See discussions, stats, and author profiles for this publication at: [Link]

net/publication/324634143

INF-UFG at FiQA 2018 Task 1: Predicting Sentiments and Aspects on Financial

Tweets and News Headlines

Conference Paper · April 2018

DOI: 10.1145/3184558.3191828

CITATIONS READS

19 1,008

2 authors:

Dayan Costa Nádia Félix Felipe Da Silva

OLX Group Federal University of Goiás
2 PUBLICATIONS 23 CITATIONS 81 PUBLICATIONS 1,019 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Nádia Félix Felipe Da Silva on 05 May 2018.

The user has requested enhancement of the downloaded file.

 INF-UFG at FiQA 2018 Task 1: Predicting Sentiments and
Aspects on Financial Tweets and News Headlines

Dayan de Franca Costa Nadia Felix Felipe da Silva

Federal University of Goias Federal University of Goias
dayancosta@[Link] nadia@[Link]

ABSTRACT Since researches has been showing that opinions expressed

This paper describes our system which participate in Task 1 in microblogs and social media discussions can make a great
of FiQA 2018. The task’s focuses was to predict sentiment impact on market[7].
and aspects of financial microblog posts and headlines. The This paper describes our systems submitted to Open Chal-
sentiment analysis for a specific company had to be predicted lenge – Financial Opinion Mining in FiQA(2018) - Task 1
using a scale between -1 and 1, while the aspect prediction which is related to aspect-based financial sentiment anal-
had to be predicted using a set of aspects which was given in ysis 3 . This task aims to identify positives(bearish) and
train data. We had used Support Vector Regression (SVR) negatives(bullish) sentiments and the aspects associated to
to predict the sentiments in both cases (microblog posts and companies stocks based on a dataset of microblogging posts
headlines). and news statements and headlines. The challenges was
that the sentiment should was scaled between -1(very neg-
ative/bullish) and 1(very positive/bearish) and don’t using
Keywords the conventional positive, negative and neutral labels, as a
Predictive sentiment analysis; Stock market; Sentiment anal- classifying challenge, in addition the aspects should be clas-
ysis in the financial domain sified using a big set of aspects given in the dataset, using
two levels of classification.
1. INTRODUCTION In this work, we extract a series of elaborately designed
features like word embeddings, n-grams, word replacements
The use and communication through microblogs as Twit-
and multiple di↵erent regression and classification algorithms
ter1 and StockTwits2 have been increasing in the last years,
such as SVM in linear kernels, Bayesian models, ensemble
making the market to pay more attention in what is said
and tree models. We also have tried a ensemble containing
about them in those microblogs [1]. This increase is given
some of these models quoted.
through the facility to use these platforms, the message for-
The rest of this paper is structured as follows. Section 2
mats and its accessibility[19]. People are using these com-
describes related work. Section 3 reports datasets, experi-
munication vehicles to express sentiments about life, busi-
ments and results discussion. Finally, Section 4 concludes
ness, work, sports and other ones, therefore the brands has
our work.
become more worried about what people say about them in
these platforms to measure how people fell about the brand,
attract more customers, etc. In financial domain, these texts 2. RELATED WORK
full of opinions are able to change the financial sector, raising Much research has focused on whether the sentiment anal-
or falling stock values. ysis of social media can be used to predict the future of stock
With the increasing use of these platforms, applications of market indicators. In this section, we give an overview of
technologies have been growing up too in last decade and it related studies, which are focused on: (i) sentiment analy-
has been caused by the applications coming from academia sis of social media as a predictor of the future stock market
to commercial domain and because sentiment the analysis indicators, and (ii) aspect-based sentiment analysis.
tasks are full of challenges. In last years, sentiment analysis
become an interest of financial research area [2, 30, 31, 4]. 2.1 Financial tweet sentiment analysis
1
The most well-known publication in this area is by Bollen
[Link] [2]. They investigated whether measurements of collective
2
[Link] moods derived from Twitter feeds are correlated to the value
of the Dow Jones Industrial Average (DJIA) over time. They
Permission to make digital or hard copies of all or part of this work for personal or analyze the tweets by two mood tracking tools, namely Opin-
classroom use is granted without fee provided that copies are not made or distributed ionFinder that measures positive vs. negative mood and
