Violence Recognition in Spanish Words using Data Mining
Adolfo Flores Moreno, Silvia B. González-Brambila and Juan G. Vargas-Rubio
Dívisión de Ciencias Básicas e Ingeniería, Universidad Autónoma Metropolitana-Azcapotzalco, México, D.F., México
Keywords: Violence, Classification, Data Mining.
Abstract: Violent behavior in our society has been studied from many points of view, yet many cause-effect relations
remain unexplained. Security personnel are normally trained to be alert and recognize potential violent
behavior, but they cannot be 100% effective in recognizing it due to the monotonous nature of their job.
This paper presents the first results of a work in progress detecting violence from the analysis of words in
conversations. We used a set of videos with two person conversations in Spanish and classified them as
violent and non violent. The audio of the conversations was extracted and converted to text. We used
“Ward”, “K-means” and “PAM” (clValid, 2014) to group words, performing a clValid analysis we found
that the hierarchical technique was the best. The percentages of frequency were computed for each term and
the SVM (Meyer, 2014) technique was applied, from which we found that there were unclassifiable terms.
In three of the tests the prediction was erroneous and in another three we obtained good predictions with
respect to the test set.
1 INTRODUCTION
One of the reasons for surveillance is to keep order
and safety in public and private places. The use of
video cameras is necessary, but the personnel in
charge of watching the images may get bored and
miss some events.
Violence is a deliberate behavior with the
intention of causing physical or psychological injury
to other person, and it may also lead to property
destruction. Violence is generally associated with
aggressiveness, but the second is not always
destructive (Villanueva et al, 2007).
Violence is hard to define due to its ambiguity
and subjectivity (Derbas & Quénot, 2014). In
technical definitions violence is defined by visual-
auditory indicators, like high-speed movements
(Gong et al, 2008) and words.
There are many factors that lead to violence, bad
mood, frustration, substance abuse, prescriptions,
also social and environmental factors like family
situations, job instability, friend circle, etc. that can
contribute to it (RStudio, 2014). Violence can appear
in any situation and consequences can be physical
and psychological, generating behaviors and
unwanted situations.
This work shows how standard video
surveillance can be improved by automatically
triggering some alarms when a conversation that is
monitored has a high probability of being violent.
When this happens some actions can be taken to
avoid violence.
Most of the work related to violence is dedicated
to detect violent scenes in videos. Violent scene
detection (VSD) is an important research problem
and promises several applications like movie/film
inspection, video on demand, semantic video
indexing and retrieval. Recent works are using low-
level and mid-level features to represent violence
concept (Lam et al, 2013) or visual and spatio-
temporal features (Derbas & Quénot, 2014).
However, at present, classification and filtering is
done manually (Fujii & Yoshimura, 2011).
The rest of this paper is organized as follows. We
describe the work in Section 2. Section 3 includes
detailed information of the experiments, results and
our analysis. Finally, conclusions and future work
are presented in Section 4.
2 DESCRIPTION OF THE WORK
For this work, we collected different video segments
where a two person conversation was taking place,
and they were classified as violent or nonviolent.
The audio track was extracted and converted to text.
210
Flores Moreno A., González-Brambila S. and Vargas-Rubio J..
Violence Recognition in Spanish Words using Data Mining.
DOI: 10.5220/0005072502100216
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2014), pages 210-216
ISBN: 978-989-758-048-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
The text files were preprocessed to build a term
matrix. The frequency of each word was computed
and presented in a graph and in a word cloud. The
grouping results were validated using “Ward”, “K-
means” and “PAM” with clValid, and we used SVM
for the classification.
2.1 Collecting Videos
We collected 100 video segments from different
sources in the internet in the formats mp4 y wmv.
The segments contain conversations of two persons
in Spanish that were classified manually as violent
or non violent (Tables 1 and 2).
Table 1: Collected video segments.
Type Qty.
Violent
53
Non violent
47
Total
100
Table 2: Duration of the videos.
