INGEOTEC at MEX-A3T: Author profiling and aggressiveness analysis in Twitter using TC and EvoMSA

Abstract

This paper describes our participation in the MEX-A3T challenge for Aggressiveness Detection and Author Profiling tasks for Mexican Spanish language. We used two approaches, TC and EvoMSA systems. The first one is a minimalistic text categorization system, and the second one is a two-level architecture for Sentiment Analysis using information from different models on the current text analyzed to get a final prediction by a consensus view.

Keywords

• emotion classification
• text categorization
• author profiling

Introduction

Author profiling and aggressiveness detection are essential tasks for marketing, digital text forensics, cyber-bullying, security, among others. Aggressiveness detection allows us to identify offenses and misbehavior expressed in text and commonly shared in social networks. Author profiling is related to extract information from author's texts such as gender, age, and other kinds of personality traits. To increase the research in those areas, several international competitions have been organized to deal with them, such as PAN [ 14 ], SemEval [ 12 ], and TASS [ 11 ]. As part of this, recently, the MEX-A3T5 [ 3 ] contest which is part of IBEREVAL'186 workshop has been launched on the research community. The purpose of MEX-A3T is deal with author profiling and aggressiveness detection in Spanish language focusing on Mexican Twitter users. MEX-A3T contest presents two tasks to classify Twitter text. The first one is aggressiveness detection task where systems have to determine whether a tweet is aggressive or not automatically. The second task is author profiling, where systems have to automatically determine the occupation and location (place of residence) of users from their tweets [ 3 ].

In the literature, several approaches have been proposed to tackle both author profiling and aggressiveness detection. Such is the case of [ 7 ] where a system based

5 https://mexa3t.wixsite.com/home

6 https://sites.google.com/view/ibereval-2018

2 on lexicon, fuzzy logic, and statistical approaches is proposed to detect aggressiveness in a text, or the proposed in [ 6 ] where a Lexical Syntactic Feature is used to detect offensive content and then be able to identify a potential offensive user in social media. Agrawal & Goncalves [ 1 ] propose a combination of classifiers to identify gender associated with a set of texts. This propose includes TFIDF representation, and a dimension reduction of it, to finally employs Naive Bayes and Linear SVM as classifiers. In [ 4 ] several stylometric features are considered for identifying males from females in several age groups. Stylometric features are also used in [ 5 ] where tri-grams and complementary-weighted Second Order Attributes are employed.

In this work, we present the methodology proposed to deal with profiling and aggressiveness detection, which includes two approaches, TC and EvoMSA systems. TC is a minimalistic text categorization system, and EvoMSA is a two-level architecture for Sentiment Analysis using information from different models getting the final prediction by consensus. Both systems will be more detailed in following sections. The rest of the paper is organized as follows. Section 2 describes our system and the general approach to model the problem. Section 3 detail the experimental methodology and the achieved results. Finally, conclusions and future work are given in Section 4. 2

System

Description

As commented, we use two systems to tackled the author profiling and the aggressiveness text detection tasks: TC and EvoMSA, respectively. On the one hand, TC is used mainly to evaluate author profiling task because in our experiments it obtained the best performance in this tasks. On the other hand, EvoMSA is used to evaluate aggressiveness task. In the following paragraphs, we describe these approaches. 2.1

EvoMSA EvoMSA7 has two stages. The first one, namely B4MSA [ 16 ], uses SVMs to predict their decision function values of a given text. On the second hand, EvoDAG [ 9, 10 ] is a classifier based on Genetic Programming with semantic operators which makes the final prediction through a combination of all the decision function values. Furthermore, EvoMSA is open to being fed with different models such as TC, and lexicon-based models, and EvoDAG. It is an architecture of two phases to solve classification tasks, see Figure 1. In the first part, a set of different classifiers are trained with datasets provided by the contests and others as additional knowledge, i.e., whatever knowledge could be integrated into EvoMSA. In this case, we used tailor-made lexicons for the aggressiveness task: aggressiveness words and affective words (positive and negative), see Section 2.3. The precise configuration of our benchmarked system is described in Section 3. 2.2

TC

TC8 (a.k.a. B4MSA) is a minimalistic system able to tackle general text classification tasks independently of domain and language. For complete details of the model see [ 17 ]. Roughly speaking, TC creates text classifiers searching for the best models in given

