Using Cluster Analysis for Author Classification of Albanian Texts: A

Study on the Effectiveness of Stop Words

DENISA KAÇORRI , ALBINA BASHOLLI , LUELA PRIFTI

Department of Mathematical Engineering,

ALBANIA

Abstract: - Cluster analysis is a statistical approach that identifies uniform clusters within data. The closeness of

data is measured quantitatively using distance functions. Specifically for text data mining, clustering serves as a

method of categorization of words based on the similarity of their occurrence within texts and classifying texts

using Ward’s method with cosine distance in the case of study by removing stop words.

Key-Words: - Clustering, text classification, Albanian text, stop words.

Received: May 5, 2022. Revised: August 9, 2023. Accepted: September 11, 2023. Available online: October 19, 2023.

1 Introduction

Clustering is a statistical technique that identifies

consistent sets within information and is applied in

various areas. Cluster analysis of data gathers akin

objects together in a cluster, as opposed to objects

located in disparate clusters which vary greatly from

one another. The similarity rate in data is

be allocated to a cluster. Fuzzy and model-based

grouping methods are frequently considered to be

soft because they make it simpler to assign units to

clusters. In text data mining, clustering is a

classification method that groups words according to

the similarity of their distribution in texts, also groups

documents by author or according to the similarity of

their topics, etc. Comparing the usage of high-

frequency features in texts is the most effective

method to distinguish between the text of different

organization, and social news clustering, [3], [4]. It

can also be used as an intermediate step for

applications such as multi-document summarization,

[5], [6], real-time text summarization, [7], sentiment

analysis, topic extraction, and labeling of

documents. In [1], [8], based on frequent feature

groups are proposed novel solutions to the

problem of text clustering, with the first focusing

on efficiency and accuracy, and the second on

hierarchical clustering and overcoming a specific

concluded that no one approach is used exclusively

for author identification; rather, researchers

apply a variety of techniques depending on the

characteristics of the understudied language, the

training data set, and the feature set. Albanian

language is classified as a unique branch on the

methods and computer science for the classification

of Albanian texts, for authorship attribution or author

identification. Our research work is focused on

developing and adapting mathematical models and

statistics methods for the identification of the authors

of Albanian texts with the aim of authorship

attribution and the detection of plagiarism in

Albanian texts, [10]. Previously we estimated the

probability of finding the correct author in Albanian

text classification using logistic and multinomial

of texts as positive and negative in short texts, [15],

[16], [17], [18], [19]. In this paper, we

present the agglomerative hierarchical

clustering to classify Albanian documents by

authors according to the similarity of their word

frequency. We apply the agglomerative

hierarchical clustering methods in a database

created from 100 Albanian documents from 10

different authors. The similarity of texts is

realized using cosine and Euclidian distances. The

with Albanian stop words included. To increase

the accuracy of classification, in this paper, we

upgrade the set of Albanian stop words in R for the

application of the hierarchical method as text

classification. We evaluate the clustering of

Albanian text by utilizing Dunn's index, thus

determining the optimal clustering.

2 Materials and Methods

Hierarchical clustering is a powerful technique

for identifying natural groupings within datasets,

distance, Manhattan distance, maximum

distance, etc. Hierarchical methods have the

advantage of the simple interpretation of the

clustering results and do

The goal of clustering is to minimize the distance

between the documents in the same cluster and to

maximize the distance between documents in

different clusters. There are two types of hierarchical

methods called agglomerative and divisive methods.

These techniques construct their hierarchy in the

opposite direction.

widely used in document classification, where large

volumes of textual data are analyzed and categorized

into groups based on their similarity. The

agglomerative hierarchical clustering algorithm

starts by treating each document as a separate cluster,

and then iteratively merges the most similar clusters

until all documents are grouped into a single cluster.

A linkage criterion, such as the average linkage,

complete linkage, or Ward's method, is used to

between each point and its corresponding centroid.

This method is sensitive to outliers, as it aims to

minimize the distance between data points and their

respective centroids. Another successful

agglomerative clustering method for text

uniform clusters in which the similarity between two

clusters is the maximum distance between

documents.

space, where each feature corresponds to a specific

term in the document. The similarity between two

documents is then measured using a distance metric.

