Automatic Web Page Classification

Jiří Materna

Natural Language Processing Centre

Faculty of Informatics, Masaryk University
Botanická 68a, 602 00, Brno, Czech Republic
[email protected]
http://nlp.fi.muni.cz

Abstract. Aim of this paper is to describe a method of automatic web

page classification to semantic domains and its evaluation. The classifica-
tion method exploits machine learning algorithms and several morpho-
logical as well as semantical text processing tools. In contrast to general
text document classification, in the web document classification there are
often problems with short web pages. In this paper we proposed two ap-
proaches to eliminate the lack of information. In the first one we consider
a wider context of a web page. That means we analyze web pages refer-
enced from the investigated page. The second approach is based on so-
phisticated term clustering by their similar grammatical context. This is
done using statistic corpora tool the Sketch Engine.

Key words: automatic classification, machine learning, web document, the-

saurus

1 Introduction
1.1 Motivation
At the present time the World Wide Web is the largest repository of hypertext
documents and is still rapidly growing up. The Web comprises billions of
documents, authored by millions of diverse people and edited by no one
in particular. When we are looking for some information on the Web, going
through all documents is impossible so we have to use tools which provide us
relevant information only. The widely used method is to search for information
by fulltext search engines like Google1 or Seznam2 . These systems process list
of keywords entered by users and look for the most relevant indexed web
pages using several ranking methods. Another way of accessing web pages is
through catalogs like Dmoz3 or Seznam4 . These catalogs consist of thousands
web pages arranged by their semantic content. This classification is usually
done manually or partly supported by computers. It is evident that building
large catalogs requires a lot of human effort and fully automated classification
1 2 3
http://www.google.com http://search.seznam.cz http://www.dmoz.org
4
http://www.seznam.cz

Petr Sojka, Aleš Horák (Eds.): Proceedings of Recent Advances in Slavonic Natural Language Processing,
RASLAN 2008, pp. 84–93, 2008. c Masaryk University, Brno 2008
 Automatic Web Page Classification 85

systems are needed. However several systems for English written documents
were developed (e.g. [1,2,3,4,5]) the approaches do not place emphasis on short
documents nor on the Czech language.

1.2 Objective

Classical methods of text document classification are not appropriate for web
document classification. Many of documents on the Web are to short or suffer
from a lack of linguistic data. This work treats with this problem in two novel
approaches:

– Experiments have proved that hypertext links in web documents usually

direct to documents with similar semantic content. This observation leads
to use these referenced web pages as an extension of the investigated one for
the purposes of processing their linguistic data as well. However there are
some restrictions. The referenced documents must be placed on the same
server (to avoid joining advertisement or other non-related material) and a
level of recursion must be limited. We experimentally set the limit to 2.
– The former method increases amount of linguistic data for the most part of
documents enough but there is another problem. To use machine learning
algorithms we need to build a high dimensional vector space where
each dimension represents one word from or phrase. In spite of the fact
that several machine learning algorithms are adjusted to high number
of dimensions, in this case the high number of dimensions decreases
algorithm accuracy and we have to proceed to dimensional clustering. The
joining of two or more dimensions (in this case words) is based on using a
special thesaurus built on training data. The method will be described more
precisely in the Section Term clustering.

2 Preprocessing

In order to use machine learning algorithms we need to build a training

data set. There were selected 11 domains (Cestování, Erotika, Hry, Informační a
inzertní servery, Kultura a umění, Lidé a společnost, Počítače a internet, Sport, Věda
a technika, Volný čas a zábava, Zpravodajství) according to the top-level domains
in http://odkazy.seznam.cz catalog and for each domain collected 1 GB of
sample data.

