# I. Introduction

eb mining is facing an important problem in measuring the semantic similarity among the words in the process of information retrieval and language processing. The most semantic based application requires the accurate measuring of semantic similarity among the document concepts and words. In information search, one of the most important problems is to semantically to get a number of documents correlated to a user's request. Semantic similarity between the words such as "word sense disambiguation" (WSD) can be an efficient assessment for the text entailment and automatic document classification, it is also important for the variety of natural language processing tasks. Automatic classification of documents is an important part of the research in the vision, and an enormous prospective for numerous applications around the text, such as search and analysis. Its purpose is to allocate a document given to the group of default to which it is in the right places. So far, applications have different types of algorithms based on the study or automatic calculation in this process and showed how much work [2], [3], [5]. However, mainly of the work functional to this task used aneffortless wordcollection representation where each attribute communicates to a particular word. That is, assume that words are independent and utilize only the distribution of content words.

Over the past few years, we've seen the Web evolve into a semantic Web. The amount of information posted with linked data has consistently increased. With this increase, annotation and classification systems have created new opportunities to reuse this data as a semantic knowledge base and can be interconnected and structured to increase the accuracy and recovery of annotation and classification mechanisms. The Semantic web aims to explain the meaning of the information posted on the Web in order to make it possible to search by understanding the meaning of the information accurately. In this regard, document text learning and classification is most common, by assigning text to one or more existing class. This development determines the class membership of a text document that has a separate set of classes with profiles and different features. Criteria for deciding appropriate features for classification are important and are determined by the priority of the classifier. Semantic classification occurs when the target document element or term of the classification represents the meaning of the document.

Measuring the semantic similarity among texts is a basic task and can be capable of being utilized for a variety of applications, together with "text clustering" [1] and "text classification" [2]. The challenge in evaluation similarities among texts is infrequent, that is, there will be no coincidence of terms between the two texts. For example, two texts "Apple's New Product" and "iPhone-6" refer to related topics, even though they do not use similar terms.

To overcome scarcity, we need to use external data or knowledge to enrich the semantic representation of text. The semantically associated words of a particular word are listed in amanually created universal dictionary vocabulary ontology such as "Word Net". In this, a synset includes a set of synonyms for a specific word sense. However, semantic similarities among individual transform more than time and across domains. For example, apples are often associated with computers on the web. However, this apple sensation is not listed in most universal thesauri or dictionaries. Users searching for apples on the web may be concerned in the meaning of "apple" and "not apple" as a fruit. Innovative words are stably generated and new senses are dispensed to existing words. Preserving ontology manually to confined these innovative words and senses is costly, if not impracticable.

In this paper, we contribute an automated semantic similarity learning (SSL) move towards to compute the probability of semantic similarity among terms or entities of documents with the class knowledge set entities. Here, we define two probabilistic scores, a semantic similarity (SS) score and N-grams pair similarity (GS) score enhancing Naive Bayes probabilistic method to aggregate the relation between document and class entities. Semantic Similarity method relates the trained class entities terms with the extracted document key terms to compute the document probable SS score against each class entities, and Ngrams pair similarity method relate a document with each trained class entity with the constructed N-grams pairs, which is constructed using most frequent terms extracted from the document and the probable GS score is the summation of all individual N-grams pairs, i.e., sum(GS 1 , GS 2 , .. , GS n ).We perform an experiment evaluation on Reuters-21578 Datasets to demonstrate the effectiveness of the proposal.

This papers organized in 6 sections. Section-1 above describes the introduction, section-2 discuss the background works, section-3 presents the proposed works outline, probabilistic semantic and N-gram pairs pattern learning, section-4 discuss the semantic similarity classification approach, section-5 present experiment methodology and results and finally section-6 presents the conclusion of the work.