for profit or commercial advantage and that copies bear this notice and the full cita-
tion on the first page. Copyrights for components of this work owned by others than Google-Profile of Mood States (GPOMS) that measures mood
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- in terms of 6 dimensions (Calm, Alert, Sure, Vital, Kind,
publish, to post on servers or to redistribute to lists, requires prior specific permission and Happy). They cross-validated the resulting mood-time
and/or a fee. Request permissions from permissions@[Link].
series by comparing their ability to detect the public’s re-
FIQA - Challenge@WWW 2018
sponse to the presidential election and Thanksgiving day
c 2018 ACM. ISBN 123-4567-24-567/08/06. . . $15.00
3
DOI: 10.475/123 4 [Link]
in 2008. The Granger causality analysis and a Fuzzy Neu- text instance adopted floating point values in the range of
ral Network were used to investigate the hypothesis that -1 (very negative/bearish) to 1 (very positive/bullish), with
public mood states, as measured by the OpinionFinder and 0 designating neutral sentiment. This task attracted a total
GPOMS mood time series, were predictive of changes in of 32 participants, with 25 participating in Track 1 and 29
DJIA closing values. Their results indicated that the accu- in Track 2. Tables 1 and 2 summarize the top five teams
racy of DJIA predictions can be improved by the inclusion that participated in task.
of specific public mood dimensions.
Smailovic et al. [30] also analyzed whether the sentiment
expressed in tweets, which discuss selected companies and Table 1: Top five teams that participated in Task
their products, can indicate their stock price changes. To 5 of Semeval – predicting sentiment from financial
address this problem, an active learning approach was de- news headlines
Authors Feature Set Approach
veloped and applied to sentiment analysis of tweet streams 1 Mansar et al. [15] word embeddings A deep learning ar-
in the stock market domain. The paper first presents a static (GloVe), lexicon chitecture
Twitter data analysis problem, explored in order to deter- 2 Kar et al. [11] word embeddings Support Vector Re-
mine the best Twitter-specific text preprocessing setting for (word2vec), lexicon gression (SVR) and a
training the Support Vector Machine (SVM) sentiment clas- deep learning archi-
tecture
sifier. In the static setting, the Granger causality test shows 3 Rotim et al. [25] String kernels, word Support Vector Re-
that sentiments in stock-related tweets can be used as in- embeddings (Glove gression (SVR)
dicators of stock price movements a few days in advance, and word2vec),
where improved results were achieved by adapting the SVM named entity recog-
classifier to categorize Twitter posts into three sentiment nition (NER)
categories of positive, negative and neutral. These findings 4 Moore and n-grams and word- Support Vector Re-
Rayson [17] replacements gression (SVR) and
were adopted in the development of a new stream-based ac- a Bidirectional Long
tive learning approach to sentiment analysis, applicable in Short-Term Memory
incremental learning from continuously changing financial (BLSTM)
tweet streams. To this end, a series of experiments was con- 5 Jiang et al. [10] linguistic, sentiment regressor ensemble
ducted to determine the best querying strategy for active lexicon, domain-
learning of the SVM classifier adapted to sentiment analy- specific and word
embedding features
sis of financial tweet streams. The experiments in analyzing (word2vec)
stock market sentiments of a particular company showed
that changes in positive sentiment probability can be used
as indicators of the changes in stock closing prices.
Pagalu et al. [18] observed how well the changes in stock Table 2: Top five teams that participated in Task 5
prices of a company, the rises and falls, are correlated with of Semeval – predicting sentiment from microblogs
the public opinions being expressed in tweets about that messages
company. Their paper employed two textual representa- Authors Feature Set Approach
tions, Word2vec and N-gram, for analyzing the public sen- 1 Jiang et al. [10] linguistic, sentiment regressor ensemble
lexicon, domain-specific
timents in tweets. The authors applied sentiment analysis and word embedding
and supervised machine learning principles to the tweets and features (word2vec)
analyzed the correlation between stock market movements 2 Ghosal et al. [6] Character ngrams, Word Deep Learning En-
of a company and sentiments in tweets. They showed that ngrams, POS-tag, Lexi- semble
cons, Pointwise Mutual
positive news and tweets in social media about a company Information (PMI), and
would encourage people to invest in the stocks of that com- Microblog Specific Fea-
pany and as a result the stock price of that company would tures