2.1 Data Cleaning

Despite of selecting restricted document content-types (HTML, XHTML) it is

necessary to remove noise from the documents. An example of unwanted data
is presence of JavaScript (or other scripting languages) as well as Cascading
Style Sheets (CSS) and the most of meta tags. Elimination of such data was
mostly done by removing head part of the document (except of content of
86 Jiří Materna

the title tag which can hold an important information about domain). As
other unwanted data were marked all n-grams (n>10) where portion of non
alphanumeric characters was grater than 50 %.
Very important issue of document preprocessing is charset encoding detec-
tion. However the charset is usually defined in the header of the document, it
is not a rule. We have used a method of automatic charset detection based on
byte distribution in the text [6]. This method works with a precision of about
99 %.
A lot of web sites allows user to chose language. Even some web pages
on the Czech internet are primarily written in foreign language (typically in
Slovak). With respect to used linguistic techniques, we are made to remove
such documents from the corpus. The detection of foreign languages is similar
to charset encoding detection based on typical 3-gram character distribution.
There has been built a training set of Czech written documents and computed
the typical distribution. Similarity of training data with the investigated docu-
ments is evaluated using cosine measure.

2.2 Corpus construction

Cleaned raw data serve as a groundwork for the training corpus construction.
To represent corpus data we use vertical text with following attributes:
– word – original word form,
– lemma – the canonical form of a word. To get lemma we have used Ajka
tagger [7] and disambiguator Desamb [8],
– tag – morphological tag of a word (obtained from Ajka).
To process data has been used corpus manager Manatee [9] which offer many
statistical functions as well as the Sketch Engine tool [10]. This system can ex-
tract so called word sketches which provide information about usual grammat-
ical context of terms in corpus and are used for the thesaurus construction.

3 Document Model
In order to use these data in machine learning algorithms we need to convert
them into appropriate document models. The most common approach is vector
document model where each dimension of vector represents one word (or
token in corpus). There are several methods of representing the words.
Let m is number of documents in the training data set, f d (t) frequency of
term t in document d for d ∈ {1, 2, . . . , m} and Terms set of terms {t1 , t2 , . . . , tn }.

3.1 Binary representation

Document d is represented as a vector (v1 , v2 , . . . , vn ) ∈ {0, 1}n , where

1 if f d (ti ) > 0
vi =
0 else
 Automatic Web Page Classification 87

3.2 Term frequency representation

Document d is represented as a vector (v1 , v2 , . . . , vn ) ∈ Rn , where

f d ( ti )
vi =
m

3.3 Term Frequency – Inverse Document Frequency (TF-IDF)

Disadvantage of previous two methods may be a fact of treating with all terms
in the same way – they are not weighted. This problem can be solved by using
IDF coefficient which is defined for all ti ∈ Terms as:
!
m
IDF (ti ) = log2
|{ j : f j (ti ) > 0}|

By combining TF and IDF we get:

!
f (t ) m
vi = d i · log2
m |{ j : f j (ti ) > 0}|

For TF and TF-IDF methods is convenient to discretize their real values. The
MDL algorithm [11] based on information entropy minimization has been used.

4 Term Clustering

The term clustering is based on a special dictionary. The dictionary is defined

as a total function
s : Terms → Rep
which assigns just one representative from Rep ⊆ Terms to each member of
Terms set. The s function defines equivalence classes on Terms by equivalence
relation σ:
( a, b) ∈ σ ⇐⇒ s( a) = s(b)
Reversely, let C ∈ Terms/σ, there always exists some function s. If r is an
arbitrary member of C, then

s( x ) = r for all x∈C

The construction of dictionary consits of following steps:

1. Finding characteristic set for each term t ∈ Terms.

2. Defining equivalence classes on Terms set based on similarity of their
characteristic set.
3. Dictionary function s definition.
88 Jiří Materna

4.1 Characteristic set

Characteristic set construction is mostly based on using the Sketch Engine
and its word sketches. Word sketches are one-page automatic, corpus-based
summaries of a word’s grammatical and collocational behavior generated
by Sketch Engine which takes as input a corpus of any language and a
corresponding grammar patterns and which generates word sketches for the
words of that language [10].
It suggest itself to look for similar word sketches and build a thesaurus.
For each lemma l with sufficient frequency we get a list of similar words
SPl = [w1 , w2 , . . . , wn ] ordered by their indexes of similarity i1 , . . . , in with
lemma l [12]. Lets define the characteristic list CHL(l ) for each lemma l from
the corpus:
– if frequency of lemma l in the corpus is less than 100:
CHL(l ) = [l ]
– else:
CHS(l ) = [w1 , w2 , . . . , wk ] : ∀i j ∈ {i1 , i2 , . . . ik } : i j ≥ 0.1
An example of characteristic list of lemma auto (car) is shown in Table 1.