7 https://github.com/INGEOTEC/EvoMSA

8 https://github.com/INGEOTEC/microTC

INGEOTEC at MEX-A3T 3 configuration space. A configuration consists of instructions to enable several preprocessing functions, a combination of tokenizers among the power set of several possible ones (character q-grams, n-word grams, and skip-grams), and a weighting scheme such as TF, TFIDF, or several distributional schemes. TC uses an SVM classifier with a linear kernel. A text transformation feature could be binary (yes/no) or ternary (group/delete/none) option. Tokenizers denote how texts must be split after applying the process of each text transformation to texts. Tokenizers generate text chunks in a range of lengths, all tokens generated are part of the text representation. In Table 1, we can see details of text transformations used in our solution for detecting aggressiveness and profiling. For example, Tokenizers used for Profiling are unigrams, bigrams, trigrams of words, and q-grams of 1 and five characters length, and skip-grams of two words with a gap between them. To introduce extra knowledge into our approach for aggressiveness task, we used two lexicon-based models. The first, Up-Down model produces a counting of affective words, i.e., for a given text, it is produced in two indexes one for positive words, and another for negative words. We created a positive-negative lexicon based on the several Spanish affective lexicons [ 2, 15, 13 ] and enriched with Spanish WordNet [ 8 ]. The other Bernoulli 4 EvoDAG9 [ 9, 10 ] is a Genetic Programming system specifically tailored to tackle classification problems on very high dimensional vector spaces and large datasets. EvoDAG uses the principles of Darwinian evolution to create models represented as a directed acyclic graph (DAG). Due to lack of space, we refer the reader to [ 9 ] where EvoDAG is broadly described. It is important to mention that EvoDAG does not have information regarding whether input Xi comes from a particular class decision function, consequently from EvoDAG point of view all inputs are equivalent. 3

Results

As mentioned, we split the dataset provided by organizers into 70-30 partition for training and test. We run several configurations of our systems. In Table 2 and Table 3 results are shown. In the case of the aggressiveness task, Table 2, we use the F1-aggressiveness score to measure the performance. The basic configuration of EvoMSA is one model based on B4MSA's predictions using the training set provided by the competition. In case of EvoMSA, plus symbol indicates the model added to the EvoMSA basic configuration.

In our experiments, the best performance we obtained is the combination of basic EvoMSA along with a Lexicon-based Bernoulli model (LexB), and a counting model of affective words (UpDown). This configuration was used to evaluate on the gold standard that our approach obtained 0.4883 in F1-aggressiveness class, see Table 4, INGEOTEC team.

In the case of author profiling task, the best performance was TC system for Occupation classes. Thus, we decided to apply the same approach to Location classes. Table 3 shows the results of author profiling in our experiments. Our best system was used to evaluate on the gold standard that our approach obtained 0.4470 of F1-score for Occupation, 0.8155 of F1-score for Location and 0.6312 of F1-score on average of both, see Table 5, INGEOTEC team .

Tables 4 and 5 list the top-final rankings for aggressiveness detection task and user profiling task, respectively, more details of all results of the contest see [ 3 ]. Our INGEOTEC team reached the first place in aggressiveness detection and the third place in the author profiling task.

9 https://github.com/mgraffg/EvoDAG

INGEOTEC at MEX-A3T 5 In this paper was presented our solution for the MEX-A3T challenge. For Aggressiveness Detection task, we applied EvoMSA system which can integrate different models as additional knowledge as we have shown. Also, we applied our generic text classifier, TC, for author profiling task. Both systems are designed to be multilingual, language and domain independent as much as possible. For the training step, we use extra knowledge coded into affective and aggressiveness lexicons our robust solution (EvoMSA) performs well for the aggressiveness task; however, there is room for further improvements in performance for author profiling task using another sort of knowledge such as semantic information into our architecture.