# II. Background Study

Semantic similarity plays an significant responsibility in "natural language processing", "information retrieval", "text summarization", "text classification", and "text clustering". Particularly, "Explicit Semantic Analysis" (ESA) [6] is extensively utilized because of its accessibility and diversity. ESA was build up to calculate word relationship as well as text comparison in natural language. ESA creates a "weighted index" that maps each phrase to the listing of articles that appears and calculates the similarity among the two words or a vector of text.

Naive Bayes [1] classification performance using semantic similarity has made various efforts. An approach that is often used to mitigate naive independent assumptions is to express attribute addiction in a graph-based model called a "Bayesian network", where nodes correspond to attributes. Oriented arch is weighted by the circumstances probability for each node specified a close relation. Because "Bayesian network learning" is NP-hard [6], numerous approaches recommend imposing model constraints to formulate it easier to deal with learning problems. Subsequent approaches in [8], [9], [17], [18] have brought considerable improvements. For example, in [21], an ensemble of Tree Augmented Naive-Bayes (TANs) was be trained, each rooted in a dissimilar attribute. It then compiles the classifications of all eligible TANs to predict class labels. In [8], we assume that the entire Bayesian network structure is learned first and all attributes are dependent. Unlike [18], the "Markov network model" is utilized to express characteristic dependencies that are estimated similar to [39] by taking advantage of the conditional log probability intention purpose. However, performing andtake advantage operation can be computationally demanding. Many methods are used to inherit the structural simplicity of Naive Bayes classifiers to keep away from the complication of the construction learning process [9], [10], [12], [13]. While the "Naive Bayes classification" is functional at the "decision tree leaves level" and is act upon on a subset of the training data, the data set properties are divided into two collections as in [11],where one group is assigned a class probability based on "Naive Bayes", and the other is supported on a "decision table".

Despite its effortlessness, the previously point out the classifier still shows a few constraints in handling very much related data. In [12], [13], the features are weighted dissimilarly depending on the involvement to the classification. A comparable approach was applied to the most effective "Bayesian Network classifier" and " Hidden Naive Bayes" [9]. In [9], the authors recommended generating a hidden close qualified that correspond to the effect of everything else on each property. The effect is computed as a linear arrangement of circumstance common information among attribute pairs, similar to [8]. Therefore, the parent correlation is ignored.

Dissimilarity like [10], [11], [19], "En Bay" [2] implements a new, uncomplicated, and useful approach that unites the generation of conditionally independent decision models and the reliable probability approximation by class. En Bay is a pattern-based Bayesian classifier that frequently uses a set of items frequently to estimate Bayesian probabilities. En Bay uses new and effective probabilistic approximation estimates that adhere to the conditional independence model. The set of extended, normal and separate items to be comprised in a class-based approximation is chosen by entropy-based heuristics, and the set of properties is conditionally mutually independent, depending on the class being evaluated. We extend En Bay probability computation methods to computes the semantic similarity probability score based on the terms dependency over the trained class terms entities, as discussed in section 3.2 below.

The "Large Bayes classifier" [11] performed the primary challenge to mitigate well-built independent 3 Year 2017 ( ) H to estimate the probability through product form approximation [12]. However, all preceding patternbased Bayesian advance create inimitable product approximations for all test cases. Thus, estimations are only tied to the considered grade. Moreover, since it is necessary to extract animmense number of long and redundant repeated item sets, the superiority of the approximation is sensitive to changes in the "support threshold", and the classification algorithm cannot cope with a large data set. We extend this constructing Ngrams pairs using frequent items to estimates the N-grams similarity probability score, as discussed in section 3.3 below.


# III. Proposed Approach a) Outline

The Semantic Similarity Learning (SSL) method, which uses probabilistic performances to describe probabilistic scores and put together scores supported on Bayes' method for accurate document classification to measure the robust discovery and semantic similarity of related entities to document.  1. outlines our the proposed approach method. Our method obtains the main points of the probabilistic analysis of associations and related documents on the basis of trained document entities. The approach performs two probabilistic score computation method. First, Semantic Similarity Method which measures the similarity of their associated entity of a document with the list of trained class entity terms by SS k = P(d k (t) | C m ), i.e., probability of a document d k (t)terms associated with a set of class C m terms by means of cosine similarity.

