TASS 2017: Workshop on Semantic Analysis at SEPLN, septiembre 2017, págs. 13-21
Overview of TASS 2017
Resumen de TASS 2017
Eugenio Martı́nez-Cámara1 , Manuel C. Dı́az-Galiano2 , M. Ángel Garcı́a-Cumbreras2 ,
Manuel Garcı́a-Vega2 , Julio Villena-Román3
1
Ubiquitous Knowledge Processing Lab (UKP-TUDA)
Department of Computer Science, Technische Universität Darmstadt
2
Grupo de Investigación SINAI
Universidad de Jaén, Jaén, Spain
3
MeaningCloud, Madrid, Spain
1
camara@ukp.informatik.tu-darmstadt.de,
2
{mcdiaz, magc, mgarcia}@ujaen.es, 3 julio.villena@sngular.team
Abstract: This paper describes TASS 2017, the sixth edition of the Workshop on
Semantic Analysis at SEPLN 2017. The main aim is to encourage the research
and development of new resources, algorithms and techniques for different tasks of
semantic analysis in Spanish. In this paper, we present the proposed tasks, the
generated datasets, and a summary of the submitted systems.
Keywords: TASS 2017, sentiment analysis, semantic analysis
Resumen: Este artı́culo describe la sexta edición del Taller de Análisis Semántico en
la SEPLN, conocido como TASS 2017. TASS tiene como objetivo principal incentivar
la investigación y desarrollo de recursos, técnicas, algoritmos y herramientas para
tareas relacionadas con el análisis semántico en español. A continuación, se describen
las tareas propuestas para la edición 2017, ası́ como los corpus creados y utilizados,
los distintos participantes y los resultados obtenidos.
Palabras clave: TASS 2017, análisis de opiniones, análisis semántico
1 Introduction ish is far away to the advance of other lan-
Since some years ago, Natural Language Pro- guages like English. Consequently, TASS2
cessing (NLP) researchers have been working (Taller de Análisis de Sentimientos en la SE-
on the discovery of the meaning of utterances PLN / Workshop on Sentiment Analysis at
from different perspectives. One of those per- SEPLN) was born in 2012 with the aim of
spectives is the understanding of the subjec- fostering the development of specific NLP
tive information or rather opinion informa- techniques for the computational treatment
tion. The task of Sentiment Analysis (SA) of opinions of text written in Spanish. The
is the result of this study, and it is defined previous editions in 2016 (Garcı́a-Cumbreras
as the computational treatment of opinion, et al., 2016), 2015 (Villena-Román et al.,
sentiment and subjectivity in text (Pang and 2015), 2014 (Villena Román et al., 2015),
Lee, 2008). 2013 (Villena-Román et al., 2014) and 2012
However, the potential semantic informa- (Villena-Román et al., 2013) have yielded
tion encoded in an utterance is so rich and outstanding linguistic resources such as the
broad that different new NLP tasks have General Corpus of TASS and some datasets
arosen, such as argumentation mining, stance for the task of polarity classification at as-
classification, irony detection or the consid- pect level, used by a great number of research
ered tasks in the different editions of our sib- groups and companies as reference for Span-
ling workshop SemEval.1 ish. Additionally, a research community has
The Spanish language is the second native been created around TASS that usually par-
language in the world and the second lan- ticipate in the workshop and contribute with
guage in number of speakers. Nevertheless, vivid discussions about the state-of-the-art
the progress of the NLP research in Span- and the next challenges in SA in Spanish.
1 2
http://alt.qcri.org/semeval2018/ http://www.sepln.org/workshops/tass
ISSN 1613-0073 Copyright © 2017 by the paper's authors. Copying permitted for private and academic purposes.