Table 1. Characteristic list of lemma auto

auto 1
automobil 0.184
autobus 0.171
vůz 0.166
vozidlo 0.153
vlak 0.141
aut 0.133
tramvaj 0.126
lod’ 0.124
letadlo 0.112
trolejbus 0.11

The table shows that the incorporated words are really semantically similar.
However, there are some problems with homonyms and tagging errors (in this
case term aut). The characteristic set is defined in the way of eliminating words
occurred in the corpus more frequently in other senses than we currently treat
with.
Let CHL(l ) = [w1 , w2 , . . . , wk ] is the characteristic list of the lemma l,
S(l ) = {w1 , w2 . . . , wk } and S p (l ) = {wi |i ≤ k/p} where p ∈ R+ is a constant
coefficient. The characteristic set is defined as
CH (l ) = {wi : q · |S(wi ) ∩ S p (l )| ≥ |S p (l )|}
where q ∈ R+ is an appropriate constant. The experiments have shown that the
best values seem to be p = 2, q = 2.
 Automatic Web Page Classification 89

4.2 Dictionary construction

When we have a characteristic set for each lemma from corpus it remains
to define clustering and dictionary function s. Intuitively, the clusters are
composed of terms with similar characteristic sets. In this work, the similarity
is measured by Jaccard index, where similarity of terms a and b is defined as
|CH ( a) ∩ CH (b)|
j( a, b) =
|CH ( a) ∪ CH (b)|
The clustering works on the principle of hierarchical clustering [13] using top-
down method. Minimal similarity for joining sets was experimentally set to
0.45. These clusters define equivalence relation σ.
Let f req( x ) is a frequency of term x. We define dictionary function s: ∀S ∈
Terms/σ, ∀ a ∈ S : s( a) = b where b ∈ S, f req(b) = max { f req( x )| x ∈ S}. In the
case of ambiguity the first possible lemma in lexicographical order is used.
Finally, when we have dictionary function s, we are able to replace all terms
t in corpus by their representatives s(t).

5 Attribute Selection
Even after application of the dictionary function there are a lot of different terms
for using machine learning algorithms in the corpus and it is necessary to select
the most convenient ones. Statistics provides some standard tools for testing if
the class label and a single term are significantly correlated with each other. For
simplicity, let us consider a binary representation of the model. Fix a term t and
let
– k i,0 = number of documents in class i not containing term t
– k i,1 = number of documents in class icontaining term t
This gives us a contingency matrix
It \C 1 2 ... 11
0 k1,0 k2,0 ... k11,0
1 k1,1 k2,1 ... k11,1
where C and It denote boolean random variable and k l,m denotes the number
of observation where C = l and It = m.

5.1 χ2 test
This measure is a classical statistic approach. We would like to test if the
random variables C and It are independent or not. The difference between
observed and expected values is defined as:

(k l,m − n · P(C = l ) P( It = m))2

χ2 = ∑ ∑ n · P(C = l ) P( It = m)
l ∈Class m∈{0,1}
90 Jiří Materna

5.2 Mutual Information Score

This measure from information theory is especially useful when the multinom-
inal document model is used and documents are of diverse length (as is usual).
The mutual information score is defined as:
k l,m k l,m /n
MI ( It , C ) = ∑ ∑ n
log
(k l,0 + k l,1 ) · (∑i∈Class k i,m )/n2
l ∈Class m∈{0,1}

6 Classification and Evaluation

We have tested the classification using four algorithms (C4.5, k-nearest neigh-
bors, Naîve Bayes classifier and Support machines) on 3,500 randomly chosen
training samples and 1,500 testing examples. For testing has been used 10-fold
cross validationi [14]. As an implementation, we have chosen open source data
mining software Weka [15] for algorithm C4.5, k-nearest neighbors and Naîve
Bayes classifier and LIBSVM [16] for Support Vector machines.
First, we compare preprocessing methods and selected machine learning
algorithms on data without clustering and document extending. Next, the
best-resulting method is chosen to test approaches presented in this paper. In
Figure 1 you can see overall accuracy graphs of all presented algorithms and
methods of document model representation. The best results with 79.04 % of
overall accuracy have been acquired using Support vector machines algorithm,
term frequency document model and MI-score selection of attributes.

