LATTICE-CELL : HYBRID APPROACH FOR TEXT CATEGORIZATION

Sundarapandian et al. (Eds) : ICAITA, SAI, SEAS, CDKP, CMCA-2013
pp. 207–216, 2013. © CS & IT-CSCP 2013 DOI : 10.5121/csit.2013.3817
LATTICE-CELL : HYBRID APPROACH FOR
TEXT CATEGORIZATION
Hichem Benfriha1
, Fatiha Barigou2
, Baghdad Atmani3
Laboratoire d’Informatique d’Oran, LIO, Oran University, Algeria
{1
hichem.benfriha,2
fatbarigou,3
atmani.baghdad}@gmail.com
ABSTRACT
In this paper, we propose a new text categorization framework based on Concepts Lattice and
cellular automata. In this framework, concept structure are modeled by a Cellular Automaton
for Symbolic Induction (CASI). Our objective is to reduce time categorization caused by the
Concept Lattice. We examine, by experiments the performance of the proposed approach and
compare it with other algorithms such as Naive Bayes and k nearest neighbors. The results
show performance improvement while reducing time categorization.
KEYWORDS
Text Categorization, Concepts Lattice, Boolean Inference Engine, Cellular Automaton, CASI.
1. INTRODUCTION
Text Categorization is considered among the most important components in a research
information system. In fact text Categorization allows the document organization in categories
which accelerates and improves research and information retrieval[1]. Many different types of
techniques have been used in text categorization, including probabilistic naive Bayesian methods,
decision trees, k nearest neighbors and support vector machines. In this paper, the method we
propose is based on concepts lattice and cellular automaton.
Concept lattice are useful to explore the voluminous information : The text/web classification [2],
image retrieval [3], Information discovering and data warehouse classification [4]. The main
advantage of the lattice concept is its completeness. But the counterpart, building and operating
hierarchy are exponentially higher. So, the temporal and spatial complexity of concepts Lattice
formulates the goal of this research. We provide a novel contribution to reduce this complexity by
modeling the concepts lattice structure using cellular automaton CASI. We called our system
LATTICE-CELL [5].
This paper is structured as follows. In Section 2, we recall basic results on concept lattice [6] and
formal concept analysis (FCA) [7,8]. A detailed study of our LATTICE-CELL is presented in
section 3. In this section we describe the architecture and the categorization methodology of the
proposed system. Experimental results and comparative study are presented in sections 4 and 5
before concluding this work in Section 6.

Computer Science & Information Technology (CS & IT) 208
2. BACKGROUND ON CONCEPTS LATTICES
The contents of Formal Concept Analysis (FCA) [9] are formal context, formal concept and the
relation between the formal concepts. Formal context is defined as a triple K= (O, A, I), where O
is the set of objects, A is the set of attributes and I the binary relationship between O and A, that
is , oIa. In the formal context K, two mapping function X and Y is defined as follow
:
‫′ܣ‬ = {ܽ ∈ ‫ܣ‬ | ‫,ܴܽ݋‬ ∀ ‫∈݋‬ ‫}ܣ‬ (1)
‫′ܤ‬ = {‫݋‬ ∈ ܱ | ‫,ܴܽ݋‬ ∀ ܽ∈ ‫}ܤ‬ (2)
They are called the Galois connection between O and A. If tuple (A, B) from P(G)×P(M) satisfied
two conditions: A= X(B) and B=Y(A), or A=B’ and B=A’, (A, B) is called an concept from
formal context K, denoted C = (A, B), B and A are called the Intension and Extension of concept
C. Assume that C1 = (A1,B1) and C2 = (A2, B2) are two concepts, the order relation“≤”is
defined as . C1 is the sub concept of C2, C2 is super concept of C1. All
concepts and their relations consist of a concept lattice L.
For example, figure 1 on the left shows the binary relation K= (O,A, I) (or formal context) and on
the right the Hasse diagram of the concepts lattice derived from K. In the lattice, the concept C3 =
{(2, 6), (d)} in its Extension contains the objects 2 and 6, which have the property "d" in its
Intension.
Figure 1. Left: Binary table K= (O = {1, 2, , . . . ,5}, A = {a,b, . . . , e}, I). Right: The Hasse diagram of the
Concepts Lattice derived from K.
A variety of efficient algorithms exist for constructing the concept lattice from a binary table
[9,10,11,12,13,14,15,16]. A classical distinction between them is made on two axes. The first one
focuses on the way the initial data is acquired, it divides the methods on batch and incremental
ones. The second axis distinguishes algorithms that construct the entire lattice, i.e., set of concepts
and order (precedence) relation, from algorithms that only focus on concept set.
Authors in [15] proposed a novel paradigm for lattice construction that may be seen as a
generalization of the incremental one. In fact, the suggested methods construct the final result in a
divide-and-conquer mode, by splitting recursively the initial context B in two part B1 and B2 and
then merging the lattices resulting (L1 and L2) from the processing of sub-contexts at various
levels.
In this paper, we follow this approach to lattice algorithmics. The algorithm represents a
complement of the well known algorithms of Noris, Chein and Ganter that compute the set of
closed sets of a binary relation. When it is jointly applied with a concept computing algorithm, is
an alternative to the batch procedures of Bordat and Lindig as well as to the incremental
procedure of Godin, which constructs both the ground set and the Hasse diagram.