The second method extracts the most frequent terms F from the extracted document terms using term frequency (tf) and using F we construct N-grams pairs. In general, an N-gram method slice a longer text into ncharacters, but we customized this to slice a pattern into number words pairs (V-Pair) based on n which we term as N-gram pattern, an illustration is shown in Fig. 2.Using the constructed pairs we compute,
???? ?? = ? ?? ?? ?? ??=1
, where n is the number of pairs and W i = P( V-Pair n | C m ) i.e., probability of N-gram pair terms related to the set of class C m terms using cosine similarity. Now, we compute the final probability of semantic similarity P sem = sum (SS k, GS k ) for each document against each trained class. To classify the document we find the max P sem among the computed probability of semantic similarity of each class. The class which has the max P sem will be considered as the document class. We describe each method mechanism in the aspect in the following sections. [1], "Decision Trees" [2], "SVMs" [3], "Rule-based" [4], and "Associative classifiers" [5].


# Unlabeled Document
Keyword
Bayesian classification methods recognized a classification supported on the of "Bayes theorem" [1].It predicted that a class based on test documents previously un seen T = { a 1 , a 2 , . . , a n }by opting the class c i that make the most of the subsequent formula: Despite the simplicity, the Bayesian approach is calculation intractable without compelling a powerful model simplification [1], [6], [7]. The most important instance of simplification is the "Naive Bayes classifier" [1], which solves the problem by assuming that all attributes are conditionally self-determined and given as the class c i . Therefore, the join probability of (1), is based on the generated Naive Bayes model, which can be approximated as, (2) Based on the approximation we combine the probabilistic semantic similarity (SS) scores extracted from the training data to find the appropriate entities for the document. Let's assume that multiple key terms are entered as input. That is, we compute P (c|T ')for the set of core key terms T' ={t 1 , t 2 , ... , t k }, where T' is a key term, which are derived using traditional Naive Bayes for any related class c i .

One possible approach to this task is a twostep method of determining the key terms first and then applying the existing Naive Bayes. However, this approach raises the question of how key terms are established. We have developed a probabilistic similarity method for finding related entities. It can be functional to a set with probability determined members. 
F = {n = 3 n = 2 n = 1 Year 2017 ( )

# H

For a particular, a set of key terms T, P (c|T' ) is calculated for all probable states T'.Fig. 3,summarizes an illustration of the probabilistic semantic similarity method for a set of key terms of a document d k as t 1 , . . , t k . SS method is utilized to calculate P(t K |C m ), which is the probability score SS k of the set of key terms, T for the class C m .


# Figure 3: Probabilistic Semantic Similarity Method

Unfortunately, the circumstances of selfdetermination supposition prepared by Naive Bayes may not true always, to obelieves high-order associations for the period of the probability estimate, a parallel proposal is made based frequent pattern learning in T, and constructing a N-grams pairs to support accurate classification.


# c) N-grams Pattern Learning

The N-gram is defined as a sequence of terms, the length is n, and the words taken are called terms. In the literature, we can see the definition of an N-gram as a concurrent set of terms, but only consecutive term sequences were used in this study. One word in the document is represented by a set of overlapping Ngrams as shown in Fig. 2. The N-gram model can be fictional by introduction a small window over a sentence or text, where only n words can be seen at the same time. So the effort less N-gram model is the so-called "unigram model". This is a one-word model at a time. For example, the "Latest application and iPhone released." sentence contains five unigrams as, "Latest", "application", "and", "iPhone" and "released" Of course, this is not very beneficial information. It is just a word that makes up the sentence. In fact, N-grams are interesting when n is greater than 2 (bigram) or more.

