could reﬂect the

readers’ emotion at reading the text. Previous stud-

ies show indeed a correlation between the emotions

of the reader and writer of a comment (Yang et al.,

2009; Liu et al., 2013).

Such comments integration approach was found

to be effective for short-text topic modeling (Alvarez-

Melis and Saveski, 2016) based on underlying as-

sumptions on topic consistency between posts and

comments. As the social emotion can be related to

the topic of the post, the assumption is that the con-

tent of the comments could be informative and useful

for social emotion prediction models to learn from,

even without knowing the emotions experienced by

Alsaedi, A., Brooker, P., Grasso, F. and Thomason, S.

Improving Social Emotion Prediction with Reader Comments Integration.

DOI: 10.5220/0010837000003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 2, pages 285-292

ISBN: 978-989-758-547-0; ISSN: 2184-433X

285

the writers of the comments.

In this paper we propose three models for social

emotion prediction using the comment integration ap-

proach. We integrate each post with its corresponding

comments as a whole document, then train our mod-

els on these integrated documents of posts and com-

ments. We report on experiments that show that, even

with a basic architecture, by integrating the comments

we can outperform established social emotion predic-

tion models.

The rest of the paper is structured as follows. Sec-

tion 2 explores related works. Section 3 presents our

comments integration models. Section 4 discusses the

experiments and analysis, and Section 5 presents the

conclusion of our work.

2 RELATED WORK

In this section, we summarize related work in two

areas: social emotion prediction and bi-perspective

emotion analysis.

2.1 Social Emotion Prediction

A shared task on Affective Text, proposed at the

2007 SemEval workshop (Strapparava and Mihalcea,

2007), served as starting point for the development

of several systems for social emotion prediction us-

ing various approaches, such as word-based methods,

topic-based methods, and deep learning-based meth-

ods (Alsaedi et al., 2021). In this section we sum-

marise the more relevant ones to the scope of this pa-

per.

Word-based systems assume that all words, even

neutral ones, can be associated with a likelihood to

provoke emotions (Strapparava and Mihalcea, 2008).

SWAT (Katz et al., 2007), one of the top-performing

word-based systems, uses a bag-of-words (BOW)

model trained to label news headlines with social

emotions. It utilizes Roget’s Thesaurus for synonym

and antonym expansion and scores each word to de-

termine the average score for the headline. Another

interesting approach is the one presented in (Chau-

martin, 2007). The authors propose UPAR7, a rule-

based system, depends on a syntactic parser that

utilises resources from WordNet, SentiWordNet, and

WordNet-Affect for social emotion prediction.

Topic-based systems suggest that social emotion

should be linked to topics rather than words (Bao

et al., 2009). Therefore, topic-based systems such as

Emotion-Topic Model (ETM)(Bao et al., 2009)(Bao

et al., 2011), Affective Topic Model (ATM) (Rao

et al., 2014b), Sentiment Latent Topic Model (SLTM)

(Rao et al., 2014a), etc. beneﬁt from the machin-

ery of topic modeling techniques and introduce an

intermediate emotion layer to popular topic mod-

els such as Latent Dirichlet Allocation (LDA) (Blei

et al., 2003). Similarly, Multi-label Supervised Topic

Model (MSTM) (Rao et al., 2014a) utilizes the Super-

vised Topic Model (STM) (Blei and McAuliffe, 2010)

that extended to model multi-labels rather than single

label to be suitable for social emotion prediction.

State-of-the-art systems usually depend on deep

learning-based methods, such as neural networks and

word-embeddings, to obtain better understanding of

sequences and overcome the data sparsity problem

that occurs with other approaches. For example, in

(Guan et al., 2019), a hierarchical model based on

long short-term memory (LSTM) was proposed. The

model utilizes the attention mechanism and attempts

to capture the semantics of long texts using three dif-

ferent levels of embeddings: words, sentences and

documents. Similarly, TESAN (Wang and Wang,

2020) combines semantic and topical features and

feeds them into a uniﬁed deep learning model. The

model consists of a neural topic model that learns

the topical embeddings of documents, and a topic-

enhanced self-attention mechanism to generate the

document vector from the semantic and topical fea-

tures. Both features are integrated in a ﬁnal gate,

achieving an improvement in performance.