Eugenio Martínez-Cámara, Manuel C. Díaz-Galiano, M. Ángel García-Cumbreras, Manuel García-Vega, Julio Villena-Román
The organization committee of the work- General Corpus of TASS3 and a new corpus,
shop has updated its name in the edition InterTASS, which was specifically developed
of 2017 because of the need of widening in 2017 for the task (see Section 3).
the gamut of semantic tasks in TASS. The Datasets were annotated with 4 different
new name of TASS is Workshop on Seman- polarity labels positive, negative, neu-
tic Analysis at SEPLN (Taller de Análisis tral and none), and systems had to iden-
Semántico en la SEPLN ), which allows to tify the intensity of the opinion expressed in
keep the acronym TASS. The change of the each tweet in any of those 4 intensity lev-
name is a call to researchers on other seman- els. For the two sets of the General Corpus
tic tasks (argumentation mining, irony de- of TASS, which was annotated in 6 polarity
tection, stance classification...) to organize tags, a direct translation from P+ into P and
a shared-task for the treatment of semantic N+ into N was performed so that the evalu-
information in Spanish for the next edition. ation is consistent with InterTASS and based
The edition of 2017 proposes two subtasks, on 4 levels of intensity of polarity.
polarity classification at document (tweet) All datasets were divided into training, de-
level (Task 1) and aspect level polarity clas- velopment and test datasets, which were pro-
sification (Task 2). Apart from reusing sev- vided to participants in order to train and
eral datasets of previous editions, a new evaluate their systems. Systems were allowed
dataset was specifically generated for this to use any set of data as training dataset, i.e.
edition. The new dataset is called Inter- the training set of InterTASS, other train-
TASS, which is composed of more than 2,000 ing sets from the previous editions of TASS
tweets annotated at four opinion intensity or other sets of tweets. However, using the
level (positive, neutral, negative and test set of InterTASS and the test set of the
none). Further details about the tasks and datasets of previous editions as training data
the datasets in Sections 2 and 3 respectively. was obviously forbidden. Apart from that,
The edition of 2017 has attracted the par- participants could use any kind of linguistic
ticipation of 11 teams, mainly from Spain resource for the development of their classifi-
and America. Most of the systems follow the cation model.
state-of-the-art of SA, which is the use of a Participants were expected to submit 3 ex-
deep learning architecture. Most of the teams periments per each evaluation set, so each
participated in Task 1, and a few of them in participant team could submit a maximum of
Task 2, which is an indication that polarity 9 files of results. Results must be submitted
classification at aspect level is a tough task. in a plain text file with the following format:
The rest of this paper is organized as
follows. Section 2 presents in more details tweet id \t polarity
the two subtasks of TASS 2017. Section 3
describes the datasets and how we created Allowed values for polarity were P, NEU,
them. Section 4 presents the submitted sys- N and NONE.
tems and the results reached by them. Fi- Accuracy and the macro-averaged ver-
nally, Section 5 concludes and points the fu- sions of Precision, Recall and F1 were used
ture work in TASS. as evaluation measures. Systems were be
ranked by the Macro-F1 and Accuracy mea-
2 Tasks sures.
TASS 2017 has proposed two tasks address- 2.2 Task 2. Aspect-based
ing the challenging task of SA in Twitter in Sentiment Analysis
Spanish. This second task proposed the development
of aspect-based polarity classification sys-
2.1 Task 1. Sentiment Analysis at tems. Two datasets from previous editions
Tweet level were used to evaluate the systems: Social-
TV and STOMPOL (see Section 3). The two
This main task focused on the evaluation of
datasets were annotated for aspect, the main
polarity classification systems at tweet level
in Spanish. Systems were evaluated on three 3
The entire test set annotated with 4 classes, the
different datasets: the two test sets of the 1k test set also annotated with 4 classes.
14
Overview of TASS 2017
category of aspect, and the polarity of the Then, the general sentiment of a random
opinion about the aspect. Systems had to selection of tweets was manually annotated
classify the opinion about the given aspect in by five annotators. We used a scale of 4 lev-
3 different polarity labels (positive, nega- els of polarity: positive (p), neutral (neu),
tive, neutral). negative (n) and no sentiment tag (none).