Each word happens in a document with a dissimilar frequency. The main thought of categorization utilized by Trenkle and Cavnar [5] is that they should have similar N-gram frequency distributions when comparing documents of the same category. We performanN-gram pattern learning through creating n pairs using frequent document terms.

For a given document d having a T terms. Let's assume the frequent terms represent as F. Using the F terms we construct N-grams pairs as V-Pair. To learn the probability of V-pair pattern associationW n of a document with a class c m we calculate P(V-Pair n |c m )as shown in Fig . 3. Here, the class must contain all the pair terms to match the association. To compute the Ngrams probable similarity GS, we done the summation of all W n as, In order to efficiently search for class-specific patterns, in the SSL training phase, an FP-growth data pattern representation [18] is separately created to store training data belonging to each class in a compressed form. The FP-growth pattern is a single-tree data structure for the minimum support (min_sup) frequent item set used in the class pattern learning context. Algorithm-1 shows the pseudo-code in the SSL training phase. Minimum supported thresholds were applied to remove infrequently used items. In this case, items that do not meet the necessitated minimum support threshold are not consisted of in the FP-growth pattern.
d k C 1 C 2 C m . . . .
The obtained FP-G of each individual class c i , will be used as a trained knowledge for the SSL classifier.


# b) SSL Classification

The SSL classification approach is one of the accomplished algorithms managing unlabeled documents. It applies two probability computation as semantic similarity SS and N-gram Similarity G Son the dataset to perform the classification using the trained FP-growth pattern knowledge as shown in Fig. 1. The SSL classifier initialized with anonly some trained class item sets. At each iteration, it chose an unlabeled document and perform the computation to compute the SS and GS score. It learns separate similarity score over each class pattern learning, and support a set of class labels for the unlabeled documents. For each class c i belonging to the training data set is corresponding to FP-growth is visited to construct the class-centric product estimation and calculated the probability P(T,c i ).

Its ultimate prediction is through by coming together both SS and GS score, P sem = sum( SS k , GS k ) which decline classification error spredictions. The massive is the P sem of the class will be predicated as the
d k C 1 C 2 C m . . . . . . W 2 = P( V-Pair 2 |C 1 ) W 1 = P( V-Pair 1 |C 1 ) W n = P( V-Pair n |C 1 ) . . . . . . . . W 2 = P( V-Pair 2 |C 2 ) W 1 = P( V-Pair 1 |C 2 ) W n = P( V-Pair n |C 2 ) . . . . . . . . W 2 = P( V-Pair 2 |C m ) W 1 = P( V-Pair 1 |C m ) W n = P( V-Pair n |C m ) . . . . . . . . ???? 1 = ? ?? ?? ?? ??=1 ???? 2 = ? ?? ?? ?? ??=1 ???? ?? = ? ?? ?? ?? ??=1 Year 2017 ( )

# H

closer association. This predictionim provises the performance of the algorithm classification accuracy. Since SSL classification uses two probability computation score with the FP-growth pattern, its presentation is better than any particular classifier.Algorithm-2briefly summarizes the SSL classification algorithm.  The "Reuters-21578 corpus" is the mainly common utilized benchmark corpus in text classification. It consists of over 20,000 Reuters news stories from 1987 to 1991, and 135 subject classes are used in the experiment. This version contains "9603 training documents", "3299 test documents", and "27,863 inimitable words" after stopping stemming and word removal. We consider only 10 topics as classes of Reuters-21578 data for experimental evaluation measurements.
d k = {T k } ?D k ; //--For all class in FP-G vector -- for all c i in FP-Gdo c i = {FT i } ?FP-G i ; SS i = P (d k ? c i ); V_SS[i] = SS i ; end for VD s [k] = V_SS; // --

# b) Performance Measure

To estimate the classification performance of the proposed method, we utilize the precision, recall, and accuracy. Let considered P is all relevant documents and N is all negative document. PC + as a positively classified, NC + as negatively classified documents. PCas a positively classified for an incorrect document, NCas negatively classified for correct documents. By constructing a confusion matrix for the above evaluation measure we compute the classifier performance.