2.2 Bi-perspective Emotion Analysis

Another set of works which are relevant to our re-

search focus on the analysis of reader’s emotion when

compared and contrasted to the writer’s emotion.

In (Yang et al., 2009) the readers’ emotion from

Yahoo!Kimo news, the Taiwan branch of Yahoo!, was

analyzed in comparison to the writers’ emotion from

the Yahoo!Kimo blog corpus. The authors built a

reader’s emotion classiﬁer trained on the news corpus,

and applied it to the blog corpus, then analyzed the

new corpus, annotated with both perspectives. They

found that the valence, which is the degree of pleas-

antness, reveals that in blog topics, readers’ and writ-

ers’ emotions tend to agree on their polarities. How-

ever, the degree of inﬂuence is affected by the topic.

Tang and Chen (Tang and Chen, 2011) used

data from the micro-blogging platform Plurk

(https://www.plurk.com/portal/) that provides emo-

tion tagging from both perspectives for posts and

their corresponding comments. Users in Plurk have

the ability to tag their own posts with emotion, hence

self reporting the writer’s emotion. Users replying

to a post can also tagged their own replied, hence

self reporting the reader’s emotion. When analysing

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

286

both perspectives, they conclude that predicting the

reader’s emotion is clearly more challenging than

predicting the writer’s emotion. The authors extended

their work (Tang and Chen, 2012) by studying the

emotion transition between writer and reader in the

Plurk platform, and analyzed the linguistic features

that signal the change of emotions between the two.

They also proposed models to predict the transition

and they suggested sentiment word mining as a useful

tool for prediction performance.

In (Liu et al., 2013), the authors studied the rela-

tionship between news and their comments, again on

thee Yahoo!Kimo News. As the reader’s emotion is

provided by Yahoo!Kimo News, they manually anno-

tated the comments with the writer’s emotion. When

comparing, the news reader’s emotion and the com-

ments writer’s emotion, they found that both sets of

emotions are strongly correlated in term of valence,

but interestingly, not in ﬁne-grained emotions such as

happiness, anger, sadness, etc.

In (Buechel and Hahn, 2017), the authors exam-

ined the data annotation process for writer’s emotion

and reader’s emotion, and they interestingly propose a

third perspective in the process of analysis, that is the

emotion from the point of view ”of the text itself”.

They conducted experiments on two English corpora

and suggest that the quality of the writer’s emotion an-

notation was the highest. The work is notable and it

is accompanied by a dataset, EmoBank, a large-scale

bi-perspective emotion annotated corpus of 10k sen-

tences, covering different genres.

Another important dataset is GoodNewsEveryone

(Bostan et al., 2019), a corpus of English news head-

lines annotated with writer’s emotion, writer’s emo-

tion intensity, reader’s emotion and semantic roles ac-

cording to the FrameNet (Baker et al., 1998) semantic

frame. The dataset consists of 5000 English headlines

annotated via crowdsourcing.

3 COMMENT INTEGRATION

MODELS

In this section we present our models for social emo-

tion prediction. We use for all models the comment

integration approach, and we merge texts with the

comments attached to that text in one single docu-

ment. We conducted a number of experiments by

varying different features, and we report the combina-

tion of features producing the best results for each of

the three basic approaches: word-based, topic-based

and deep learning based.

3.1 Comments Integration Word-based

Model

Word-based methods aim to detect the emotion in

content through ﬁnding the reader’s emotional trigger,

as it is noticed that any words, even the neutral ones,

can provoke emotions. Our simplest model combine

posts and their comments and used a basic Na

ıve-

Bayes classiﬁer trained on the BOW representation of

the documents resulting from the combination. A so-

cial emotion for the post is then attached to the whole

document.