209 Computer Science & Information Technology (CS & IT)
Our goal is to build a hierarchy of concepts, to use it for text categorization. Algorithm 1
presented in figure 2 shows the assembly phase when L1=B1≤1 and L2 = B2 ≤2 represents the
two lattices corresponding to each part of the formal context and L = B ≤ represents the global
Concepts Lattice.
Figure 2. Algorithm 1 : Build the Global Lattice
3. LATTICE-CELL ARCHITECTURE
Our system called LATTICE-CELL consists of two modules as shown in figure 3. The first, is the
preprocessing module that produces an index of words which is the vector representation or the
formal concept. The outputs of the first module are the input of the second module which is
going to generate the categorization model based on the Boolean representation for the lattice
(cellular lattice).
Figure 3. LATTICE-CELL Architecture
3.1. Preprocessing
The learners cannot operate on the documents as they are; the documents must be given internal
representations that the learners can make sense of, once the learners build them. It is thus
customary to transform all the documents (i.e., those used in the training phase, in the testing
phase, or in the operational phase of the classifier) into internal representations. The step of
preprocessing build a representation of the selected documents, all through a process of indexing
is to extract terms representing the corpora [1].

a) Linguistic processing : This linguistic processing will be performed in 2 steps:
• Segmentation and lemmatization : This step is done by Tree Tagger1
which will split
the account into a set of units (words, numbers, punctuation, symbols) and reduce each
word (name, verbs, adverbs, adjectives) into entities called first or canonical forms
lemmas.
• Filtering and elimination of stopwords: Cleaning is performed to remove the non-
representative to retain only text content.
b) Dimension Reduction : This step can be performed by using a selection of words (lemmas)
which keep only the words that are considered useful in the categorization according to some
evaluation function, assigning a score to each word according to its discriminatory power , it
suffices to keep only the words with the highest score and to significantly reduce the size of the
space. In our case we chose to use the Information gain (IG) [17] as a measure of selection terms.
For illustrative purposes we selected 9 documents that represent 3 categories (S: Sport, E:
Economie and T: Television), and selected 6 words for index. Once the preprocessing step
completed we obtain a vector representation with binary weighting for the 9 documents presented
in table 1, which will represent the formal context.
Table 1. Vector representation with binary weighting
3.2. Boolean categorization model
3.2.1. Cellular Automaton
CASI (Cellular Automaton for Symbolic Induction) [18,19] is a cellular Automaton that is made
of two finite arbitrary long layers of finite state machines (cells) that are all identical. The
operation of the system is synchronous, and the state of each cell at time t+1 depends only on the
state of its vicinity cells, and on its own state at time t. This automaton, simulates the functioning
of the basic cycle of an inference engine by using two finite layers of finite automata. The first
layer, called CELFACT, is for representing the fact base, and the second layer, called CELRULE,
is for representing the rule base. In each layer the content of a cell determines whether and how it
participates in each inference step. At every step, a cell can be active or passive, and can take part
in the inference or not. The states of cells are composed of two parts: EF and SF, and ER and SR,
which are the input and output parts of the CELFACT cells, and of the CELRULE cells,
respectively. Any cell i in the CELFACT layer with input EF (i) = 1 is regarded as representing
an established fact. If EF (i) = 0, the represented fact has to be established. Any cell j of the
CELRULE layer with input ER (j) = 0 is regarded as a candidate rule. When ER(j) = 1, the rule
1
http://www.ims.uni-Stuttgart.de/projekte/corplex/TreeTagger
Stade Pays Personnage Ministre Puissance Visage
Doc 1 1 1 0 0 0 0
Doc 2 1 1 0 0 0 0
Doc 3 0 0 0 0 0 1
Doc 4 0 0 0 0 0 0
Doc 5 0 0 0 1 1 0
Doc 6 0 0 0 1 1 0
Doc 7 1 0 0 0 0 1
Doc 8 0 0 0 1 0 0
Doc 9 0 0 1 0 0 0