To measure the classifier precision rate CP, the classifier recall rate CR and the classifier accuracy rate CA the following equation are used.


# c) Evaluation Results

In the Reuters-21578 datasets we do consider both labeled and unlabeled documents, the effect of using two probabilistic semantic similarity learning is given in Table 1. We initially evaluate with Semantic Similarity Score (SS), then with N-grams patterns pairs Score (GS) and finally with both. The classification performance using both the Semantic Similarity and the N-gram pattern pairs learning outperforms over the one using any single learning for most classes.
???? = ???? + ???? + + ???? +(4)???? = ???? + ???? + + ???? ?(5)???? = ???? + + ???? + ?? + ??(6)

# Global Journal of Computer Science and Technology

Volume XVII Issue II Version I
8 Year 2017 ( ) H Table 1:
The accuracy enhancement by using semantic similarity learning on "Reuters-21578 corpus".

We found that the greater the number of related documents in the training set, the higher the accuracy of using N-Naive Bayes has a low error rate and high accuracy when there are many documents in the class. The classification comparison result is shown in Fig. 4. At first, we performed comparisons with stateof-the-art Bayesian classifiers. And because our approach is pattern-based, we compare it with the wellknown associative classifiers SVM and the new improved Bayesian approach known as En Bays [2].

Finally, we performed a comparative assessment of precision, recall, and accuracy rates as a classifier for classifiers.   The rate of precision and recall in Fig. 5 and 6 shows an improvisation in compared to SVM and En Bays method. The effects of both SS and GS score in probability similarity measure shows SSL precision improvisation. Fig. 7 shows the classifier accuracy measures comparison. It also shows an improvisation of SSL approach in compare to others. The falling of accuracy with increasing of the document due to the limitation of trained class knowledge. As both the method has a dependency on the trained data knowledge for performing probability similarity computation cause the falling of the rate.


# VI. Conclusion

In this paper, we propose a semantic similarity and N-gram pattern learning method based on the Bayesian classifier, which approximates Bayesian probability using frequent itemsets. It utilized new and more efficient probability approximations that adhere to the conditional independence model. A long, frequent, and separate set of items to be included in a classbased approximation is selected. It is based on the Baye's theorem and semantic similarity computation approach. Our method is a sort of probabilistic semantic similarity learning (SSL) that uses vectors to generate vectors of related entities as semantic representations of specific text and to measure semantic similarities. SSL combines vectors using expanded Naive Bayes, while SSL simply adds up the vectors for each term occurring in the text based on the majority of rules. This method uses both Semantic Similarity Learning for SSL algorithms and N-gram pattern learning and applies algorithms to unstructured document classification.

Experiments on the Reuters-21578 document show that the SSL approach improves classification performance, and unlabeled documents are a good resource to overcome documents with a limited number of labels.

Future developments in this work will address the integration of generalized item aggregation mining algorithms to further improve classification and accuracy in noise-prone areas of data where there liability of probability estimation is particularly important.
1![Figure 1: Outline of the proposed Approach Fig.1. outlines our the proposed approach method. Our method obtains the main points of the probabilistic analysis of associations and related documents on the basis of trained document entities. The approach performs two probabilistic score computation method. First, Semantic Similarity Method which measures the similarity of their associated entity of a document with the list of trained class entity terms by SS k = P(d k (t) | C m ), i.e., probability of a document d k (t)terms associated with a set of class C m terms by means of cosine similarity.The second method extracts the most frequent terms F from the extracted document terms using term frequency (tf) and using F we construct N-grams pairs. In general, an N-gram method slice a longer text into ncharacters, but we customized this to slice a pattern into number words pairs (V-Pair) based on n which we term as N-gram pattern, an illustration is shown in Fig.2.Using the constructed pairs we compute, ???? ?? = ? ?? ??](image-2.png "Figure 1 :")
2![Figure 2: An illustration of N-gram pairs](image-3.png "Figure 2 :")
1![Where P (T| c i )indicates the provisional possibility of the test document T of a given class c i . Probability is approximate commencing from the training set. Since classification focuses on choosing the class that takes advantage based on the equation (1), relatively than assigning an unambiguous probability to each class, denominator P(T) in (1) can be misplaced because it does not influence the comparative class instruct.](image-4.png "( 1 )")