Figure 1 shows the architecture of the proposed

word-based model.

Figure 1: Architecture of the comments integration word-

based model.

Results will be discussed in Section 4, but we

anticipate here that, compared to the Na

ıve-Bayes

model trained on posts content only, our model im-

proved the performance in terms of accuracy, preci-

sion and recall. The model also achieved a signiﬁcant

improvement compared to other popular models, in-

cluding topic-based models.

3.2 Comments Integration Topic-based

Model

In this model we used the merged document of post

and related comments as an input to an LDA model to

extract the topics distribution for each document. We

used EmoLex (Mohammad and Turney, 2013), which

is an emotion lexicon annotated with classes from

Plutchick’s emotion model (Plutchik, 1980). EmoLex

was used to extract the emotion-word frequency in

each document in order to produce a vector of emo-

tion frequency. Both the LDA output and the emotion

frequency vector were then used as an input to a max-

imum entropy classiﬁer. Figure 2 shows the architec-

ture of the proposed topic-based model.

Improving Social Emotion Prediction with Reader Comments Integration

287

Figure 2: Architecture of the comments integration topic-

based model.

3.3 Comments Integration Deep

Learning-based Model

The deep learning model utilizes a Term Fre-

quency–Inverse Document Frequency (TF-

IDF)(Salton and Buckley, 1988) representation

that gives more importance to words that are more

relevant to the document. We found that the use of

a TF-IDF representation of the merged document

of posts and comments improved the performance

signiﬁcantly. The model consists of a multi-layer

neural network with one hidden layer and a softmax

activation in the output layer. Figure 3 shows the

architecture of the proposed deep learning-based

model.

4 EXPERIMENTS

4.1 Dataset

The evaluation was conducted using FacebookR

(Krebs et al., 2018), a dataset of Facebook posts with

their social emotion and comments. Facebook intro-

duced the ”reactions” feature in 2016 to enable its

users to express their emotion toward posts, represent-

ing their social emotion. The reactions set includes

Like, Love, Care, HAHA, Wow, Sad and Angry.

FacebookR is scraped from customer service

pages of popular supermarkets from the United States

and the United Kingdom. To the best of our knowl-

Figure 3: Architecture of the comments integration deep

learning-based model.

edge, this is the only available English dataset that

provides comments as well as posts with the social

emotion labels. The dataset consists of over 70,000

posts. Of these, we considered those that excluded the

”like” reaction. As pointed out in (Krebs et al., 2018),

the like reaction makes the dataset very imbalanced

and inaccurate, as Facebook users tend to use ”like”

for any positive emotion or, sometimes, to show that

they have read the post. The resulting dataset when

excluding the ”like” reaction consists of 8103 posts.

As manual labeling for social emotion is challeng-

ing and usually produces low quality labels, Face-

bookR is scraped from real data that is annotated by

Facebook users themselves. Also, based on statistical

analysis conducted by the dataset publishers, user la-

beling seems consistent and the reactions that appear

together tend to agree in a high degree. We use the

dataset to predict the top social emotion for posts. Ta-

ble 1 shows the number of posts with the reaction of

the top number of votes in the dataset.

Table 1: The number of posts with their top social emotion.

Top reaction Number of posts

ANGRY 2276

HAHA 2253

LOVE 1648

WOW 1237

SAD 686

There are on average six comments for each post

and among 8103 posts only 150 are without any com-

ment. Table 2 shows the comments statistics for Face-

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

288

bookR and Table 3 shows the documents length be-

fore and after merging the posts with their comments.

Table 2: Comments statistics for FacebookR posts.

Measure Number of comments

mean 6.31

std 37.56

min 1

max 2605

Table 3: Number of words in posts versus posts merged with

their comments.

Number of words Posts

Posts merged

with comments

mean 111.66 274.62

std 125.23 624.15

min 1 2

max 2093 38004

4.2 Baselines

We compare our models with the following popular

models from the literature:

• SVM: Support vector machine classiﬁer.

