Supervised WSD Using Master- Slave Voting Technique

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661,p-ISSN: 2278-8727, Volume 17, Issue 2, Ver. V (Mar – Apr. 2015), PP 66-74
www.iosrjournals.org
DOI: 10.9790/0661-17256674 www.iosrjournals.org 66 | Page
Supervised WSD Using Master- Slave Voting Technique
Boshra F. Zopon AL_Bayaty1
, Dr. Shashank Josh2
1
Bharati Vidyapeeth University, 2
Bharati Vidyapeeth University
1
Department of Computer Science,Yashwantrao Mohite College,
2
Department of Computer Engineering, Engineering College India, Pune, Pin no.411046
1
AL- Mustansiriyah University, Baghdad, Iraq
Abstract: The Word sense disambiguation approaches contain number of methods such as stacking, voting, in
this paper we combined three approaches, Decision List as Master approach and Naïve Bayes, Adaboost a
Slaves approaches, for combining models to increase the accuracy and WSD performance.
Keywords: Decision List, Naïve Bayes, Adaboost, Senseval-3, WSD, WordNet, Master-Slave technique, Voting,
combination.
I. Introduction
Number of systems can be used for the ability of an algorithm to continue operating despite
abnormalities in input, calculation etc, means it improves robustness. It there may be possibilities to create
independent module for WSD, in that case we act each module individually for better performance. If there is
combination of number WSD systems, the errors are find out and they are detected by a factor of 1/N. The main
task of WSD is to assign sense to word in context. The senses of a word can be typically taken from dictionary.
Various machine learning (ML) approaches are explained or evaluate to produce successful Word sense
disambiguation systems [1]
. But how the performance between different algorithms can measure still remains the
question. Decision List and Naïve Bayes are used improve the performance. This performance is improved by
collecting the voting. After collecting the voting accuracy of finding correct sense will get increased. Word
Sense Disambiguation is open problem, so the output of any approach will depend on your particular data.
Disambiguation means choosing one meaning from pre-specified set. The main idea is to determine similarity
between every meaning and the context.
II. Master – Slave Technique
In WSD there are two main methods voting and stacking, the voting method can be weighted or non-
weighted, the weighted approach done by adding more weight to the classifier which is selected by votes and
got more accuracy among some classifiers. Here in the figure below show you our suggestion which called
Master- Slave technique. In this model several classifier as slaves suns separately, and one or more can select by
the Master. The selection depends on the accuracy, in case found two classifiers got same results, the master has
control and decision to select according the reputation each one.
The Master-Slave technique [2]
is a technique to achieve improvement in WEB Search engine results,
by combination one or more of supervised classifiers, figure (1), shows the master-slave technique. In this
experiment we combine two supervised classifier, Decision List as master approach [3]
, and Naïve Bayes,
Adaboost as Slaves classifiers.
Figure 1: Mater – Slave technique
The Reputation

2.1 Decision List
A Decision List is an ordered list with conjunctive rule. It consists of sequence of tests, means for
output result or finally obtained result the tests applied for each input. This type of iteration can be done until
first test is applicable i.e. first test become classified as true or false, or negative or positive, assumes n Boolean
attributes to be considered, we denote the set of such variables as:
Vn= {x1, x2, xn} [4]
. Decision List was selected to be master approach in our model Master-slave technique.
2.2. Naïve Bayes
An advantage to use Naïve Bayes is that it requires data with, very small size to estimate the parameters
necessary for classification. The technique is based on Bayesian theorem.
Given set of variable , we construct the probability for the event cj from the set of
outcomes, . Her P is predictor and C is set of categorical levels, which act as dependent
variable. Using Byes rule
Where:
That means p belongs to cj. We can use Maximum A posteriori (MAP). Mainly is also known as decent
classifier, so the probability outputs from predict- probability are not be taken too seriously.
Naïve Bayes, indicates as a strong independence assumptions between the features. Naïve Bayes
requiring number of parameters, Bayes theorem is a technique for constructing classifiers. In that not a single
algorithm is used, but a combination of algorithm which based on common principle. For example- a fruit may
be considered to be orange, if it is orange, round and about 3 inch diameter, a naïve bayes classifier by
considering each of these features [5]
.
2.3. Adaboost
In some case, the weak classifier need combine in such a way to improve the accuracy and create
strong one. Ensemble method is an strategy to combine learning algorithms that have different methodology
together. An application to WSD combination method is Adaboost, which is a general approach to create and
contract a strong classifier from weak classifiers. The actual process carried out is as mentioned below(6)
.
Box (1): Adaboost Algorithm implemented
To make learning process easier members of training data are weighted equally. Adaboost Algorithm
treats it as an input. For X components, it is iterated y times one turn is allotted for each classifier. In case of
master – slave technique the algorithms are selected on the basis of the accuracy which is decision List and the
algorithms which is needs some boost in terms of accuracy are slaves, in this case Naïve bayes and Adaboost are
acting as a slaves, in such way combining improve the accuracy of these slaves.
2.4. Ensemble Methods
For ensemble methods use more can one learning algorithm to obtain predictive performance as
comparing constituent learning algorithms. Predictive performance means accuracy typically used in inductive
learners. Robustness over single estimator the original ensembles method is Bayesian averaging [7]
.
Some method for constructing ensembles manipulates the example to finalize multiple hypotheses such as:
 Manipulate the set of input features.
 Manipulate the output, to obtain the good ensemble of classifiers for obtaining the values.
For x =1; x< m; x++)
Fetch weight αx from classifier cx
))(()(
1
XxxsignxH
y
x