Participants were expected to submit up Each tweet was finally annotated at least by
to 3 experiments for each corpus, each in a three annotators. When a tweet has the same
plain text file with the following format: tag by two of more annotators, the process
end. If not, each annotator revised the tweet
tweetid \t aspect \t polarity again, until it has the same tag by two of
more annotators. The annotation resulted in
Allowed polarity values were p, neu and n. a corpus of 3,413 tweets, which was split into
For evaluation, exact match with a single three datasets: training, development and
label combining “aspect-polarity” was used. test. Table 1 shows the size of each dataset
Similarly to Task 1, the macro-averaged ver- of InterTASS corpus.
sion of Precision, Recall and F1, and Ac-
curacy were the evaluation measures, and Corpus #Tweets
Macro-F1 were used for ranking the systems. Training 1,008
Developement 506
3 Datasets Test 1,899
TASS 2017 provides four datasets to the par- Total 3,413
ticipants for the evaluation of their systems.
Three of the datasets were used in previous Table 1: Number of tweets in each dataset of
editions, and a new dataset was created for InterTASS
TASS 2017, namely InterTass. Each tweet includes its ID (tweetid), the
The datasets will be made freely available creation date (date) and the user ID (user).
to the community after the workshop.4 Due to restrictions in the Twitter API Terms
of Service,6 it is forbidden to redistribute a
3.1 InterTASS
corpus that includes text contents or informa-
International TASS Corpus (InterTASS ) is tion about users. However, it is valid if those
a new corpus released this year for general fields are removed and instead IDs (includ-
task (Task 1). The goal of the organiza- ing Tweet IDs and user IDs) are provided.
tion of TASS is the creation of a corpus The actual message content can be easily ob-
of tweets written in the Spanish language tained by making queries to the Twitter API
spoken in Spain and in different Hispano- using the tweetid.
American countries. We release the first ver- The training set was released, so the par-
sion of InterTASS in TASS 2017, which is ticipants could train and validate their mod-
only composed of tweets posted in Spain and els. The test corpus was provided without
written in the Spanish language spoken in any annotation and has been used to evalu-
Spain. ate the results. The InterTass statistics are
More than 500,000 tweets were collected, in Table 2.
from July 2016 to January 2017, using some
keywords. The downloaded set of tweets was Training Dev. Test
filtered out in order to meet the following re- P 317 156 642
quirements: NEU 133 69 216
N 416 219 767
• The language of the tweets must be NONE 138 62 274
Spanish5 , Total 1,008 506 1,899
• each tweet must contain at least one ad-
Table 2: Number of tweets in each dataset
jective,
and class of InterTASS
• the minimum length of each tweet must
be four words. The three datasets of the corpus are three
4
XML files. Figure 1 shows an example of an
Further information for requesting the datasets InterTASS XML file.
in: http://www.sepln.org/workshops/tass/.
5 6
We used the python library langdetect. https://dev.twitter.com/terms/api-terms
15
Eugenio Martínez-Cámara, Manuel C. Díaz-Galiano, M. Ángel García-Cumbreras, Manuel García-Vega, Julio Villena-Román
3.3 Social-TV Corpus
768212591105703936 the 2014 Final of Copa del Rey champi-
martitarey13 onship in Spain between Real Madrid and
@estherct209 jajajaja la F.C. Barcelona, played on 16 April 2014 at
tuya y la d mucha gente seguro Mestalla Stadium in Valencia. After filter-
!! Pero yo no puedo sin mi ing out useless information a subset of 2.773
melena me muero tweets was selected. The details of the corpus
2016-08-23 22:25:29 are described in (Villena-Román et al., 2015;
es Garcı́a-Cumbreras et al., 2016).
All tweets were manually annotated with
31 different aspects and its sentiment polar-
N ity. It was randomly divided into training
AGREEMENT set (1.773 tweets) and test set (1.000 tweets),
with a similar distribution of both aspects
and sentiments.
Figure 3 shows a tweet from the Social-TV
corpus.
Figure 1: A tweet from the XML file of In-
terTASS corpus
Para mi, ISCO sentiment>
0000000000 ha hecho un partidazo.
q es adicto al drama! Ja ja ja "P">El mejor partido
2011-12-02T02:59:03 desde que llego al
es Real MadridP+ sentiment>.