• NB: Naive Bayes classiﬁer.

• SWAT: One of the best performing models pro-

posed in SemEval-2007 task 14, which depends

on scoring the document words.

• ET (Bao et al., 2011): A straightforward model

that depends on a Naive Bayes classiﬁer to learn

the word-emotion associations.

• ETM: The ﬁrst topic-based model that introduced

an emotion layer to LDA for social emotion pre-

diction.

• ATM: A topic-based model that generates social

emotion lexicon and predicts unlabelled docu-

ments.

• MSTM: A supervised topic model that learns the

association of word and topic, then predicts the

emotions from each topic.

• SLTM: A topic-based model that associates words

and emotions to topics to predict the social emo-

tion from unseen documents.

In addition, we perform a comparison with the fol-

lowing deep learning models:

• LSTM (Hochreiter and Schmidhuber, 1997): A

long short-term memory model, which is a popu-

lar type of recurrent neural network, usually used

for learning from ordered sequences such as sen-

tences.

• AttBiLSTM (Wang and Yang, 2020): A widely-

used bidirectional long short-term memory net-

work with an attention mechanism.

4.3 Experimental Design

We arbitrarily set the number of topics K to 20 for

the following models: ETM, ATM, MSTM, SLTM

and our topic-based model. For LSTM, AttBiLSTM,

and our deep learning model we used the Adam op-

timizer (Kingma and Ba, 2015) for training with a

learning rate of 0.001, batch size 100, 20 epochs, and

early stopping on validation loss to prevent overﬁtting

(SJ

OBERG and LJUNG, 1995). Both LSTM and At-

tBiLSTM consist of one layer with 32 units, and use

a pre-trained glove (Pennington et al., 2014) word-

embedding with 200-dimensions trained on a Twitter

dataset as the language on Twitter and Facebook share

similar characteristics.

4.4 Evaluation Metrics

We use Acc@1 as a metric for model evaluation in

our experiments, as it has been proved to represent

the correct prediction for the social emotion with the

highest number of ratings and is considered the most

important metric according to (Bao et al., 2009; Bao

et al., 2011).

To determine Acc@1, the correctly predicted doc-

uments are counted against the whole number of doc-

uments as follows:

Acc

(

1 if e

= EM

top

0 otherwise

Acc@1 =

∑

d∈D

Acc

|D|

Put simply, if there are many reactions for a post,

we will consider the reaction with the top ratings as

the social emotion of that post. If two or more reac-

tions have the same number of ratings we will con-

sider either of them.

4.5 Prediction Performance

Here we evaluate our models’ ability to predict each

social emotion by measuring the precision, recall, and

F1-score. Table 4 shows the evaluation for our word-

based model, topic-based model, and deep learning-

based model.

We can notice that the models performances vary

depending on the predicted emotion. As the dataset

is unbalanced, the prediction for some emotions is af-

fected negatively, such as the SAD and WOW emo-

tions, especially for the topic-based and the deep

Improving Social Emotion Prediction with Reader Comments Integration

289

Table 4: Evaluation of the proposed models.

Model Emotion Precision Recall F1

Word-

based

ANGRY 0.47 0.83 0.60

HAHA 0.56 0.55 0.56

LOVE 0.74 0.60 0.66

SAD 0.15 0.02 0.03

WOW 0.32 0.07 0.11

Topic-

based

ANGRY 0.45 0.67 0.54

HAHA 0.45 0.62 0.52

LOVE 0.59 0.56 0.58

SAD 0.00 0.00 0.00

WOW 0.19 0.01 0.02

Deep

Learning-

based

ANGRY 0.50 0.77 0.61

HAHA 0.51 0.75 0.61

LOVE 0.82 0.61 0.70

SAD 0.00 0.00 0.00

WOW 0.33 0.00 0.01

learning-based models. Interestingly, the F1 for the

LOVE emotion was always the highest, even though

the number of training samples for the ANGRY and

HAHA are higher which needs further investigation.