 
}
Where H(x) sign is function for linear combine of
weak learner to boost the performance.

 Injecting randomness, used for generating ensembles of classifiers to inject randomness into the algorithm.
III. Experimental Setup
Experiments are conducted by using an approach to resolve word sense disambiguation. Input is
nothing but 10 nouns and 5 verbs along with WordNet repository to know POS. Innovative approach which is
based on Master-Slave model. Results are calculated on the basis of the said set up.
1. Data Set: five verbs and ten nouns are selected to perform the experiments of word sense disambiguation [8]
.
2. Data Source: WordNet 2.19]
referred to several the details related with a particular word like part of speech
(POS). This data source is used to resolve the disambiguation of various meanings related with given data
set.
3. Training Set: To train the algorithm to identify correct sense of given word context is used. This context is
in the form of snseval-3[10]
. This means with reference of the context given with respective word. Training
Phase plays important role in identification of correct meaning of a word from data set. Result of training
phase is to make disambiguation task much easier.
4. Testing Set: Thus the calculated meaning of a word is verified in this testing phase.
5. Algorithms: Algorithms are written in java [11]
which drives the meaning identification process. Master –
Slave Voting Algorithm, this is an extension of algorithm process mentioned above where two algorithms or
more are clubbed to deliver the maximum performance acting as a slave.
6. Attributes: Attributes is nothing but factors taken into the consideration for making the decision related with
word sense disambiguation. There are various stages of this attribute, means while deciding the weight
allotted for given meaning feature acts as attribute.
While dealing with Master-Slave model main task is to decide a particular algorithm as Master and
other/ others as a Slave. So in this decision making process overall accuracy or F-measure of all algorithm acts
as an attribute.
About topical and lexical context analysis, for example Suppose w-3, w-2, w-1, w, w+1, w+2, w+3 is
the context of words before and after given word w (to solve word sense disambiguation). All information
related with respective part of speech (-3≤POS≤3), and consider various combinations like (w-1, w+1), (w+1,
w+2), (w,w+1,w+2), (w-1, w, w+1, w+2, w+3), these could be many more combinations of these words
mentioned in a context. Together the set of all W, POS, is known as sample set for the attribute of given word
environment [12]
.
7. Combination algorithm applied: the steps of combination algorithm we implemented as below:
 Input – data set of 15 words is used to disambiguation a word along with data source of WordNet and
context.
 Process- java code is used to implement master- Slave model to improve the accuracy of an algorithm. Data
processing, classification and accuracy calculation is carried out by this code.
 Output-Accuracy of this master- Slave model is decided by precision, recall and f-measure. Box1. Below
shows the steps of combination algorithm implemented.
Box (2): Master-Slave combination Algorithm implemented steps
Master – Slave model deals with combination of algorithms to improve the result. This combination helps to
increase the performance of an algorithm by boosting the accuracy of given algorithm. Algorithm is designed to
implement Master-Slave technique to improve the performance of Naïve Bayes and Adaboost algorithms.
Step1. Accuracy of Master X % is collected.
Step2. Accuracy of Slave y %
Step3. Collect voting to improve X by using factor
F= (X - f)/100.
Step4. Accuracy of Word=old Accuracy + F
Step5. Apply this factor for all words, X1, X2,
X3…, and X15.
Step6. Calculate precision, Recall, and f-measure.