Fig. 1. Preprocessing and classification algorithms

Figure 2 shows dependency of overall accuracy on attribute number with-

out clustering, with clustering based on same lemmas and with clustering
 Automatic Web Page Classification 91

Fig. 2. Clustering methods

Fig. 3. Extending by referenced documents

based on selected lemmas. In the third case, only nouns, adjectives, verbs and
adverbs have been selected. You can see that overall accuracy in all cases grows
92 Jiří Materna

till about 12,000 attributes. After this threshold the overall accuracy does not
vary significantly. The best result (83.4 %) was acquired using clustering based
on same lemmas.
Finally, Figure 3 shows result of experiments with extended documents,
clustering based on same lemmas and on both lemmas and dictionary. The
overall accuracy growth from previous experiment is about 5.9 % for lemma
based clustering and 8.2 % for dictionary based clustering.

7 Conclusion
We have presented a method of automatic web page classification into given
11 semantic classes. Special attention has been laid on treating with short
documents which often occur on the internet. There have been introduced
two approaches which enable classification with overall accuracy about 91 %.
Several machine learning algorithms and preprocessing methods have been
tested. The best result has been acquired using Support vector machines with
linear kernel function (followed by method of k-nearest neighbors) and term
frequency document model with attribute selection by mutual information
score.

Acknowledgments. This work has been partly supported by the Academy of

Sciences of Czech Republic under the projects 1ET100300419 and 1ET200610406,
by the Ministry of Education of CR within the Center of basic research LC536
and in the National Research Programme II project 2C06009 and by the Czech
Science Foundation under the project 407/07/0679.

References
1. Asirvatham, A.P., Ravi, K.K.: Web page categorization based on document structure
(2008) http://citeseer.ist.psu.edu/710946.html.
2. Santini, M.: Some issues in automatic genre classification of web pages. In: JADT
2006 – 8èmes Journèes internationales d’analyse statistiques des donnés textuelles,
University of Brighton (2006).
3. Mladenic, D.: Turning Yahoo to automatic web-page classifier. In: European
Conference on Artificial Intelligence. (1998) 473–474.
4. Pierre, J.M.: On automated classification of web sites. 6 (2001)
http://www.ep.liu.se/ea/cis/2001/000/.
5. Tsukada, M., Washio, T., Motoda, H.: Automatic web-page classification by using
machine learning methods. In: Web intelligence: research and development,
Maebashi City, JAPON (23/10/2001) (2001).
6. Li, S., Momoi, K.: A composite approach to language/encoding detection. 9th
International Unicode Conference (San Jose, California, 2001).
7. Sedláček, R.: Morphemic Analyser for Czech. Ph.D. thesis, Faculty of Informatics,
Masaryk University, Brno (2005).
8. Šmerk, P.: Towards Morphological Disambiguation of Czech. Ph.D. thesis proposal,
Faculty of Informatics, Masaryk University, Brno (2007).
 Automatic Web Page Classification 93

9. Rychlý, P.: Korpusové manažery a jejich efektivní implementace (in Czech). Ph.D.
thesis, Faculty of Informatics, Masaryk University, Brno (2000).
10. Kilgarriff, A., Rychlý, P., Smrž, P., Tugwell, D.: The Sketch engine in practical
lexicography: A reader. (2008) 297–306.
11. Fayyad, U.M., Irani, K.B.: On the handling of continuous-valued attributes in
decision tree generation. Machine Learning 8 (1992) 87–102.
12. Kilgarriff, A.: Thesauruses for natural language processing. Proc NLP-KE (2003).
13. Berka, P.: Dobývání znalostí z databází. Academia (2003).
14. Kohavi, R.: A study of cross-validation and bootstrap for accuracy estimation and
model selection. In: IJCAI. (1995) 1137–1145.
15. Witten, I.H., Frank, E.: Data mining: Practical machine learning tools and tech-
niques. Technical report, Morgan Kaufmann, San Francisco (2005).
16. Chang, C.C., Lin, C.J.: LIBSVM: a Library for Support Vector Machinesi. Technical
report, Department of Computer Science National Taiwan University, Taipei 106,
Taiwan (2007).