4.6 Comparison with Baselines

In this subsection we compare our social emotion pre-

diction models to the baselines mentioned in 4.2. All

models were evaluated on the FacebookR dataset and

two evaluation metrics have been used in the compari-

son: Accuracy of the top social emotion (Acc@1), and

F1-score. We focus on comparing models with sim-

ilar methods and divide the models into word-based

models, topic-based models, and deep learning-based

models. Table 5 summarises the performance of all

models.

The table shows that our proposed models per-

form best among models with the same method in

term of accuracy, and also in F1 except for the topic-

based models, where ETM is better and achieved a

comparable accuracy to our topic-based model. Our

deep learning model outperforms all of other models

in terms of accuracy and F1. The word-based model

outperforms SWAT and ET both in accuracy and F1.

However, the ET model performs much better than

SWAT, which has a very low F1. Interestingly, the

performance of our word-based model was better than

all of the topic-based models, including our model,

and that might be related to the topics distribution the

extent of overlapping topics in our dataset.

When comparing the topic-based models we can

see that their performances vary signiﬁcantly. ATM,

SLTM, and MSTM accuracies are low compared to

ETM and our model. However, with regard to F1

score, the differences between them are much bigger

since ATM, SLTM, and MSTM have very low F1.

Deep learning models provide a relatively close

accuracy for F1. There are no signiﬁcant differences

between the accuracy and F1 score, which sometimes

appears in some of the word-based and topic-based

models. However, our deep learning-based model

boosts the accuracy by 10% compared to the LSTM

model. The F1 score also increased by 0.08. The

AttBiLSTM model improved the performance by 2%

over the LSTM model the F1 score raised by 0.02.

All of the deep learning models were learning

quickly before suffering from overﬁtting after the ﬁrst

few epochs. We tried to avoid that by applying the

early stopping technique to force the training to stop

when there was no improvement. Figure 4 presents

the accuracy of all deep learning models over 20

epochs. It is noteworthy that the models’ accuracy

began to reduce after the ﬁrst few epochs, which is a

sign of overﬁtting.

Figure 4: Accuracy of deep-based models with different

number of epochs.

5 CONCLUSION

In this paper, we have proposed a comment integra-

tion method for social emotion prediction. We de-

veloped word-based, topic-based, and deep learning-

based models that use our proposed methods and

compared them to popular social emotion prediction

methods. Experiments show that models that use

the comments integration method outperform popu-

lar models in terms of Acc@1 and F1. We found

that merging social media comments with their re-

lated posts added valuable data about the readers’

emotions and enhanced the ability to predict the social

emotion for posts. In the future, we will utilize the

comment integration methods to improve the models

from two aspects. On the one hand, we will develop

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

290

Table 5: A summary of models’ performances in terms of accuracy and F1.

Method Model Accuracy F1

Word-based

SVM 48.41 0.44

NB 50.08 0.46

SWAT 18.73 0.06

ET 43.41 0.34

Our Word-based model 52.59 0.48

Topic-based

ETM 46.21 0.44

ATM 21.16 0.06

SLTM 27.43 0.22

MSTM 19.40 0.06

Our Topic-based model 48.12 0.42

Deep Learning-based

LSTM 45.34 0.41

AttBiLSTM 47.22 0.43

Our Deep learning-based model 55.55 0.49

more advanced models that depend on deep learning-

based methods and neural topic models. On the other

hand, we will attempt to improve the prediction per-

formance for emotions with low precision and recall.

REFERENCES

Alsaedi, A., Brooker, P., Grasso, F., and Thomason, S.

(2021). A survey of social emotion prediction meth-

ods. In Proceedings of the 10th International Confer-

ence on Data Science, Technology and Applications.

SCITEPRESS-Science and Technology Publications.