increase. In addition, this study demonstrated a strong cor- 3 Deborah et al. [26] bag-of-words- based fea- Gaussian Process
tures
relation between the rise and falls in stock prices with the
4 Cabanski et al. [3] word embedding and lex- recurrent neural
public sentiments in tweets. icon features network
Given the link between sentiment and market dynamics, 5 Kumar et al. [12] Word Embedding an ensemble of
the analysis of public sentiment becomes a powerful method (GloVe), Tf-Idf Score, Support Vector
to predict the market reaction. However, the accuracy of Sentiment Lexicon Classifier and Lo-
gistic Regression.
machine learning-based sentiment analysis approaches rarely
exceeds seventy percent [31].
It is still important to mention the Task 5 of Semeval4 [4] 2.2 Aspect-based sentiment analysis
which discusses the “Fine-Grained Sentiment Analysis on
Aspect-based sentiment analysis systems treats a set of
Financial Microblogs (subtask a) and News (subtask b)”
texts (e.g., product reviews or messages from social media)
is specifically under the “Detecting sentiment, humor, and
discussing a particular entity (e.g., a new restaurant) to de-
truth” theme. This task contains two tracks, where the first
tect the main (e.g., the most frequently discussed) aspects
one concerns Microblog messages and the second one cov-
(features) of this entity (e.g., ‘food’, ‘service’) to estimate
ers News Statements and Headlines. The main goal behind
the average sentiment of the texts per aspect (e.g., how pos-
both tracks was to predict the sentiment score for each of the
itive or negative the opinions are on average for each aspect)
mentioned companies/stocks. The sentiment scores for each
[13].
One of the earliest studies on Aspect-based sentiment anal-
4
[Link] ysis is [9], which followed a frequency-based approach. The
basic idea is that frequently mentioned nouns are more likely for the microblog posts. Into these classes, the most often
to be aspects. To compensate for the errors resulting from class in news statements and headlines’ dataset are ’corpo-
ignoring infrequent names, the authors suggested exploiting rate’, ’m&a’ and ’stock’ with 327, 106 and 101 classifications
opinion words to find aspects. For this part, they proposed respectively. The most often classes into microblog posts’
to consider the nearest opinion word. This idea was used in dataset are ’stock’, ’price action’ and ’bullish’ with 546, 379
papers [32, 24]. and 203 classifications respectively. The datasets has lot
The main approaches use classifiers with expensive hand- of classes that are classified just once in whole dataset, the
crafted features based on n-grams, parts-of-speech, and sen- news statements and headlines’ dataset has 28 classes and
timent lexicon. For example, Popescu and Etzioni [23] sug- in the microblog posts’ dataset has 29 classes that were used
gested the use of part-of relationship to remove frequent just once.
noun phrases that are not aspects. Benefiting from statistics The instances of training dataset contains the followed
about the use of nouns in the English language, Scaffidi et attributes:
al. [29] improved the general approach of relying frequent
nouns to extract aspects. Long et al. [14] added the use of • Target: In microblog posts it refers to stocks compa-
information distance, and dependent words (adjectives). nies without the cashtag (that is present in sentence).
The studies about aspect-based sentiment analysis has In headlines it refers to the main company in sentence
gained more propulsion in some editions of SemEval567 [22, • Sentence: The news statements and headlines or the
21, 20]. These Semeval competitions provided datasets of microblog posts which sentiment is expressed
English reviews annotated at the sentence level with aspect
terms (e.g., “mouse”, “pizza”) and their polarity for specific • Snippets: The main piece of the sentence
domains, as well as aspect categories (e.g., “food”) and their
polarity. • Aspects: The aspects associated to the sentences
It is interesting to present that a one of the best scores
• Sentiment Score: A scaled sentiment between -1 and
[21] was achieved based on a Liblinear model [8] and the sec-
1, where 1 is very positive, -1 is very negative and
ond best score [5] was achieved with an SVM model trained
0 represents a neutral sentiment. It’s related to the
on the restaurants training data. The model used features
sentences and targets.
based on unigrams, sentiment lexica and PMI scores learnt
from TripAdvisor data. The team of EliXa [28] (one of the 3.2 Feature Engineering
challenge winners) used a multiclass SVM and features based
In this system, we tested a set of features and parame-
on word clusters, lemmas, n-grams, POS tagging, and sen-
ters trying get the best data fitting. The features used are
timent lexica.
described below:

3. EXPERIMENTAL EVALUATION 1. N-Grams

It’s a sequence of N words in a text/sentence. Was
3.1 Datasets used uni-grams, bi-grams, and tri-grams in the grid
The dataset provided by organization consists of 675 mi- search to see which one would perform better.
croblog messages, and 438 news statements and headlines
for training stage (see Table 3). There are 99 microblog 2. Tokenization
messages, and 93 news statements and headlines for testing Was used a library called Unitok11 to help tokenize
stage. The data became from web sites as Stocktwits8 , Red- the sentences. This library recognizes URLs, e-mail
dit9 , Wikinews10 , and other finance domain web pages. We addresses, DNS domains and IP addresses, and an in-
used just the dataset available for participating systems. teresting characteristics of that is the recognition of
abbreviations made in English.

Table 3: Statistis about training dataset 3. Word Replacements

Instances Positives Negatives Neutrals It was tested replacing or not the company’s stock sym-
Microblog Posts 675 440 234 1 bols presented in the sentences as well as positives and
Headlines 438 282 144 12 negatives words. To replace negatives and positives
words, it was used the N most similar words based on
The Table 3 refers to statistics about the sentiments of a cosine similarity, in our case N was 10. With all
training dataset. Positives, Negatives and Neutrals corre- the stock symbols we tested to replace with a common
spond to instances with sentiment score positive, negative or word (e.g., replacing $AAPL by ’company’ string) and
zero, respectively. The Instances columns refers to amount removing it from the sentence.
of sentences available in the dataset.
About the aspects, the datasets contains 95 aspect classes 4. Word Embeddings
for the news statements and headlines and 83 aspect classes As many tasks involving natural language processing,
5 we decided to use word embeddings that performs bet-
[Link] ter than a natural bag-of-words, which hardly capture
6
[Link] the words’s semantic. As it was widely used by the
7
[Link]
8 winners of Task 5 of Semeval 2017 we decided to use
[Link]
9 this technique in our experiments through the use of
[Link]
10 11
[Link] [Link]
 n
X
word2vec12 model. We have chosen to use word2vec
RStot = (yi ȳ)2
because of its size (3 millions vocabulary entries) and
i=1
because it is trained on Google News using the Miko-
liv’s [16] method method. tp
Precision =
tp + f p
5. TF-IDF (Term Frequency - Inverse Document
Frequency) tp
Recall =
Using TF-IDF weighting [27] was used to scale down tp + f n
the impact of frequencies words in both cases, regres-
sion and classification. Was tested using both l1 (a 1 precision · recall
F1 = 2 · 1 1 =2·
simple sum of the vector’s components) and l2 (based recall
+ precision precision + recall
on Euclidean Distance) normalizations and using or In our tests we have got the following results:
not inverse-document-frequency to reweighting.