IV. Methodology
Master – Slave model deals with combination of algorithm to improve the result. This combination helps to
increase the performance of an algorithm by boosting the accuracy of given algorithm.
To select Master and Slave experiment is conducted. After conducting the experiment and performing the
necessary literature survey related with it following options or approaches are considered.
1.Select the Master, generally an algorithm with high accuracy, good history and utilization.
2.Combine existing algorithms to improvise the accuracy.
3.Variable Factor selection: The Master- Slave Architecture adds a factor to boost the performance of system if
this factor is designed an fixed format, the value to be added will not be added differently for different words.
How to treat a word with accuracy=100%.
In the factor that we add is decided . Where x is the accuracy of an algorithm which is
lagging and all the time we add to ensure addition in the accuracy. Hence all the time performance of
the algorithm will get improved by referring Master- Slave model. This model is boosting the performance
ensures the rise in the overall accuracy provided selection of adequate algorithm, with high accuracy should be
made. Java code improved algorithm, which is written in Java [13]
, improves the accuracy by using delimiter
function which is mention at step 2.3. This function will internally invoke several programs to conduct voting
and find the correct sense.
V. The Experiments
5.1 First Experiment
The first combination deals with Naïve Bayes as classifier and Decision List as Master, experiment is
conducted by considering decision list as a master and Naïve Bayes algorithm as a slave, after completing this
experiment the accuracy of Naïve Bayes model (individual) got increased. To effect was possible only due to
decision list (which is acting as a master).
Table 1: Data Set of Words and Results of Naïve Bayes and Decision list Combination
Word POS
# First Combination
Sense Recall Precision F-Measure
Praise n 2 1000 500 1500
Name n 6 1000 764 2292
Worship v 3 1000 763 2289
Worlds n 8 1000 702 2106
Lord n 3 500 500 1500
Owner n 2 500 500 1500
Recompe-nse n 2 333 333 999
Trust v 6 1000 143 429
Guide v 5 1000 1000 3000
Straight n 3 1000 1000 3000
Path n 4 473 412 1236
Anger n 3 922 500 1500
Day n 10 250 250 750
Favored v 4 167 167 501
Help v 8 125 125 375
684.6667 510.6 1531.8
This table shows the F-measure which is calculated by knowing precision and recall with the help of
following formula.
Even after implementing Master-Slave model, the accuracy is not 100%.

Fig.2. The first combinationGraph
When we look at the performance of the combination that we have selected, we can observe the
considerable hike in the performance of Naïve Bayes algorithm.
This hike could be well interpreted by looking at the table of individual contribution of the Naïve Bayes
algorithm [14]
.
5.2 Second Experiment
Experiment conducted accuracy is increased, this combination experiment deals with Adaboost as slave
classifier, to improve the accuracy more and more. Experiment is conducted and it is observed that this
combination gives better result.
Table 2: Data Set of Words and Results of Adaboosts and Decision list Combination
Word POS
# Second Combination
Praise n 2 899 1000 3000
Name n 6 1000 1000 3000
Worship v 3 996 1000 3000
Worlds n 8 141 1000 3000
Lord n 3 465 1000 3000
Owner n 2 942 1000 3000
Recompe-nse n 2 963 1000 3000
Trust v 6 167 167 501
Guide v 5 500 510 1530
Straight n 3 500 500 1500
Path n 4 333 333 999
Anger n 3 500 500 1500
Day n 10 111 1000 3000
Favored v 4 250 250 750
Help v 8 125 125 375
526.1333 692.3333 2077
This table shows the F-measure which is calculated by knowing precision and recall, and below the results
graph of the combination between Adaboost and decision list.
Fig.3. The second combinationGraph
0
1000
2000
3000
4000
5000
6000
1 3 5 7 9 11 13 15
First
combina
tion F-
Measure
First
combina
tion
Precision
First
combina
tion
Recall
0
1000
2000
3000
4000
5000
6000
1 3 5 7 9 11 13 15
Second
Combination
F-Measure
Second
Combination
Precision
Second
Combination
Recall