Alvarez-Melis, D. and Saveski, M. (2016). Topic Modeling

in Twitter: Aggregating Tweets by Conversations. In

Proceedings of the Tenth International AAAI Confer-

ence onWeb and Social Media.

Baker, C. F., Fillmore, C. J., and Lowe, J. B. (1998). The

berkeley framenet project. In 36th Annual Meeting

of the Association for Computational Linguistics and

17th International Conference on Computational Lin-

guistics, Volume 1, pages 86–90.

Bao, S., Xu, S., Zhang, L., Yan, R., Su, Z., Han, D., and

Yu, Y. (2009). Joint emotion-topic modeling for so-

cial affective text mining. In Proceedings - IEEE In-

ternational Conference on Data Mining, ICDM, pages

699–704. IEEE.

Bao, S., Xu, S., Zhang, L., Yan, R., Su, Z., Han, D., and

Yu, Y. (2011). Mining Social Emotions from Affec-

tive Text. IEEE Transactions on Knowledge and Data

Engineering, 24(9):1658–1670.

Blei, D. M. and McAuliffe, J. D. (2010). Supervised topic

models. arXiv preprint arXiv:1003.0783.

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent

dirichlet allocation. Journal of machine Learning re-

search, 3(Jan):993–1022.

Bostan, L., Kim, E., and Klinger, R. (2019). GoodNew-

sEveryone: A Corpus of News Headlines Annotated

with Emotions, Semantic Roles, and Reader Percep-

tion. arXiv:1912.03184v3 [cs.CL].

Buechel, S. and Hahn, U. (2017). EMOBANK: Studying

the impact of annotation perspective and representa-

tion format on dimensional emotion analysis. In 15th

Conference of the European Chapter of the Associa-

tion for Computational Linguistics, EACL 2017, vol-

ume 2, pages 578–585.

Chaumartin, F. R. (2007). UPAR7: A knowledge-based

system for headline sentiment tagging. In Agirre, E.,

arquez, L., and Wicentowski, R., editors, Proceed-

ings of the 4th International Workshop on Semantic

Evaluations (SemEval-2007), pages 422–425. ACL.

Fan, W. and Gordon, M. D. (2014). The power of social me-

dia analytics. Communications of the ACM, 57(6):74–

81.

Guan, X., Peng, Q., Li, X., and Zhu, Z. (2019). Social

Emotion Prediction with Attention-based Hierarchical

Neural Network. In 2019 IEEE 4th Advanced Infor-

mation Technology, Electronic and Automation Con-

trol Conference (IAEAC), pages 1001–1005.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural Comput., 9(8):1735–1780.

Katz, P., Singleton, M., and Wicentowski, R. (2007).

SWAT-MP: The SemEval-2007 systems for task 5 and

task 14. In ACL 2007 - SemEval 2007 - Proceedings

of the 4th International Workshop on Semantic Evalu-

ations, pages 308–313.

Kingma, D. P. and Ba, J. (2015). Adam: A Method for

Stochastic Optimization. In Bengio, Y. and LeCun,

Y., editors, 3rd International Conference on Learn-

ing Representations, ICLR 2015, San Diego, CA, USA,

May 7-9, 2015, Conference Track Proceedings.

Krebs, F., Lubascher, B., Moers, T., Schaap, P., and

Spanakis, G. (2018). Social emotion mining

techniques for facebook posts reaction prediction.

ICAART 2018 - Proceedings of the 10th Interna-

tional Conference on Agents and Artiﬁcial Intelli-

gence, 2:211–220.

Lerner, J. S. and Keltner, D. (2000). Beyond valence: To-

ward a model of emotion-speciﬁc inﬂuences on judge-

ment and choice. Cognition & emotion, 14(4):473–

493.

Improving Social Emotion Prediction with Reader Comments Integration

291

Lin, K. H.-Y., Yang, C., and Chen, H.-H. (2007). What emo-

tions do news articles trigger in their readers? In Pro-

ceedings of the 30th annual international ACM SIGIR

conference on Research and development in informa-

tion retrieval, pages 733–734.