Minutes
Minimum
1:12
Maximum
8:39
2.2 Audio Extraction and Processing
Once the video segments were classified, the audio
was extracted in mp3 format using “Adobe Premiere
Pro CC” (Adobe, 2014). Using the same program,
the audio was converted to text with the Speech
Analysis Model for Spanish. The results were saved
in a .txt file.
There were some problems with the conversion.
The conversion of each audio segment lasted from 3
to 10 minutes, but it was not always successful.
Some of the conversations required manual
transcription because the model could not identify
some words correctly. The number of files
successfully converted by Adobe was 83 while the
other 17 had to be converted manually.
2.3 Pre-Processing
The file “stopwords.txt” that contains all 617
meaningless words like prepositions, articles and
conjunctions was created to filter the texts. Also the
gender was removed manually from some words to
obtain better results.
The following steps were used to pre-process the
text files for both, violent and non violent
conversations:
Load the file.
Build the corpus.
Convert to lowercase.
Eliminate spaces.
Eliminate punctuation signs
Load the file with meaningless words
(stopwords.txt).
Remove generic words (used in R).
Remove meaningless words (from
“stopwords.txt” and other meaningless
words).
Build the term matrix.
Table 3 shows how the number of terms is
reduced due to the pre-processing. The 100 more
frequent words for violent and non violent
conversations are shown in Figures 1 to 4.
Table 3: Number of terms before and after pre-processing.
Type of
conversations
Terms before
Pre-processing
Terms after
Pre-processing
Non violent
1971 1549
Violent
1388 1133
Total
3359 2682
From Figures 1 to 4 we observe that there are
some words that appear frequently in both, violent
and non violent conversations. The difference is that
they are preceded by different words in each case. In
the case of violent conversations, those preceding
words are normally “bad words”.
2.4 Clustering
To validate the best grouping using the clustering
methods “Ward”, “K-means” and “PAM” we used
the connectivity, silhouette and Dunn index. Figure 5
shows the hierarchical grouping with 8 clusters is
the best choice and it is validated with Dunn for the
file with all the conversations.
Using the clValid package from R we
determined that 8 groups is the best solution, and it
is shown in the Dendrogram of Figure 6.
From the grouping generated by “k-means” we
observe that in each cluster we have the words that
appear frequently together in conversations. In
clusters 1, 2, and 7 we have the words used in
violent conversations, and in the others are the
words used in non violent conversations (see Fig. 7).
ViolenceRecognitioninSpanishWordsusingDataMining
211
Figure 1: The most frequent 100 words in non violent conversations.
Figure 3: The most frequent 100 words in violent conversations.
Figure 2: A cloud of the 100 more frequent words for non
violent conversations.
Figure 4: A cloud of the 100 more frequent words for
violent conversations.
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
212
Figure 5: Grouping validation.
Figure 6: Dendrogram of 8 groups.
Figure 7: The most important 5 words in each cluster using K-Means.
Using “K-medoids” (see Figure 8) the clusters 1 to 7
contain the words used in both, violent and non
violent conversations, clusters 8 and 9 contain the
words most probably used in violent conversations,
and cluster 10 contains the words most used in non
violent conversations.
In order to classify words with SVM the “e1071”
package for R was used (SVM, 2014). The term
matrix was saved along with the frequency of each
term. If a term appeared more than 50% in non
violent conversations (1 to 47), it was classified as
non violent. On the other side, if a term appeared
more than 50% in violent conversations (48 to 100),
it was classified violent. If the percentage was the
ViolenceRecognitioninSpanishWordsusingDataMining
213
Figure 8: Clusters obtained with the K- Medoids
algorithm.
same (50%=V=NV=50%) it was marked not
classified (SIN).
Data was saved in a text file that contains:
Terms, percentage in non violent conversations
(NV), percentage in violent conversations (V), and
the classification (Clas). Figure 9 shows the results
of the classification.
Figure 9: Results of SVM classes and levels.
Among the 454 non violent terms in the test set,
398 were predicted as non violent and 56 as violent.