should not take part in the inference. Two incidence matrices called RE and RS define the
neighborhood of cells. They represent the facts input relation respectively and the facts output
relation. They are used in forward chaining.
The input relation, noted iREj, is formulated as follows: if (fact i ∈ Premise of rule j) then iREj
=1 else iREj = 0.
The output relation, noted iRSj, is formulated as follows: if (fact i ∈ Conclusion of rule j) then
iRSj =1 else iRSj =0.The cellular automaton as a cycle of an inference engine made up of two
local transitions functions δfact and δrule, where δfact corresponds to the evaluation, selection,
and filtering phases and δrule corresponds to the execution phase.
3.2.2. Construction of the cellular Lattice
The vector presented in table 1 is used as input for the construction of cellular lattice . For this,
we tuned algorithm 1 [15] by adding instructions to generate the cellular model representing the
lattice. Algorithm 2 presented in figure 5 is the modified version of algorithm 1 to construct the
cellular lattice.

To construct cellular lattice, algorithm 2 uses for this purpose four functions :
1) TCell _F : for eatch intention Ik and extension Ek, TCell_F creates two cells in the
CELFACT layer: (EF, IF, SF) [Ik] ← (0, 0, 0) to represent the fact Ik , and (EF, IF, SF)
[Ek] ← (0, 0, 0) to represent the fact Ek,(For the part Ek, TCell_F calculates the majority
class of each extension), and where the concept has generated an empty word is in its
intention or extension, TCell_F does not consider this concept.
2) TCell _R : for each concept Ck not empty, TCell_R creates a cell in the CELRULE layer
: (ER, IR, SR) ← (0, 1, 1) to represent the rule RC.
3) TCell_RE : initialize the incidence matrice of input : RE [Ik, RC] ← [1].
4) TCell_RS : initialize the incidence matrice of output: RS ← [Ek, RC] ← [1].
By applying algorithm 2 for LATTICE-CELL system on the same representation of table 1, we
obtain a cellular lattice represented by two layers of cellular automata CELFACT and CELRULE
and incidences matrices RE and RS shown in figure 6.
4. EXPERIMENTATION AND DISCUSSION
LATTICE-CELL system has been tested on Corpora about (sport, economy and television)2
.
Tables 2 summarizes the documents distribution .
2
http://deft08.limsi.fr/corpus-desc.php