AGREEMENT
Figure 3: A tweet from the XML file of the
entretenimiento Social-TV corpus
3.4 STOMPOL
Figure 2: A tweet from the XML file of the STOMPOL (corpus of Spanish Tweets for
General Corpus of TASS Opinion Mining at aspect level about POLi-
tics) is a corpus of Spanish tweets developed
for the research in opinion mining at aspect
3.2 General corpus level. Each tweet in the corpus has been
manually annotated by two annotators, and
The General Corpus of TASS has 68,000 a third one in case of disagreement, with the
tweets, written in Spanish by about 150 well- sentiment polarity at aspect level.
known personalities and celebrities of the The corpus is composed of 1,284 tweets,
world of politics, economy, communication, and has been divided into training set (784
mass media and culture, between November tweets), which is provided for building and
2011 and March 2012. The details of the cor- validating the systems, and test set (500
pus are described in (Villena-Román et al., tweets) that will be used for evaluation.
2015; Garcı́a-Cumbreras et al., 2016). Fig- The details of the corpus are described
ure 2 shows a tweet from the General Corpus in (Villena-Román et al., 2015; Garcı́a-
of TASS. Cumbreras et al., 2016). Figure 4 shows a
16
Overview of TASS 2017
tweet from the STOMPOL corpus. current neural network (CNN) and the third
one with a long-short term memory (LSTM)
recurrent neural network (RNN). The perfor-
@rosadiezupyd lamenta que el # mance of each configuration depends on the
empleo semble classifier system for the first task.
no termine de estabilizarse y The author generated quantitative features
dice que el from the tweets, such as the number of
#paro " used lists of opinion bearing words like iSOL
sigue siendo dram~
A¡tico" http://t (Molina-González et al., 2013), as well as the
.co/1xdS3UjJWk #EPA inversion of the polarity of words following a
window shifting approach for negation han-
dling. The base classifiers of the ensemble
Figure 4: STOMPOL XML example system were Logistic Regression and SVM.
The system followed two ensemble strategies,
namely stacking and maximum classification
4 Participants and Results confidence. The maximum confidence strat-
Most of the systems submitted in TASS 2017 egy outperformed the stacking strategy and
are based on the use of deep learning tech- it reached the highest accuracy value with the
niques as the state-of-the-art in SA in Twit- test set of the InterTASS dataset.
ter. However, some of the systems are based Montañés Salas et al. (2017) used the
on traditional machine learning methods and classifier FastText (Joulin et al., 2016) for
others are meta-classifiers whose inputs are only classifying the test set of the InterTASS
the output of deep learning systems and tra- dataset. The authors performed a traditional
ditional machine learning algorithms. We de- pre-processing to the input tweets, however
pict the main features of the systems submit- the substitution of words with a emotional
ted in the subsequent paragraphs. meaning by their synonyms from a list of
Table 3, Table 4 and Table 5 show the words with a emotional meaning (Bradley
results reached by the submitted systems in and Lang, 1999) stands out.
Task 1, using the test sets of InterTASS cor- Rosá et al. (2017) participated in the two
pus and the General Corpus (full test and 1k tasks. Concerning the first task, the authors
test). Table 6 and Table 7 shows the results submitted three systems: 1) a SVM classifier
reached by the submitted systems in Task 2, with word embeddings and quantitative lin-
using the test sets of Social-TV corpus and guistic properties as features; 2) a deep neu-
STOMPOL corpus respectively. ral network grounded on the use of a CNN for
Hurtado, Pla, and González (2017) par- encoding the input tweets; and 3) the combi-
ticipated in the two tasks. They submitted nation of the two previous classifiers by the
the same system for both tasks, and the only selection of the output class with a higher
difference between the tasks lies in the char- probability mean from the two previous clas-
acteristics of the input. The input of the sifiers. The third strategy outperformed the
first task is the entire tweet, meanwhile the other ones in two test sets of Task 1. Re-
input in the second task is the context of garding the Task 2, the authors submitted
the aspects, which is previously determined. two SVM classifiers mainly grounded on the
The authors created a set of domain-specific use word embeddings.