Also there are 31 terms marked not classified from
which 28 were predicted non violent and 3 as
violent. On the other side there are 278 violent terms
in the test set. From those, 60 were predicted as non
violent and 218 as violent.
3 INTERPRETATION OF
RESULTS
Table 4 shows the number of terms before and after
pre-processing. Form the table we observe that many
words are removed in order to obtain better results
when classifying a conversation.
Table 4: Data from the term matrix, violent conversations
vs. non violent conversations.
Analyzed
conversations
Terms
before pre-
processing
Terms after
pre-
processing
Non
violent
47 1971 1549
Violent
53 1388 1133
Total
100 3359 2682
Table 5 shows the most frequent terms in violent
and non violent conversations. We observe that there
are words that appear in both like “Quieres” among
others.
Table 5: Most frequent terms, violent conversations vs.
non violent conversations.
Conversation type Term
Repetitions
Non violent
Casa 25
Quiero 23
Vas 22
Quieres 17
Miedo 14
Violent
Maldit@ 25
Pendej@ 22
Dinero 20
Favor 18
Quieres 14
Table 6 shows the grouping validation between 2
and 10 clusters obtained with “clValid” where we
can compare the results with “Ward”, “K-means”
and “PAM”. Also, the validation measures
connectivity, silhouette and Dunn are shown.
Observing the results in Table 6 it is possible to
choose the most efficient technique given that the
smaller values indicate better grouping.
From Table 7, the validation of results, it is
observed that the hierarchical grouping is the best in
each case, and the optimum is with the Dunn
validation with 8 clusters.
4 CONCLUSIONS AND FUTURE
WORK
After analyzing the results we consider pre-
processing an important step to obtain the words
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
214
Table 6: Comparison of results using: “Ward”, “K-means” and “PAM”.
Table 7: Optimum results for grouping.
Score Method Clusters
Connectivity
2.9290 hierarchical 2
Dunn
0.5586 hierarchical 8
Silhouette
0.8218 hierarchical 2
most used in violent and non violent conversations.
The word clouds and term graphs were very useful
to visualize the most used words and their frequency
in each type of conversationIt was observed that in
both, violent and non violent conversations, there are
terms that are used with similar frequencies, but in
the violent conversations these words are
accompanied by other words considered “bad
words” or “rudeness”.
Once the word grouping was performed, eight
groups of words that appear frequently together in
conversations (violent or non violent) were obtained.
Not all the grouping techniques used were adequate,
but the hierarchic technique ward was the most
efficient given that the closeness of their words was
much better than k-means and PAM.
When classifying terms in violent or non violent
using SVM it was observed that some terms cannot
be classified due to the fact that they appear with
similar frequency in both types of conversations. In
the training sets of SVM it was observed that the
performance depends on the size of the test set.
Using the procedure presented in this work it is
convenient to experiment with a larger number of
video segments, and also use a better pre-processing
that can include synonyms and removal of gender in
most words. Also it is convenient to try more data
mining techniques in order to make a thorough
comparison and obtain better results due to the
larger number of terms.
With the procedure presented in this work it is
possible to design a system capable of classifying
automatically conversations as violent and non
violent. And this system can evolve to make this
classification in real-time in order to trigger some
alarm when a conversation turns violent in order to
alert security personnel to take measures.
Another idea that can be explored is to pre-
classify the video segments in categories like sports,
political, family, commercial, etc. and also by region
or social context in order to help the classification
process.
REFERENCES
clValid: An R Package for Cluster Validation, [Online],
Available at: http://www.jstatsoft.org/v25/i04/paper
[Retrieved January 2014]
Meyer, D., “Support Vector Machines: The Interface to
libsvm in package e1071”, September 2012, [Online],
Available at: http://cran.r-project.org/web/packages/
e1071/vignettes/svmdoc.pdf [Retrieved January 2014]
Brun, R. E., Senso, J. A., Minería textual, [Online],
Available at:
http://www.elprofesionaldelainformacion.com/conteni
dos/2004/enero/2.pdf [Retrieved January 2014]
Montes-y-Gómez, M., Minería de texto: Un nuevo reto
computacional, [Online], Available at: http://ccc.
inaoep.mx/~mmontesg/publicaciones/2001/MineriaTe
xto-md01.pdf [Retrieved January 2014]
Villanueva, V. J., Escribano, M., Isorna, M., Pellicer, J.,
Alapont, L., Pellicer, P., Programa de apoyo al ámbito
familiar: Agresividad y violencia, Editorial IES Pablo
Serrano. Andorra (Teruel), España, 2007.