3.3 Experimental Setup Table 4: Aspects Classification

In training stage, we have split the dataset using ran- Precision Recall F1-Score
domly 80%-20% proportion, where the 80% part was used Microblog Posts 0.6673 0.5592 0.5775
to train the algorithms and 20% was used to test and vali- Headlines and Statements 0.4992 0.4 0.4240
date our algorithms according to competition measures. The
results was evaluated using Mean Squared Error (MSE) and
R Squared (R2) to measure the sentiment analysis score and
precision, recall and F1-score for aspect classification task. Table 5: Sentiment Analysis
We have used, in both cases Support Vector Machine al- MSE R Squared
gorithms from sklearn13 , so we applied the Support Vector Microblog Posts 0.1281 0.2402
Regressor in sentiment analysis’ task and Support Vector Headlines and Statements 0.1052 0.2817
Classifier in aspect’s classification task.
In our experiments, we have used a grid search cross vali-
dation with 10-folds for both cases. On regression case (sen- 4. CONCLUSIONS
timent) we tested the values 10, 1 and 0.1 for the penalty
In this paper we have presented our solution implemented
parameter C of LinearSVR algorithm, and we validated with
for FiQA 2018 Task 1 using Linear Kernel(SVR and SVC)
the R Squared and Mean Squared Error, both metrics from
models from sklearn. We have tested almost all regressors,
sklearn metrics’ package. We have ran almost all sklearn’s
classifiers and some ensembles of them, but our best results
regressors to this case, including an ensemble of some of
were just using linear models alone for each part of the task.
them and our best results were with LinearSVR14 only.
The enhancing of the results can be made by using di↵er-
On classification case (aspects) we first binarized the as-
ent approaches of learning like simple neural networks or
pects based on a set of all possible aspects in training set.
LSTMs(Long Short Term Memory). In the future we can
To extract this set, we simple split the aspects given in ”/”
improve our solution with some of these approaches.
and cleared the values to remove punctuation and white
spaces. Then we ran the OneVsRest15 meta Classifier with
LinearSVC algorithm to classify the multi labeled aspects 5. REFERENCES
and the scorer function we used the precision, recall and F1- [1] S. Asur and B. A. Huberman. Predicting the future
score algorithms provided by sklearn, asking to refit the grid with social media. In Proceedings of the 2010
using F1-score. As it was done in regression, we tested the IEEE/WIC/ACM International Conference on Web
same parameters for the penalty parameter C of LinearSVC. Intelligence and Intelligent Agent Technology - Volume
01, WI-IAT ’10, pages 492–499, Washington, DC,
3.4 Results and Discussion USA, 2010. IEEE Computer Society.
In this section we will present the results that we have [2] J. Bollen, H. Mao, and X. Zeng. Twitter mood
got using the sklearn metric package, based on the metrics predicts the stock market. CoRR, abs/1010.3003,
given by the organization of FiQA(2018) for both cases of 2010.
sentiment analysis and aspect classification. The metrics [3] T. Cabanski, J. Romberg, and S. Conrad. Hhu at
used were Mean Squared Error(MSE) and R Squared for semeval-2017 task 5: Fine-grained sentiment analysis
the sentiment analysis and precision, recall and F1-score for on financial data using machine learning methods. In
the aspects classification and the formulas is shown below. Proceedings of the 11th International Workshop on
Semantic Evaluation (SemEval-2017), pages 832–836,
n Vancouver, Canada, August 2017. Association for
1X
MSE = (Yi Ŷi )2 . Computational Linguistics.
n i=1 [4] K. Cortis, A. Freitas, T. Daudert, M. Huerlimann,
12
[Link] M. Zarrouk, S. Handschuh, and B. Davis.
13