5.3 Third Experiment
Now after two experiments above, we combined the three approaches Naïve Bayes and Adaboost as
slaves with master approach which is Decision list, and as per the anticipation highest accuracy is received.
Table 3: Data Set of Words and Results of Naïve Bayes, Adaboosts and Decision list Combination
Word POS
# Third Combination
Praise n 2 771 1000 3000
Name n 6 1000 1000 3000
Worship v 3 494 676 2028
Worlds n 8 142 1000 3000
Lord n 3 483 1000 3000
Owner n 2 848 1000 3000
Recompe-nse n 2 882 1000 3000
Trust v 6 167 167 501
Guide v 5 500 971 2913
Straight n 3 500 500 1500
Path n 4 333 333 999
anger n 3 500 500 1500
Day n 10 111 1000 3000
Favored v 4 250 250 750
Help v 8 125 125 375
473.7333 701.4667 2104.4
Fig.4. The third combination Graph
VI. Comparison Approaches of Master – Slave model combination
By looking to the graphs (2, 3, 4), and make Comparative analysis of three experiments of Master-
Slave model to observe rise in the performance of Naïve Bayes algorithm. So this model gives hike in the
individual performance of second and third combination experiments. The graph below
Fig. 5 Master-Slave thechinue
The Comperative Combination Recall Graph
0
1000
2000
3000
4000
5000
6000
1 3 5 7 9 11 13 15
Third
Experiment
Third
Combination
F-Measure
Third
Experiment
Third
Combination
Precision
Third
Experiment
Third
Combination
Recall
0
200
400
600
800
1000
1200
Praise
Name
Worship
Worlds
Lord
Owner
Recompe-nse
Trust
Guide
Straight
Path
anger
Day
Favored
Help
COMPARATIVE ANALYSIS OF RECALL
1st
Experiment
Recall
2nd
Experiment
Recall
3rd
Experiment
Recall

And table (7) at end of paper shows the comparative results of Master- Slave technique.
Fig.6. The Comperative Combination precision Graph
Fig.7. The Comperative Combination f-measure Graph
VII. Conclusion
In this paper, we presented Master- Slave technique suggested, in the first experiment, Decision list acts a
Master and Naïve Bayes act as slave. Individually each algorithm gives good values of precision and f-measure.
When they are combined together recall is enhanced which might be useful application like search engine which
requires more coverage of sample space, but word sense disambiguation it is less useful.
In the second experiment, we Decision list as a master and Adaboost as a slave. There is increase in
precision by (1.0733) and f-measure (3.2). Unlike to the first experiment recall is decreased. This is
enhancement in precision to resolve word sense disambiguation problem.
In the third experiment combination, the decision list as master, call the Naïve Bayes and Adaboost
together. It is observed that there in increases in precision and f-measure by (48.7367) And (146.2) respectively,
this combination gives all round performance for precision.
At final the Master – Slave technique worked well to increase the performance of Slave algorithms by
boosting the accuracy of the algorithms. Type of Slave, context will play very crucial role in the growth of f-
measure. These experiments motivate to consider number of Slaves and type of Slaves carefully to make the
disambiguation process more and more accuracy. Since the emphasis is more on precision and f-measure effort
are not highlighted in the values of recall.
Table 4: The Results of three approaches before combination
No
.
Approach Before Combination
Recall Precision F- measure
1 N.Bayes 305.73 628.6 1885.8
2 D. List 440.33 691.26 2073.8
3 Adaboost 459.2 652.73 1958.2
0
500
1000
1500
2000
2500
3000
COMPARATIVE ANALYSIS OF PRECISION
3rd
Experiment
Precision
2nd
Experiment
Precision
1st
Experiment
Precision
0
2000
4000
6000
8000
10000
Praise
Name
Worship
Worlds
Lord
Owner
Recompe-nse
Trust
Guide
Straight
Path
anger
Day
Favored
Help
COMPARATIVE ANALYSIS OF F-MEASURE
3rd
Experi
ment
F-
Measu
re
2nd
Experi
ment
F-
Measu
re