Adobe Premiere Pro CS6, [Online], Available at:
http://www.adobe.com/mena_en/products/premiere.ht
ml [Retrieved January 2014]
Modelos de análisis de voz para Adobe Premiere Pro CS6,
[Online], Available at: http://www.adobe.com/es/
products/premiere/extend.displayTab3.html,
[Retrieved January 2014]
RStudio v0.97.551, [Online], Available at: http://www.
rstudio.com/ide/download/desktop [Retrieved January
2014]
ViolenceRecognitioninSpanishWordsusingDataMining
215
Support Vector Machines in R, [Online], Available at:
http://www.jstatsoft.org/v15/i09/paper [Retrieved
February 2014]
An Introduction to R, [Online], Available at: http://cran.r-
project.org/doc/manuals/R-intro.pdf [Retrieved Janua-
ry 2014]
R Data Import/Export, [Online], Available at: http://cran.r-
project.org/doc/manuals/r-release/R-data.html [Retrie-
ved January 2014]
Grün, B., Hornik, K., “Topicmodels: An R Package for
Fitting Topic Models”, 2011, [Online], Available at:
http://cran.r-project.org/web/packages/topicmodels/
vignettes/topicmodels.pdf [Retrieved January 2014]
Feinerer, I., Hornik, K., “Package ‘tm’”, August 2013,
[Online], Available at: http://cran.r-project.org/
web/packages/tm/tm.pdf [Retrieved January 2014]
Wainschenker R., Doorn, J., Castro M., “Medición
Cuantitativa de la Velocidad del Habla”, 2002,
[Online], Available at: http://www.sepln.org/
revistaSEPLN/revista/28/28-Pag99.pdf [Retrieved
March 2014]
Zhao, Y., R and Data Mining: Examples and Case Studies,
2013, [Online], Available at: http://cran.r-
project.org/doc/contrib/Zhao_R_and_data_mining.pdf
[Retrieved March 2014]
Data Preprocessing Techniques for Data Mining, [Online],
Available at: http://iasri.res.in/ebook/win_school_aa/
notes/Data_Preprocessing.pdf [Retrieved March
2014]
Hastie T., Tibshirani R., Friedman J., The Elements of
Statistical Learning. Data Mining, Inference, and
Prediction. Springer, 2001.
Bourel, M., Support Vector Machines, [Online], Available
at: http://www.iesta.edu.uy/wiki/images/7/71/SVM
SemEstadistica.pdf [Retrieved March 2014]
Derbas, N., Quénot, G., "Joint Audio-Visual Words for
Violent Scenes Detection in Movies", Internation
Conference on Multimedia Retrieval ICMR'14
Glasgow, United Kingdom, April 01-04, 2014
Gong, Y., Wang, W., Jiang, S., Huang, Q., Gao, W.,
"Detecting violent scenes in movies by auditory and
visual cues. Advances in Multimedia Information
Processing PCM 2008, 317-326, Spring Berlin
Heidelberg, 2008
Lam, V., Phan, S., Ngo, T., Le, D., Duong, D., Satoh, S.,
"Violent Scene Detection Using Mid-level Feature",
SolCt'13, Danang, Vietman. December 5-6, 2013.
Fujii, Y.,, Yoshimura, T., Ito, T., "Filtering Harmful
Sentences based on Three-Word Co-occurrence",
CEAS'11, Perth, Western Australia, Australia,
September 1-2, 2011.
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
216