[Link] Semeval-2017 task 5: Fine-grained sentiment analysis
14
[Link] on financial microblogs and news. In Proceedings of
[Link] the 11th International Workshop on Semantic
15 Evaluation (SemEval-2017), pages 519–535.
[Link]
[Link] Association for Computational Linguistics, 2017.
 [5] O. De Clercq, M. Van de Kauter, E. Lefever, and phrases and their compositionality. In Advances in
V. Hoste. Lt3: Applying hybrid terminology neural information processing systems, pages
extraction to aspect-based sentiment analysis. In 3111–3119, 2013.
Proceedings of the 9th International Workshop on [17] A. Moore and P. Rayson. Lancaster A at semeval-2017
Semantic Evaluation (SemEval 2015), pages 719–724, task 5: Evaluation metrics matter: predicting
Denver, Colorado, June 2015. Association for sentiment from financial news headlines. CoRR,
Computational Linguistics. abs/1705.00571, 2017.
[6] D. Ghosal, S. Bhatnagar, M. S. Akhtar, A. Ekbal, and [18] V. S. Pagolu, K. N. R. Challa, G. Panda, and
P. Bhattacharyya. Iitp at semeval-2017 task 5: An B. Majhi. Sentiment analysis of twitter data for
ensemble of deep learning and feature based models predicting stock market movements. CoRR,
for financial sentiment analysis. In Proceedings of the abs/1610.09225, 2016.
11th International Workshop on Semantic Evaluation [19] A. Pak and P. Paroubek. Twitter as a corpus for
(SemEval-2017), pages 899–903. Association for sentiment analysis and opinion mining. In LREc,
Computational Linguistics, 2017. volume 10, 2010.
[7] R. Goonatilake, A. Herath, S. Herath, S. Herath, and [20] M. Pontiki, D. Galanis, H. Papageorgiou,
J. Herath. Intrusion detection using the chi-square I. Androutsopoulos, S. Manandhar, M. AL-Smadi,
goodness-of-fit test for information assurance, M. Al-Ayyoub, Y. Zhao, B. Qin, O. De Clercq,
network, forensics and software security. Journal of V. Hoste, M. Apidianaki, X. Tannier,
Computing Sciences in Colleges, 23(1):255–263, 2007. N. Loukachevitch, E. Kotelnikov, N. Bel, S. M.
[8] H. Hamdan, P. Bellot, and F. Bechet. lsislif: Feature Jiménez-Zafra, and G. Eryiğit. Semeval-2016 task 5 :
extraction and label weighting for sentiment analysis aspect based sentiment analysis. In Proceedings of the
in twitter. In In In Proceedings of the 9th International 10th International Workshop on Semantic Evaluation
Workshop on Semantic Evaluation (SemEval, 2015. (SemEval-2016), pages 19–30. Association for
[9] M. Hu and B. Liu. Mining and summarizing customer Computational Linguistics, 2016.
reviews. In Proceedings of the Tenth ACM SIGKDD [21] M. Pontiki, D. Galanis, H. Papageorgiou,
International Conference on Knowledge Discovery and S. Manandhar, and I. Androutsopoulos. SemEval-2015
Data Mining, KDD ’04, pages 168–177, New York, Task 12: Aspect Based Sentiment Analysis, pages
NY, USA, 2004. ACM. 486–495. The Association for Computational
[10] M. Jiang, M. Lan, and Y. Wu. Ecnu at semeval-2017 Linguistics, 2015.
task 5: An ensemble of regression algorithms with [22] M. Pontiki, D. Galanis, J. Pavlopoulos,
e↵ective features for fine-grained sentiment analysis in H. Papageorgiou, I. Androutsopoulos, and
financial domain. In Proceedings of the 11th S. Manandhar. Semeval-2014 task 4: Aspect based
International Workshop on Semantic Evaluation sentiment analysis. In Proceedings of the 8th
(SemEval-2017), pages 888–893. Association for International Workshop on Semantic Evaluation
Computational Linguistics, 2017. (SemEval 2014), pages 27–35, Dublin, Ireland, August
[11] S. Kar, S. Maharjan, and T. Solorio. Ritual-uh at 2014. Association for Computational Linguistics and
semeval-2017 task 5: Sentiment analysis on financial Dublin City University.
data using neural networks. In Proceedings of the 11th [23] A.-M. Popescu and O. Etzioni. Extracting product
International Workshop on Semantic Evaluation features and opinions from reviews. In Proceedings of