Liu, H., Li, S., Zhou, G., Huang, C. R., and Li, P. (2013).

Joint modeling of news reader’s and comment writer’s

emotions. In ACL 2013 - 51st Annual Meeting of the

Association for Computational Linguistics, Proceed-

ings of the Conference, volume 2, pages 511–515.

Mohammad, S. M. and Turney, P. D. (2013). Crowdsourc-

ing a word-emotion association lexicon. Computa-

tional Intelligence, 29(3):436–465.

Pennington, J., Socher, R., and Manning, C. D. (2014).

Glove: Global vectors for word representation. In

Proceedings of the 2014 conference on empirical

methods in natural language processing (EMNLP),

pages 1532–1543.

Plutchik, R. (1980). Psychoevolutionary Theory of Basic

Emotions. American Scientist.

Rao, Y., Li, Q., Mao, X., and Wenyin, L. (2014a). Sentiment

topic models for social emotion mining. Information

Sciences, 266:90–100.

Rao, Y., Li, Q., Wenyin, L., Wu, Q., and Quan, X. (2014b).

Affective topic model for social emotion detection.

Neural Networks, 58:29–37.

Rao, Y., Quan, X., Wenyin, L., Li, Q., and Chen, M. (2012).

Buildingword-emotion mapping dictionary for online

news. In SDAD@ ECML/PKDD, pages 28–39.

Ruths, D. and Pfeffer, J. (2014). Social media for large

studies of behavior. Science, 346(6213):1063–1064.

Salton, G. and Buckley, C. (1988). Term-weighting ap-

proaches in automatic text retrieval. Information pro-

cessing & management, 24(5):513–523.

OBERG, J. and LJUNG, L. (1995). Overtraining, reg-

ularization and searching for a minimum, with appli-

cation to neural networks. International Journal of

Control, 62(6):1391–1407.

Strapparava, C. and Mihalcea, R. (2007). SemEval-2007

task 14: Affective text. In ACL 2007 - SemEval 2007

- Proceedings of the 4th International Workshop on

Semantic Evaluations.

Strapparava, C. and Mihalcea, R. (2008). Learning to iden-

tify emotions in text. In Proceedings of the ACM Sym-

posium on Applied Computing, pages 1556–1560.

Tang, Y.-j. and Chen, H.-h. (2011). Emotion Modeling

from Writer / Reader Perspectives Using a Microblog

Dataset. In Proceedings of the Workshop on Sentiment

Analysis where AI meets Psychology (SAAIP), pages

11–19.

Tang, Y. J. and Chen, H. H. (2012). Mining sentiment words

from microblogs for predicting writer-reader emotion

transition. In Proceedings of the 8th International

Conference on Language Resources and Evaluation,

LREC 2012, pages 1226–1229.

Wang, C. and Wang, B. (2020). An End-to-end Topic-

Enhanced Self-Attention Network for Social Emotion

Classiﬁcation. In The Web Conference 2020 - Pro-

ceedings of the World Wide Web Conference, WWW

2020, volume 2, pages 2210–2219.

Wang, Z. and Yang, B. (2020). Attention-based bidi-

rectional long short-term memory networks for

relation classiﬁcation using knowledge distillation

from bert. In 2020 IEEE Intl Conf on De-

pendable, Autonomic and Secure Computing, Intl

Conf on Pervasive Intelligence and Computing, Intl

Conf on Cloud and Big Data Computing, Intl

Conf on Cyber Science and Technology Congress

(DASC/PiCom/CBDCom/CyberSciTech), pages 562–

568.

Yang, C., Lin, K. H. Y., and Chen, H. H. (2009). Writer

meets reader: Emotion analysis of social media from

both the writer’s and reader’s perspectives. In Pro-

ceedings - 2009 IEEE/WIC/ACM International Con-

ference on Web Intelligence, WI 2009, volume 1,

pages 287–290.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

292