if n = 3;1-grams2-grams3-gramsapplication(application,(application, software, screen)softwaresoftware)(application, software, phone)screen(application, screen)(application, software, iphone)phone(application, phone)(software, screen, phone)iphone(application, iphone)(software, screen, iphone)(software, screen)(screen, phone, iphone)(software, phone)(software, iphone)(screen, phone)(screen, iphone)
			© 20 7 Global Journa ls Inc. (US)
			© 2017 Global Journals Inc. (US)
		
		
			

				


* 
	
		Wikipedia-Based Semantic Similarity Measurements for Noisy Short Texts Using Extended Naive Bayes
		
			MasumiShirakawa
		
		
			KotaroNakayama
		
		
			TakahiroHara
		
		
			ShojiroNishio
		
	
	
		IEEE Transactions On Emerging Topics In Computing
		
			3
			2
			June 2015
		
	




* 
	
		En Bay: A Novel Pattern-Based Bayesian Classifier
		
			ElenaBaralis
		
		
			LucaCagliero
		
		
			PaoloGarza
		
	
	
		IEEE Transactions On Knowledge And Data Engineering
		
			25
			12
			December 2013
		
	




* 
	
		Learning Bayesian Network Classifiers by Maximizing Conditional Likelihood
		
			DGrossman
		
		
			PDomingos
		
		10.1145/1015330.1015339
		
	
	
		Proc. 21st Int'l Conf. Machine Learning (ICML '04)
				21st Int'l Conf. Machine Learning (ICML '04)
		
			2004
			46
		
	




* 
	
		Scaling up the Accuracy of Naive Bayes Classifiers: A Decision-Tree Hybrid
		
			RKohavi
		
	
	
		Proc. Second Int'l Conf. Knowledge Discovery Data Mining (KDD '96)
				Second Int'l Conf. Knowledge Discovery Data Mining (KDD '96)
		
			1996
			
		
	




* 
	
		Combining Naive Bayes and Decision Tables
		
			MHall
		
		
			EFrank
		
	
	
		Proc. 21st Int'l Florida Artificial
				21st Int'l Florida Artificial
		
	




* 
	
		
	
	
		Intelligence Research Soc. Conf
		
			
			2008
		
	




* 
	
		Weighted Naive Bayes Modelling for Data Mining
		
			JT A SFerreira
		
		
			DG TDenison
		
		
			DJHand
		
		
			2001
		
	




* 
	
		A Decision Tree-Based Attribute Weighting Filter for Naive Bayes
		
			MHall
		
		XXIII, M. Bramer, F. Coenen, and A. Tuson
		
			2007
			Springer
			
		
	
	Research and Development in Intelligent Systems




* 
	
		Mining Association Rules between Sets of Items in Large Databases
		
			RAgrawal
		
		
			TImielinski
		
		
			ASwami
		
	
	
		ACM SIGMOD Record
		
			22
			
			1993
		
	




* 
	
		Axiomatic Derivation of the Principle of Maximum Entropy and the Principle of Minimum Cross Entropy
		
			RJohnson
		
		
			JShore
		
	
	
		IEEE Trans. Information Theory
		
			26
			1
			
			Jan. 1980
		
	




* 
	
		Support Driven Opportunistic Aggregation for Generalized Itemset Extraction
		
			EBaralis
		
		
			LCagliero
		
		
			TCerquitelli
		
		
			VD' Elia
		
		
			PGarza
		
	
	