Table 2. documents Distribution
Sport Economie Television Sample
Learning 300 300 300 900
Categorization 154 154 154 462
4.1. Experiments results
Once the cellular lattice built, we proceed to the categorization phase which automatically
determine documents category. To do this, we illustrate this step with an example of a document
labeled “Economie”, by following these steps: preprocessing, computing the similarity or
initialization (Search similar documents), inference engine and majority voting to determine the
document class.
4.1.1. Preprocessing
The categorization set will undergo the same preprocessing as learning set. We obtained the
following vector.
Table 3. Vector of terms
4.1.2. Computing the similarity (Initialization)
Document has been represented in vector form as points in a n-dimensional space, we determine
the most similar vectors by calculating the distance between these points. There are different
measures to calculate similarity as Jaccard, Cosine, Dice and Inner. Once one of these similarity
measures used to calculate the distance between the vector of words in the table 3 and the word
vectors in the layer CELFACT of figure 6. EF cells most similar vectors are initialized to 1, as
shown in Table 4 and with the use of the similarity measure Inner:
Table 4. Initialization of the layer CELFACT.
Facts EF IF SF
[Pays, Stade] 0 1 0
[S0 (100% S), (0% E), (0% T)] 0 1 0
[Visage] 0 1 0
[S3 (50% S), (50% E), (0% T)] 0 1 0
[Puissance, Ministre] 1 1 0
[S4 (0% S), (67% E), (33% T)] 0 1 0
[Visage, Puissance, Ministre] 1 1 0
[S5 (0% S), (100% E), (0% T)] 0 1 0
[Stade] 0 1 0
[S6 (67% S), (0% E), (33% T)] 0 1 0
[Personnage] 0 1 0
[S7 (0% S), (0% E), (100% T)] 0 1 0
Stade Pays Personnage Ministre Puissance Visage
0 0 0 1 1 0

4.1.3. Inference engine
Once the cells of the EF layer CELFACT instantiated to 1, the inference engine is initiated using
the two transition functions δrule and δfact in forward chaining. After the execution of the
transition function δrule, R4 and R5 are triggered as shown in the table 5 represents the layer
CELRULE.
Table 5. The layer CELRULE after executing of δrule transition function
Rules ER IR SR
R1 0 1 1
R2 0 1 1
R3 0 1 1
R4 1 1 1
R5 1 1 1
R6 0 1 1
Once the execution of δrule transition function, the inference engine using the second function
δfact it used for the execution phase. We note in table 6 that the EF cells peaks S4 and S5 are
instantiated to 1.
Table 6. The layer CELFACT after the inference engine.
Facts EF IF SF
[Pays, Stade] 0 1 0
[S0 (100% S), (0% E), (0% T)] 0 1 0
[Visage] 0 1 0
[S3 (50% S), (50% E), (0% T)] 0 1 0
[Puissance, Ministre] 0 1 0
[S4 (0% S), (67% E), (33% T)] 1 1 0
[Visage, Puissance, Ministre] 0 1 0
[S5 (0% S), (100% E), (0% T)] 1 1 0
[Stade] 0 1 0
[S6 (67% S), (0% E), (33% T)] 0 1 0
[Personnage] 0 1 0
[S7 (0% S), (0% E), (100% T)] 0 1 0
4.1.4. Majority Vote
After the inference engine, a majority vote is launched to calculate the class membership of the
document. If we take our example the two vertices are triggered: [S4 (0% S) (67% E) (33% T)]
and [S5 (0% S), (100% E) (0% T )]. The system calculates the percentage of each categories and
return to the user the highest 83.5% E (Economie) : for Sport : ((0 +0) / 2) = 0 ℅, for Economie:
((67 +100) / 2) = 83.5 ℅ and for Television: ((33 +0) / 2) = 16.5 ℅.
4.2. Comparative experiments
To study the effect of the similarity measure of the performance of LATTICE-CELL, we tested
four scenarios. In these scenarios, we varied the similarity measure, using Jaccard, Cosine, Dice
and Inner and use GI as a function of reduction equal to 500 words. Table 7 presents the results of
these experiments.