word embeddings following the approach of Garcı́a-Vega et al. (2017) submitted four
Tang (2015). The former word embeddings systems for the classification of the test set
set is jointly used with a general-domain set of the InterTASS dataset. The first two sys-
of embeddings to represent the tokens of the tems are a SVM classifier that uses word-
tweets. The authors evaluated three different embeddings as features. The difference be-
neural networks architectures, the first one is tween these two systems lies in the use of ad-
a multilinear perceptron (MLP), the second ditional tweets from the users of the training
encodes the tweets with a convolutional re- set. The intention of the authors was the in-
17
Eugenio Martínez-Cámara, Manuel C. Díaz-Galiano, M. Ángel García-Cumbreras, Manuel García-Vega, Julio Villena-Román
System M-F1 Acc. System M-F1 Acc.
ELiRF-UPV-run1 0.493 0.607 INGEOTEC- 0.577 0.645
RETUYT-svm cnn 0.471 0.596 evodag 003
ELiRF-UPV-run3 0.466 0.597 jacerong-run-1 0.569 0.706
ITAINNOVA-model4 0.461 0.576 jacerong-tass 2016- 0.568 0.705
jacerong-run-2 0.460 0.602 run 3
jacerong-run-1 0.459 0.608 ELiRF-UPV-run2 0.549 0.659
INGEOTEC- 0.457 0.507 ELiRF-UPV-run3 0.548 0.725
evodag 001 RETUYT-svm cnn 0.546 0.674
RETUYT-svm 0.457 0.583 jacerong-run-2 0.545 0.701
tecnolengua-sent only 0.456 0.582 ELiRF-UPV-run1 0.542 0.666
ELiRF-UPV-run2 0.450 0.436 RETUYT-cnn 0.541 0.638
ITAINNOVA-model3 0.445 0.561 RETUYT-cnn3 0.539 0.654
RETUYT-cnn3 0.443 0.558 tecnolengua-run3 0.528 0.657
SINAI-w2v-nouser 0.442 0.575 tecnolengua-final 0.517 0.632
tecnolengua-run3 0.441 0.576 tecnolengua- 0.508 0.652
tecnolengua- 0.441 0.595 531F1 no ngrams
sent only fixed INGEOTEC- 0.447 0.514
ITAINNOVA-model2 0.436 0.576 evodag 001
LexFAR-run3 0.432 0.541 OEG-victor2 0.389 0.496
LexFAR-run1 0.430 0.539 INGEOTEC- 0.364 0.449
jacerong-run-3 0.430 0.576 evodag 002
SINAI-w2v-user 0.428 0.569 OEG-laOEG 0.346 0.407
INGEOTEC- 0.403 0.515 GSI-64sent99ally 0.324 0.434
evodag 002
OEG-victor2 0.395 0.451 Table 4: Task 1 General Corpus of TASS (full
OEG-victor0 0.383 0.433 test) results
OEG-laOEG 0.377 0.505
LexFAR-run2 0.372 0.490 ditionally they used EvoDAG, a GP system
GSI-sent64-189 0.371 0.524 that combines all decision values predicted by
SINAI-embed-rnn2 0.333 0.391 B4MSA systems. They also used two exter-
GSI-sent64-149-ant-2 0.306 0.479 nal datasets to train the B4MSA algorithm.
GSI-sent64-149-ant 0.000 0.000
Navas-Loro and Rodrı́guez-Doncel (2017)
Table 3: Task 1 InterTASS corpus results participated only on Task 1. They experi-
mented with two classifier algorithms, Multi-
nominal Naı̈ve Bayes and Sequential Minimal
troduction of the use of language of each user Optimization for SVM. Furthermore they
in the classification. The two last systems are used morphosyntactic analyses for negation
deep neural networks grounded on the use of detection, along with the use of lexicons and
LSTM RNN for the encoding of the mean- dedicated preprocessing techniques for de-
ing of the input tweets. The first neural ar- tecting and correcting frequent errors and ex-
chitecture uses word embeddings as features, pressions in tweets.