M. Graff. et al.

References

1. Agrawal , M. , Goncalves , T. : Age and gender identification using stacking for classification . notebook for pan at clef 2016 ( 2016 )

2. de Albornoz, J.C. , Plaza , L. , Gervas , P. : Sentisense: An easily scalable concept-based affective lexicon for sentiment analysis . In: Proceedings of LREC 2012 . pp. 3562 { 3567 ( 2012 )

3. Alvarez-Carmona , M.A. , Guzman-Falcon , E. , Montes-y Gomez , M. , Escalante , H.J. , Villasen~or- Pineda , L. , Reyes-Meza , V. , Rico-Sulayes , A. : Overview of mex-a3t at ibereval 2018: Authorship and aggressiveness analysis in mexican spanish tweets . In: Notebook Papers of 3rd SEPLN Workshop on Evaluation of Human Language Technologies for Iberian Languages (IBEREVAL) , Seville, Spain, September ( 2018 )

4. Bilan , I. , Zhekova , D. : Caps: A cross-genre author profiling system ( 2016 )

5. Bougiatiotis , K. , Krithara , A. : Author profiling using complementary second order attributes and stylometric features ( 2016 )

6. Chen , Y. , Zhou , Y. , Zhu , S. , Xu , H. : Detecting offensive language in social media to protect adolescent online safety . In: 2012 International Conference on Privacy, Security, Risk and Trust and 2012 International Confernece on Social Computing . pp. 71 { 80 (Sept 2012 ). https://doi.org/10.1109/SocialCom-PASSAT. 2012 .55

7. Del Bosque , L.P. , Garza , S.E. : Aggressive text detection for cyberbullying . In: Gelbukh, A. , Espinoza , F.C. , Galicia-Haro , S.N . (eds.) Human-Inspired Computing and Its Applications . pp. 221 { 232 . Springer International Publishing, Cham ( 2014 )

8. Fernandez-Montraveta , A. , Vazquez , G. , Fellbaum , C. : The spanish version of wordnet 3.0 . Text Resources and Lexical Knowledge . Mouton de Gruyter pp. 175 { 182 ( 2008 )

9. Graff , M. , Tellez , E.S. , Miranda-Jimenez , S. , Escalante , H.J.: Evodag: A semantic genetic programming python library . In: 2016 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC) . pp. 1 { 6 (Nov 2016 ). https://doi.org/10.1109/ROPEC. 2016 .7830633

10. Graff , M. , Tellez , E.S. , Escalante , H.J. , Miranda-Jimenez , S. : Semantic Genetic Programming for Sentiment Analysis . In: Sch A~ 14 tze, O., Trujillo , L. , Legrand , P. , Maldonado , Y . (eds.) NEO 2015 , pp. 43 { 65 . No. 663 in Studies in Computational Intelligence , Springer International Publishing ( 2017 ), dOI: 10.1007/978-3-319-44003-3 2

11. Mart nez-Camara, E., D az- Galiano , M. , Garc a-Cumbreras, M. , Garc a-Vega, M. , Villena-Roman , J. : Overview of tass 2017 . In: Proceedings of TASS 2017: Workshop on Semantic Analysis at SEPLN (TASS 2017 ). vol. 1896 ( 2017 )

12. Mohammad , S.M. , Bravo-Marquez , F. , Salameh , M. , Kiritchenko , S. : Semeval-2018 Task 1: Affect in tweets . In: Proceedings of International Workshop on Semantic Evaluation (SemEval-2018) . New Orleans, LA, USA ( 2018 )

13. Perez-Rosas , V. , Banea , C. , Mihalcea , R. : Learning sentiment lexicons in spanish . In: LREC . vol. 12 , p. 73 ( 2012 )

14. Rangel , F. , Rosso , P. , Potthast , M. , Stein , B. : Overview of the 5th author profiling task at pan 2017: Gender and language variety identification in twitter . Working Notes Papers of the CLEF ( 2017 )

15. Sidorov , G. , Miranda-Jimenez , S. , Viveros-Jimenez , F. , Gelbukh , A. , Castro-Sanchez , N. , Velasquez , F. , D az- Rangel , I. , Suarez-Guerra , S. , Trevin~o, A. , Gordon , J.: Empirical study of machine learning based approach for opinion mining in tweets . In: Proceedings of the 11th Mexican International Conference on Advances in Artificial Intelligence - Volume Part I . pp. 1 { 14 . MICAI' 12 , Springer-Verlag, Berlin, Heidelberg ( 2013 )

16. Tellez , E.S. , Miranda-Jimenez , S. , Graff , M. , Moctezuma , D. , Suarez , R.R. , Siordia , O.S.: A simple approach to multilingual polarity classification in Twitter . Pattern Recognition Letters 94 , 68 { 74 ( 2017 )

17. Tellez , E.S. , Moctezuma , D. , Miranda-Jimenez , S. , Graff , M.: An automated text categorization framework based on hyperparameter optimization . Knowledge-Based Systems 149 , 110 { 123 ( 2018 )