This document describes a text change of representation approach to the task of Aggressiveness Identi cation in Twitter , as part of IberLEF2019. The task consists in classifying tweets as being aggressive or not aggressive. Tweets have been written by Mexican authors who come from a wide variety of backgrounds. Our approach consists of a change of the space of representation of text into statistical descriptors which characterize the text. In addition, dimensional reduction is performed to 6 characteristics per class in order to make the method suitable for a Big Data environment. Frequency Analysis Interpolation (FAI) is the approach we use to achieve rank 12th among 24 submissions.
Social media has become a new standard of communications in the last years. Every year more and more people actively participate in the content creation, sometimes under the shield of anonymity. Social media has become a complex communication channel in which usually o ensive contents are written. Supervising the content and banning o ensive messages currently is a subject of high interest for social media administrators. In this task we address the problem of detecting aggressive comments in tweets from Mexican users. Spanish is a language plenty with nuances, a characteristic which excels in Mexican Spanish and their usage of "albur", where nothing means what it seems. This problem will be considered as an Author Pro ling task, since the main goal is building a system which would ideally detect author whose content is o ensive to women and/or immigrant.
Author Pro ling is widely studied and some new ideas arise from time to time
[
Training Test
7,700 3,156
dimensionality of the information for each author to 6 characteristics per class.
This representation, Frequency Analysis Interpolation, is used to codify the texts
for each user and this codi ed information is used as input data to support vector
machines with linear kernel. In a Big Data environment, reducing the number
of characteristics from thousands to 6 per class allows an e cient way to deal
with high volumes at high speed. With this will in mind a previous method was
tested which can be checked at [
Whereas a complete description of the corpus used in this task can be found at
[
For this task, the nal score corresponds to the F 1 -measure for the aggressive class. 3
Our goal was to develop a method that was language independent and that required no prior knowledge of the language used by the authors. We started implementing Term Frequency (TF) representation for each tweet in the corpus, counting how many times each word appears in each author, each tweet in this case, and globally for all tweets. We denote TFa as the term frequency vector for author a.
T F a = T F(w1;a); T F(w2;a); : : : ; T F(wm;a) (1) TF is used since this way we could represent a priori class dependent probability for each term for each class simply by counting the number of times a term occurs for each class, and dividing this amount by the number of times this term shows for all classes. Let F be the frequency term vector for all classes.
F =
X T F a a2A In order to achieve that, one vector per class is generated. The vector length is the number of words in the vocabulary. For each word, we divide the number of times this word shows for this class, and divide it by the number of times the word shows in all classes. We denote Ck as the term frequency vector for class k that belong to the set of all classes K.
Ck =
X T F a8k 2 K a2Ak These vectors are then used to codify the texts. Each word in the text is substituted by the a priori probability for each class in as many arrays as classes. Once we have codi ed the text, six statistic values are calculated for each of the classes: 1. Mean. 2. Standard Deviation. 3. Skewness. 4. First Tertile's length. 5. Second Tertile's length.
6. Third Tertile's length.
At this point, for every author, 6 characteristics per class are calculated and concatenated in a single vector. This vector is used to feed the Support Vector Machines with Linear kernel. LinearSVC support vector machine from Pythons Sklearn library is used to train the model and, of course, to predict the results. One vector is created for each author. This vector contains the six characteristic mentioned above for every class, concatenated.
Although the FAI representation was developed and mainly tested for Author
Pro ling tasks, it has previously been used for agressiveness detection at HatEval
in SemeEval 2019 with good results for multi-class classi cation [
This task is evaluated and ranked using F1-score for the aggressive class.
Acc = number of correctly predicted instances
total number of instances P = number of correctly predicted instances number of predicted labels (2) (3) (4) (5)
Team INGEOTEC task aggressiveness run 1
Casavantes Aggressiveness Text GLP-run2 Aggressiveness Text
GLP-run4 Aggressiveness Text mineriaUNAM aggressiveness secondaryRun mineriaUNAM aggressiveness primaryRun
GLP-run3 Aggressiveness Text GLP-run1 Aggressiveness Text
Baseline (Trigrams) LyRAggressiveness T ext Run3 LyRAggressiveness T ext Run6
Victor run1
OscarGaribo run1 LyR Aggressiveness Text Run5 LyR Aggressiveness Text Run2 LyR Aggressiveness Text Run1
Baseline (BoW ) OscarGaribo run2 LASTUS-UPF run2
LASTUS-UPF run1 mdmolina agressive detection
Victor run2
Aspie96 secondary LyR Aggressiveness Text Run4 hzegheru Aggressiveness Text
Aspie96 primary number of correctly predicted instances number of labels in the gold standard
F 1 = 2 * P * R
P + R
(6)
(7)
FAI is usually penalized by the fact of having two classes. As we could see at
[
We have used FAI, a method developed under the scope of Author Pro ling tasks to approach HatEval Task. Prior testing performed with our method has been done under di erent conditions, since there were always more tweets (minimum 100) per author. Thus, there was much more vocabulary to learn from, and more vocabulary per author. We have to point that our method can easily be updated with new data, since the only required task to be done is recomputing the a priori probability vectors once the new labeled data is available, and train the machine learning algorithm, support vector machines in this speci c case. As future work we think of exploring new con gurations of our method. One of the immediate ones is to remove some of the vocabulary from the vocabulary we use to codify the tweets. We have seen in our in house testing that some problems require the more the better vocabulary, for example age identi cation, whereas some others work better if low used words are removed from the vocabulary, for example removing words used by less than 1% of the authors.