and the second one the TF-IDF value of each Araque et al. (2017) have proposed, for
word of the tweets. Task 1, a RNN architecture composed of
Moctezuma et al. (2017) participation LSTM cells followed by a feed-forward net-
was based on an ensemble of SVM classi- work. The architecture makes use of two
fiers combined into a non-linear model cre- different types of features: word embeddings
ated with genetic programming to tackle and sentiment lexicon values. The recurrent
the task of global polarity classification at architecture allows them to process text se-
tweet level. They used B4MSA algorithm, quences of different lengths, while the lexicon
a proposed entropy-based term weighting inserts directly into the system sentiment in-
scheme, which is a baseline supervised learn- formation. Two variations of this architec-
ing system based on the SVM classifier, an ture were used: a LSTM that iterates over
entropy-based term-weighting scheme. Ad- the input word vectors, and on the other
18
Overview of TASS 2017
System M-F1 Acc. System M-F1 Acc.
RETUYT-svm 0.562 0.700 ELiRF-UPV-run1 0.537 0.615
RETUYT-cnn4 0.557 0.694 RETUYT-svm2 0.508 0.590
RETUYT-cnn2 0.555 0.694 ELiRF-UPV-run3 0.486 0.578
INGEOTEC- 0.526 0.595 ELiRF-UPV-run2 0.486 0.541
evodag 003 C100T-PUCP-run3 0.445 0.528
tecnolengua-run3 0.521 0.638 C100T-PUCP-run1 0.415 0.563
ELiRF-UPV-run1 0.519 0.630 C100T-PUCP-run2 0.414 0.517
jacerong-tass 2016- 0.518 0.625 RETUYT-svm 0.377 0.514
run 3
jacerong-run-1 0.508 0.678 Table 7: Task 2 STOMPOL corpus results
jacerong-run-2 0.506 0.673
ELiRF-UPV-run2 0.504 0.596 lexicons are represented by the bag-of-word
tecnolengua-final 0.488 0.618 model and they are weighted using Term Fre-
tecnolengua-run4 0.483 0.612 quency measure at tweet level.
ELiRF-UPV-run3 0.477 0.588 Moreno-Ortiz and Pérez Hernández
INGEOTEC- 0.439 0.431 (2017) have proposed, for Task 1, a clas-
evodag 002 sification model based on the Lingmotif
INGEOTEC- 0.388 0.486 Spanish lexicon, and combined this with a
evodag 001 number of formal text features, both general
OEG-victor3b 0.367 0.386 and CMC-specific, as well as single-word
OEG-victor2 0.366 0.412 keywords and n-gram keywords. They use
OEG-laOEG 0.346 0.448 logistic regression classifier trained with the
GSI-run-1 0.327 0.558 optimal set of features, SVM classifier on
GSI-64sent99ally 0.321 0.499 the same features set. Sentiment features
are obtained with Lingmotif SA engine
Table 5: Task 1 General Corpus of TASS (1k) (sentiment feature set, text feature set and
results keywords feature set).
System M-F1 Acc. 5 Conclusion and Future work
ELiRF-UPV-run3 0.537 0.615
TASS was the first workshop about sentiment
ELiRF-UPV-run2 0.513 0.600 analysis focused on the processing of texts
ELiRF-UPV-run1 0.476 0.625 written in Spanish. In this edition, 11 teams
RETUYT-svm2 0.426 0.595 participated with a total of 123 runs, most of
RETUYT-svm 0.413 0.493 them in the InterTASS task.
Anyway, the released corpora and the re-
Table 6: Task 2 Social-TV corpus results
ports from participants will for sure be help-
ful for other research groups approaching
hand a combination of the input word vectors these tasks.
and polarity values from a sentiment lexicon. The future work will mainly go in two
Tume Fiestas and Sobrevilla Cabezudo directions. On the one hand, the organiza-
(2017) have proposed, for Task 2, an ap- tion of one o more shared-tasks for the treat-
proach based on word embeddings for polar- ment of semantic information in Spanish like
ity classification at aspect-level. They used those mentioned above (argumentation min-
word embeddings to get the similarity be- ing, irony detection and stance classification).