		Proc. Fifth Int'l Conf. Intelligent Systems
				Fifth Int'l Conf. Intelligent Systems
		
			2010
		
	




* 
	
		Clustering short texts using Wikipedia
		
			SBanerjee
		
		
			KRamanathan
		
		
			AGupta
		
	
	
		Proc. Int. ACM SIGIR Conf. Res. Develop. Inf. Retr. (SIGIR)
				Int. ACM SIGIR Conf. Res. Develop. Inf. Retr. (SIGIR)
		
			Jul. 2007
			
		
	




* 
	
		Towards effective short text deep classification
		
			XSun
		
		
			HWang
		
		
			YYu
		
	
	
		Proc. Int. ACM SIGIR Conf. Res. Develop. Inf. Retr. (SIGIR)
				Int. ACM SIGIR Conf. Res. Develop. Inf. Retr. (SIGIR)
		
			Jul. 2011
			
		
	




* 
	
		Case and Relation (CARE) based Page Rank Algorithm for Semantic Web Search Engines
		
			MsNPreethi
		
		
			Devi
		
		
			DrT
		
	
	
		IJCSI International Journal of Computer Science Issues
		
			9
			1
			May 2012
		
	




* 
	
		The semantic web
		
			TBLee
		
		
			JHendler
		
		
			OLassila
		
	
	
		Scientific American
		
			284
			5
			May 2001
		
	




* 
	
		EIIS: an educational information intelligent search engine supported by semantic services
		
			Ch.-QinHuan
		
		
			Ru-LinDuan
		
		
			YTang
		
		
			Zhi-TingZhu
		
		
			Y.-JianYan
		
		
			Yu-QingGuo
		
	
	
		international Journal of Distance Education Technologies
		
			January 1, 2011
		
	




* 
	
		Web Page Indexing through Page Ranking for Effective Semantic Search
		
			RobinSharma
		
		
			AnkitaKandpa
		
		
			PriyankaBhakuni
		
		
			RashmiChauhan
		
		
			RHGoudar
		
		
			AsitTyagi
		
	
	
		Proceedings of7'h International Conference on Intelligent Systems and Control
				7'h International Conference on Intelligent Systems and Control
		
			2013
		
	




* 
	
		A Cluster-based Resource Correlative Query Expansion in Distributed Information Retrieval
		
			LinYuan
		
		
			LinHongfei
		
		
			LiHe
		
	
	
		Journal of Computational Information Systems
		
			8
			1
			
			2012
		
	




* 
	
		Textual document indexing and retrieval via knowledge sources and data mining
		
			WWChu
		
		
			ZLiu
		
		
			WMao
		
	
	
		Commun. Inst. Inf. Comput. Mach. (CIICM)
		
			5
			
			2002
		
	




* 
	
		Towards an on tology for global software development
		
			AVizcaíno
		
		
			FGarcía
		
		
			ICaballero
		
		
			JCVillar
		
		
			MPiattini
		
	
	
		IET Softw
		
			6
			3
			
			2012
		
	




* 
	
		Weighted Page Rank Algorithm Based on Number of Visits of Links of Web Page
		
			NTyagi
		
		
			SSharma
		
	
	
		In International Journal of Soft Computing and Engineering (IJSCE)
		2231- 2307
		
			2
			July 2012
		
	




* 
	
		Page Ranking Algorithms: A Survey
		
			NDuhan
		
		
			AKSharma
		
		
			KKBhatia
		
	
	
		proceedings of the IEEE International Advanced Computing Conference (IACC)
				the IEEE International Advanced Computing Conference (IACC)
		
			2009
		
	




* 
	
		Ontology Based Information Retrieval in Semantic Web: A Survey
		
			DrMayankVishal Jain
		
		
			Singh
		
	
	
		I.J. Information Technology and Computer Science
		
			10
			
			2013