Table 5: The Results of three approaches after combination
Approach After Combination
1st
Experiment
(N.Bayes +
D.L) 684.6667 510.6 1531.8
2nd
Experiment
(D.L+ Ada)
526.1333 692.3333 2077
3rd
Experiment
(N.Bayes +
Ada +D.L) 473.7333 701.4667 2104.4
The table below shows the final improvement on supervised approaches we implemented.
Table 6: The enhancement combination achieved
Approach Enhancement
1st
Experiment (N.Bayes +
D.L)
378.9367 -118 -354
2nd
Experiment (D.L+ Ada) 85.8033 1.0733 3.2
3rd
Experiment (N.Bayes +
Ada +D.L)
14.5333 48.7367 146.2
Acknowledgment
I would like to thank the ministry of higher education/Iraq; I will never forget for ever my research
guide respected Dr. Shashank Joshi (Professor at Bharati Vidyapeeth University, College of Engineering) for
his cooperation to inspire me always.
References
[1]. Nitin Indurkhya and Fred J. Damerau “HANDBOOK OF NATURAL LANGUAGE PROCESSING” SECOND EDITION. Chapman
& Hall/CRC, USA, 2010.(book style).
[2]. Boshra F. Zopon AL_Bayaty, Dr. Shashank Joshi, Conceptualisation of Knowledge Discovery from Web Search, Bharati Vidyapeeth
University, International Journal of Scientific & Engineering Research, Volume 5, Issue 2, February-2014, pages 1246- 1248.
[3]. Boshra F. Zopon AL_Bayaty, Dr. Shashank Joshi, Decision List Master Approach to Solve WSD Problem, INTERNATIONAL
JOURNAL FOR ADVANCE RESEARCH IN ENGINEERING AND TECHNOLOGY, Webpage: www.ijaret.org Volume 3, III,
March 2015 ISSN 2320-6802,pages.29-33.
[4]. Arindam Chatterjee, Roll No: 09305905, Under the guidance of Prof. Pushpak Bhattacharyya, Word Sense Disambiguation: Literature
Survey, Chapter 1, page 1-2, Department of Computer Science and Engineering, Indian Institute of Technology, Bombay, 2012.
[5]. Boshra F. Zopon AL_Bayaty, Shashank Joshi, , Bharati Vidyapeeth University, Comparative Analysis between Naïve Bayes
Algorithm and Decision Tree to Solve WSD Using Empirical Approach, International Conference on Knowledge and Software
Engineering, http://www.lnse.org, LNSE 2016 Vol.4( 1), 2016, 82-86.
[6]. Boshra F. Zopon AL_Bayaty, Shashank Joshi, Empirically implementation Adaboost to Solve Ambiguity, paper accepted in “National
Conference on, Modelling, Optimization and Control, 4th
-6th
March 2015, NCMOC – 2015”
[7]. Navigli, R. 2009.Word sense disambiguation: A survey. ACM Compute. Survey. 41, 2, Article 10), 69 pages DOI =
10.1145/1459352.1459355, February 2009.
[8]. http://www.e-quran.com/language/english
[9]. http://wordnet.princeton.edu.
[10]. http://www.senseval.org/senseval3.
[11]. Steve Holzner, Programming Java Application, Eclipse, Coverage of 3.0, Third Indian Reprint: May 2007.
[12]. Gerard Escudero, Lluis Marquez, and German Rigau, An Empirical Study of the Domain Dependence of Supervised Word Sense
Disambiguation System, paper available in http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.13.5717.
[13]. Patrick Niemeyer and Jonathan Knudsen, Learning Java, O’REILLY, Second Edition, 2002.
[14]. Boshra F. Zopon AL_Bayaty, Shashank Joshi, Empirical Implementation Naive Bayes Classifier for WSD Using WordNet., Bharati
Vidyapeeth University, international journal of computer engineering & technology (IJCET), ISSN 0976 – 6367(Print), ISSN 0976 –
6375, Volume 5, Issue 8, August (2014), pp. 25-31,© IAEME: ww.iaeme.com/IJCET.asp, Journal Impact Factor (2014): 8.5328
(Calculated by GISI), [Online]. Available: www.jifactor.com.
[15]. http://www.lnse.org, LNSE 2016 Vol.4 (1), 2016, 82-86.
Table 7: The final Comparative Result of Master- Slave Technique
Word
PO
S
# First Combination Second Combination Third Combination
Sens
e Recall
Precisio
n
F-
Measur
e Recall
Precisio
n
F-
Measur
e Recall
Precisio
n
F-
Measur
e
Praise n 2 1000 500 1500 899 1000 3000 771 1000 3000
Name n 6 1000 764 2292 1000 1000 3000 1000 1000 3000
Worship v 3 1000 763 2289 996 1000 3000 494 676 2028
Worlds n 8 1000 702 2106 141 1000 3000 142 1000 3000
Lord n 3 500 500 1500 465 1000 3000 483 1000 3000

Owner n 2 500 500 1500 942 1000 3000 848 1000 3000
Recompe
-nse n 2 333 333 999 963 1000 3000 882 1000 3000
Trust v 6 1000 143 429 167 167 501 167 167 501
Guide v 5 1000 1000 3000 500 510 1530 500 971 2913
Straight n 3 1000 1000 3000 500 500 1500 500 500 1500
Path n 4 473 412 1236 333 333 999 333 333 999
anger n 3 922 500 1500 500 500 1500 500 500 1500
Day n 10 250 250 750 111 1000 3000 111 1000 3000
Favored v 4 167 167 501 250 250 750 250 250 750
Help v 8 125 125 375 125 125 375 125 125 375
473.733
3
701.466
7 2104.4
526.133
3
692.333
3 2077
473.733
3
701.466
7 2104.4

Supervised WSD Using Master- Slave Voting Technique

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to Supervised WSD Using Master- Slave Voting Technique (20)

More from iosrjce (20)

Recently uploaded (20)

Supervised WSD Using Master- Slave Voting Technique