tween words selected from a training set and On the other hand, the extension and im-
make a model to classify each polarity of each provement of the InterTASS corpus. This
aspect for each tweet. Their results show that corpus has been received with great inter-
the more tweets are used, the better accuracy est, almost 90% of the experiments have been
is obtained. developed in the first task, so an exhaustive
Reyes-Ortiz et al. (2017) have proposed, analysis of the behavior of the corpus in this
for Task 1, a system that uses machine learn- task will shows the right way for a new ver-
ing, vector support machines algorithm and sion of the corpus.
lexicons of semantic polarities at the level of TASS corpora will be released after the
lemma for Spanish. Features extracted from workshop for free use by the research com-
19
Eugenio Martínez-Cámara, Manuel C. Díaz-Galiano, M. Ángel García-Cumbreras, Manuel García-Vega, Julio Villena-Román
munity. Joulin, A., E. Grave, P. Bojanowski, and
T. Mikolov. 2016. Bag of tricks for ef-
Acknowledgement ficient text classification. arXiv preprint
arXiv:1607.01759.
This research work is partially supported by
the project REDES (TIN2015-65136-C2-1-R) Moctezuma, D., M. Graff, S. Miranda-
and a grant from the Fondo Europeo de De- Jiménez, E. S. Tellez, A. Coronado, C. N.
sarrollo Regional (FEDER). Sánchez, and J. Ortiz-Bejar. 2017. A ge-
netic programming approach to sentiment
References analysis for twitter: Tass’17. In Proceed-
ings of TASS 2017: Workshop on Senti-
Araque, O., R. Barbado, J. F. Sánchez-Rada,
ment Analysis at SEPLN co-located with
and C. A. Iglesias. 2017. Applying recur-
33nd SEPLN Conference (SEPLN 2017),
rent neural networks to sentiment analysis
volume 1896 of CEUR Workshop Proceed-
of spanish tweets. In Proceedings of TASS
ings, Murcia, Spain, September. CEUR-
2017: Workshop on Sentiment Analysis
WS.
at SEPLN co-located with 33nd SEPLN
Conference (SEPLN 2017), volume 1896 Molina-González, M. D., E. Martı́nez-
of CEUR Workshop Proceedings, Murcia, Cámara, M.-T. Martı́-Valdivia, and J. M.
Spain, September. CEUR-WS. Perea-Ortega. 2013. Semantic orientation
for polarity classification in spanish re-
Bradley, M. M. and P. J. Lang. 1999. Af- views. Expert Systems with Applications,
fective norms for english words (anew): 40(18):7250 – 7257.
Stimuli, instruction manual, and affective
ratings. Technical report, Center for Re- Montañés Salas, R. M., R. del Hoyo Alonso,
search in Psychophysiology, University of J. Vea-Murguı́a Merck, R. Aznar Gimeno,
Florida. and F. J. Lacueva-Pérez. 2017. FastText
como alternativa a la utilización de deep
Cerón-Guzmán, J. A. 2017. Classier ensem- learning en corpus pequeños. In Proceed-
bles that push the state-of-the-art in sen- ings of TASS 2017: Workshop on Senti-
timent analysis of spanish tweets. In Pro- ment Analysis at SEPLN co-located with
ceedings of TASS 2017: Workshop on Sen- 33nd SEPLN Conference (SEPLN 2017).
timent Analysis at SEPLN co-located with
33nd SEPLN Conference (SEPLN 2017). Moreno-Ortiz, A. and C. Pérez Hernández.
2017. Tecnolengua lingmotif at tass 2017:
Garcı́a-Cumbreras, M. A., J. Villena-Román, Spanish twitter dataset classification com-
E. Martı́nez-Cámara, M. C. Dı́az-Galiano, bining wide-coverage lexical resources and
M. T. Martı́n-Valdivia, and L. A. Ureña text features. In Proceedings of TASS
López. 2016. Overview of tass 2016. 2017: Workshop on Sentiment Analysis
In TASS 2016: Workshop on Sentiment at SEPLN co-located with 33nd SEPLN
Analysis at SEPLN, pages 13–21. Conference (SEPLN 2017), volume 1896
of CEUR Workshop Proceedings, Murcia,
Garcı́a-Vega, M., A. Montejo-Ráez, M. C.