(SemEval-2017), pages 877–882. Association for the Conference on Human Language Technology and
Computational Linguistics, 2017. Empirical Methods in Natural Language Processing,
[12] A. Kumar, A. Sethi, M. S. Akhtar, A. Ekbal, HLT ’05, pages 339–346, Stroudsburg, PA, USA, 2005.
C. Biemann, and P. Bhattacharyya. Iitpb at Association for Computational Linguistics.
semeval-2017 task 5: Sentiment prediction in financial [24] G. Qiu, B. Liu, J. Bu, and C. Chen. Expanding
text. In Proceedings of the 11th International domain sentiment lexicon through double propagation.
Workshop on Semantic Evaluation (SemEval-2017), In Proceedings of the 21st International Jont
pages 894–898. Association for Computational Conference on Artifical Intelligence, IJCAI’09, pages
Linguistics, 2017. 1199–1204, San Francisco, CA, USA, 2009. Morgan
[13] B. Liu. Sentiment Analysis and Opinion Mining. Kaufmann Publishers Inc.
Morgan & Claypool Publishers, 2012. [25] L. Rotim, M. Tutek, and J. Snajder. Takelab at
[14] C. Long, J. Zhang, and X. Zhut. A review selection semeval-2017 task 5: Linear aggregation of word
approach for accurate feature rating estimation. In embeddings for fine-grained sentiment analysis of
Proceedings of the 23rd International Conference on financial news. In Proceedings of the 11th
Computational Linguistics: Posters, COLING ’10, International Workshop on Semantic Evaluation,
pages 766–774, Stroudsburg, PA, USA, 2010. SemEval@ACL 2017, Vancouver, Canada, August 3-4,
Association for Computational Linguistics. 2017, pages 866–871, 2017.
[15] Y. Mansar, L. Gatti, S. Ferradans, M. Guerini, and [26] A. D. S, S. M. Rajendram, and T. T. Mirnalinee.
J. Staiano. Fortia-fbk at semeval-2017 task 5: Bullish Ssn mlrg1 at semeval-2017 task 4: Sentiment analysis
or bearish? inferring sentiment towards brands from in twitter using multi-kernel gaussian process
financial news headlines. CoRR, abs/1704.00939, 2017. classifier. In Proceedings of the 11th International
[16] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and Workshop on Semantic Evaluation (SemEval-2017),
J. Dean. Distributed representations of words and pages 709–712, Vancouver, Canada, August 2017.
 Association for Computational Linguistics.
[27] G. Salton and C. Buckley. Term-weighting approaches
in automatic text retrieval. Inf. Process. Manage.,
24(5):513–523, Aug. 1988.
[28] I. n. San Vicente, X. Saralegi, and R. Agerri. Elixa: A
modular and flexible absa platform. In Proceedings of
the 9th International Workshop on Semantic
Evaluation (SemEval 2015), pages 748–752, Denver,
Colorado, June 2015. Association for Computational
Linguistics.
[29] C. Scaffidi, K. Bierho↵, E. Chang, M. Felker, H. Ng,
and C. Jin. Red opal: Product-feature scoring from
reviews. In Proceedings of the 8th ACM Conference on
Electronic Commerce, EC ’07, pages 182–191, New
York, NY, USA, 2007. ACM.
[30] J. Smailovic, M. Grcar, N. Lavrac, and M. Znidarsic.
Stream-based active learning for sentiment analysis in
the financial domain. Information Sciences, 285:181 –
203, 2014. Processing and Mining Complex Data
Streams.
[31] P. Takala, P. Malo, A. Sinha, and O. Ahlgren.
Gold-standard for topic-specific sentiment analysis of
economic texts. In N. C. C. Chair), K. Choukri,
T. Declerck, H. Loftsson, B. Maegaard, J. Mariani,
A. Moreno, J. Odijk, and S. Piperidis, editors,
Proceedings of the Ninth International Conference on
Language Resources and Evaluation (LREC’14),
Reykjavik, Iceland, may 2014. European Language
Resources Association (ELRA).
[32] L. Zhuang, F. Jing, and X.-Y. Zhu. Movie review
mining and summarization. In Proceedings of the 15th
ACM International Conference on Information and
Knowledge Management, CIKM ’06, pages 43–50, New
York, NY, USA, 2006. ACM.

View publication stats