The appropriate distance for text classification is

and gives more weight to document similarity

measures rather than distance metrics. However,

these methods can be effective depending on the

specific context and dataset. In this paper, we apply

these methods in different distance functions to

classify Albanian texts by the author.

processing steps. Initially, all non-textual elements

such as symbols and punctuation are eliminated from

the documents, and capital letters are converted to

lower letters. Every author has a unique style of

writing that stems from an unconscious habit. This is

reflected in their distinctive usage of grammar,

words, and punctuation which are different features

patterns. Among the most frequent words in different

languages, conjunctions, pronouns, and stop words

are extensively documented as functional but non-

informative words that perform a crucial role in

sentence structure. Removing stop words can lead to

[22]. While removing stop words can lead to

improved accuracy in some cases, it may not always

be beneficial. For example, in sentiment analysis,

removing stop words may not lead to significant

sentiment. As we mentioned above the Albanian

language is a unique branch in the family of Indo-

European languages, so we should consider the set of

Albanian stop words and study their impact in text

converting all capitalized letters to lowercase.

among the most frequent words in the corpus

presented below:

"per","deri","këtij","nëse","këto","siç","çdo","ose","

disa","është","ishte","kishte","sikur","kishin","janë",

"kanë","ishin","gjë","duke","prej","mund","kështu",

"nga","nuk","kur","kjo","që","dhe","të","në","se","s

words we create a term document matrix from the

corpus with a matrix consisting of 19146 terms and

97% sparsity.

The complete linkage method at first was applied

using cosine distance function. The optimal

clustering is achieved for 16 clusters with the highest

Dunn’s index of 0.7521. As result, 82% of texts are

well-classified by the author.

using the Euclidian distance function the optimal

clustering was achieved for 17 clusters with a

classification accuracy of 82%. The method gives the

same accuracy for the different distance functions.

The comparison of cluster dendrograms of Albanian text

classification with the complete linkage method in

cosine and Euclidian distance functions is presented

the right of Figure 2, we notice that the average

linkage method in the Euclidean distance function

gives the same accuracy of 81% with the highest

Dunn’s index value of 0.849. As it is shown in Figure

2 the entanglement coefficient keeps the same value

as in the complete linkage method but the number of

clusters is reduced to 10.

Using Ward’s method, we get different results for

different distance functions, as it is shown in Figure

3. Applying the cosine distance function, we get the

value of 0.849 and the overall percentage of text

classification by the author from the clusters is with

a value of 84%. Although Ward’s method, in both

distance functions, determines the same number of

clusters, the entanglement coefficient value of 0.57

As conclusion, the optimal clustering for Albanian

texts in the case of removing the most frequent stop

words is achieved by Ward’s method in cosine

distance with the highest Dunn index value of 0.72

3.2 Results in the second study case

In this case, we apply agglomerative methods in the

corpus without removing the stop words list we

mentioned above. After the preprocessing of the

texts, we created a term document matrix from the

corpus with a matrix consisting of 19283 terms and

97% sparsity.

cosine distance function, as a result, 62% of texts are

well-classified by the author. We get the same result

of text classification when we apply the complete

optimal clustering because Dunn’s index increases

when the number of clusters is increased. As result,

62% of texts are well-classified by the author.

Figure 5, presents the comparison of dendrograms for

the average linkage method. Using the average

linkage method in the cosine distance function we get

the dendrogram presented on the left of Figure 5. The

best classification is in 10 clusters with an accuracy

of 64% with the highest Dunn index value of 0.4642.

As it is shown from the dendrogram on the right of

classification with 15 clusters is achieved with the

highest Dunn index value of 0.72 with the overall

percentage of text classification by the author from

the clusters with a value of 75%. Ward’s method in

Euclidean distance function gives the dendrogram on

the right of Figure 6. The best classification with 16

clusters is achieved with the highest Dunn index

value of 0.849. As is seen in Figure 6, the overall

percentage of text classification by the author from

the clusters is with a value of 75%. Although Ward’s

Albanian texts by authors according to the similarity

of their word frequency are summarized in Table 3.

As conclusion, the optimal clustering for Albanian

texts in the second study case is achieved by Ward’s

method in cosine distance with the highest accuracy

of 75%.

Table 3. Results of agglomerative hierarchical clustering

in the second study case

3.3 Discussion

Results presented in Table 2 and Table 3 show that

the first case of study. From these results, we

conclude that removing the most frequent stop words

possible to determine the optimal clustering for the

complete linkage method because Dunn's index can

only keep increasing until all the texts are separated

into 100 clusters. In [12], using the logistic

regression model as a classification method to

estimate the probability of finding the correct

author in Albanian text was concluded that the

multinomial logistic regression model for

Albanian text has more advantages than the

function was more efficient. Comparing the

achieved outcomes, the best results for

Albanian text classification by the author are

obtained using Ward's method in the cosine distance

function, increasing the accuracy to 87%.

4 Conclusion

Using agglomerative hierarchical clustering

methods in a corpus with 100 Albanian texts of 10

different authors and the rich packages of text

mining in R we realized different classifications of

index. Applying agglomerative hierarchical

clustering methods in Albanian texts in two cases

we conclude that removing the most frequent

stop words improves the accuracy in Albanian

text classification. The best classification is

achieved using Ward's method of cosine

distance. The best model evaluated with the

maximum value of Dunn’s index of 0.7209,

separated the Albanian texts into 16 different

clusters based on the frequency of words. From the

are opportunities for improvement and expansion

development in information retrieval, pp 127–