Spain, September. CEUR-WS.
Dı́az-Galiano, and S. M. Jiménez-Zafra.
2017. Sinai en tass 2017: Clasificación Navas-Loro, M. and V. Rodrı́guez-Doncel.
de la polaridad de tweets integrando in- 2017. Oeg at tass 2017: Spanish sentiment
formación de usuario. In Proceedings analysis of tweets at document level. In
of TASS 2017: Workshop on Sentiment Proceedings of TASS 2017: Workshop on
Analysis at SEPLN co-located with 33nd Sentiment Analysis at SEPLN co-located
SEPLN Conference (SEPLN 2017). with 33nd SEPLN Conference (SEPLN
2017), volume 1896 of CEUR Workshop
Hurtado, L.-F., F. Pla, and J.-A. González. Proceedings, Murcia, Spain, September.
2017. Elirf-upv en tass 2017: Análisis de CEUR-WS.
sentimientos en twitter basado en apren-
dizaje profundo. In Proceedings of TASS Pang, B. and L. Lee. 2008. Opinion mining
2017: Workshop on Sentiment Analysis at and sentiment analysis. Foundations and
SEPLN co-located with 33nd SEPLN Con- Trends in Information Retrieval, 2(1-2):1–
ference (SEPLN 2017). 135.
20
Overview of TASS 2017
Reyes-Ortiz, J. A., F. Paniagua-Reyes, Jiménez Zafra. 2015. Tass 2014 -
B. Priego-Sánchez, and M. Tovar. 2017. the challenge of aspect-based sentiment
Lexfar en la competencia tass 2017: analysis. Procesamiento del Lenguaje
Análisis de sentimientos en twitter basado Natural, 54(0):61–68.
en lexicones. In Proceedings of TASS
2017: Workshop on Sentiment Analysis
at SEPLN co-located with 33nd SEPLN
Conference (SEPLN 2017), volume 1896
of CEUR Workshop Proceedings, Murcia,
Spain, September. CEUR-WS.
Rosá, A., L. Chiruzzo, M. Etcheverry, and
S. Castro. 2017. Retuyt en tass 2017:
Análisis de sentimientos de tweets en
español utilizando svm y cnn. In Proceed-
ings of TASS 2017: Workshop on Senti-
ment Analysis at SEPLN co-located with
33nd SEPLN Conference (SEPLN 2017).
Tang, D. 2015. Sentiment-specific repre-
sentation learning for document-level sen-
timent analysis. In Proceedings of the
Eighth ACM International Conference on
Web Search and Data Mining, WSDM
’15, pages 447–452, New York, NY, USA.
ACM.
Tume Fiestas, F. and M. A. Sobre-
villa Cabezudo. 2017. C100tpucp at
tass 2017: Word embedding experiments
for aspect-based sentiment analysis in
spanish tweets. In Proceedings of TASS
2017: Workshop on Sentiment Analysis
at SEPLN co-located with 33nd SEPLN
Conference (SEPLN 2017), volume 1896
of CEUR Workshop Proceedings, Murcia,
Spain, September. CEUR-WS.
Villena-Román, J., J. Garcı́a-Morera, M. A.
Garcı́a-Cumbreras, E. Martı́nez-Cámara,
M. T. Martı́n-Valdivia, and L. A. Ureña
López. 2015. Overview of tass 2015.
In TASS 2015: Workshop on Sentiment
Analysis at SEPLN, pages 13–21.
Villena-Román, J., J. Garcı́a-Morera,
S. Lana-Serrano, and J. C. González-
Cristóbal. 2014. Tass 2013 - a second
step in reputation analysis in spanish.
Procesamiento del Lenguaje Natural,
52(0):37–44, March.
Villena-Román, J., S. Lana-Serrano,
E. Martı́nez-Cámara, and J. C. González-
Cristóbal. 2013. Tass - workshop on
sentiment analysis at sepln. Proce-
samiento del Lenguaje Natural, 50:37–44.
Villena Román, J., E. Martı́nez Cámara,
J. Garcı́a Morera, and S. M.
21