Table 7. Impact of similarity measures
We note, in Table 7 that the use of the Cosine measure gave a good performance in terms of
accuracy and error rate. Table 8 shows the comparison between the LATTICE-CELL using the
Cosine measure and other algorithms in the WEKA platform3
.
Table 8. Comparison of results
Precision Recall accuracy Error F-Mesure
NaiiveBays 0,41 0,42 0,40 0,60 0,38
ID3 0,38 0,36 0,36 0,64 0,35
K-PPV 0,38 0,35 0,36 0,64 0,29
SVM 0,38 0,37 0,38 0,62 0,36
LATTICE-CELL 0,39 0,33 0,60 0,40 0,36
LATTICE-CELL system allows us to obtain good results compared to other algorithms in terms
of classification accuracy and error rate. We note that the system LATTICE-CELL improves
nearly 7% K-PPV results in terms of F-measure and the proportion correctly classified documents
(Acc) is higher with a 60% success rate, and the proportion of misfiled documents (Err) is
significantly lower than the other algorithms as 40% error rate.
6. CONCLUSION
In this paper, we presented the functional architecture of our system LATTICE-CELL and the
experimental study. LATTICE-CELL system can automatically acquire, represent and process
knowledge extracted from sample texts in boolean form. The results show that the system
LATTICE -CELL gives good results compared to other algorithms of text categorizations such
ID3, K-nearest neighbor, Naïve Bayes and support vector machines (SVM) in terms of accuracy
and a much lower error rate. We managed to reduce the time categorization generated by the
Concepts Lattice, with the new representation. The advantages of this modeling based on cellular
automaton can be summarized as follows :
• The transition functions are easy to use, of low complexity, effective and robust regarding
extreme values. Moreover, they are well adapted to situations with many concepts.
• The cellular model amounts to a simple set of transition functions and production rules, which
not only make it possible to describe the problem at hand but also to build a classification
function for class prediction.
• The incidences matrices RE and RS, facilitates the rules transformation into Boolean
equivalent expressions and makes it possible, thereafter, to rely on elementary boolean algebra
to test other simplifications.
3
http://switch.dl.sourceforge.net/project/weka/weka-3-6-windows-jre/3.6.4/weka-3-6-4jre.exe
Precision Recall accuracy Error F-Mesure
Jaccard 0,34 0,39 0,54 0,45 0,36
Cosinus 0,39 0,33 0,60 0,40 0,36
Inner 0,33 0,39 0,54 0,45 0,36
Dice 0,32 0,69 0,42 0,57 0,44

REFERENCES
[1] Sebastiani. F. (2006) « Classification of text, automatic ». (K.Brown, Ed) The Encyclopedia of
Language and linguistics, 14, pp. 457-462.
[2] Géraldine, P. (2008) « Une mesure de similarité contextuelle pour l'aide à la navigation dans un
treillis ». 8 éme édition de la Conference de l'Extraction et la Gestion des Connaissances.
[3] Ould Djibril , M. & Oulad, H. (2007) « Un système pour l'indexation et la recherche des images de
l'art géométrique islamique », WOTIC’07,ENSIAS, Rabat.
[4] Messai, N., Devignes M.D.,Samil Tabone, M., Napoli A. (2005) « Treillis de Galois et ontologie pour
l’interrogation d’un annuaire de source de données biologiques (BioRegistry) » . Dans actes du
XXIIIème congrès informatique des organisation et systèmes d'information et de décision -
INFORSID.
[5] Benfriha, H., Barigou, F. & Atmani, B. (2011) « LATTICE-CELL : approche hybride pour la
catégorisation des comptes rendus hospitaliers ». The 1st International Conference on Information
system and technologies, Tébessa, Algeria.
[6] Barbut, M. & Monjardet, B. (1970) « Ordre et Classification: Algèbre et combinatoire», Hachette,
Paris.
[7] Ganter, B., Wille, R. (1999) « Formal Concept Analysis, Mathematical Foundations » , Springer,
Berlin.
[8] Wille, R. (1982) « Restructuring lattice theory: an approach based on hierarchies of concepts », in: I.
Rival (Ed.), Ordered Sets, Reidel, Dordrecht, Boston, pp. 445–470.
[9] Godin, R., Missaoui, R. & Alaoui, H. (1995) « Incremental concept formation algorithms based on
Galois (concept) lattices, Comput. Intell. 11 (2) 246–267.
[10] Bordat, J.-P. (1986) « Calcul pratique du treillis de Galois d’une correspondance ». Mathématiques et
Sciences Humaines, 96:31–47.
[11] Ganter, B. (1984) « Two basic algorithms in concept analysis» . preprint 831, Technische
Hochschule, Darmstadt.
[12] Godin, R., Missaoui , R. & Alaoui, H. (1995) « Incremental concept formation algorithms based on
Galois (concept) lattices. Computational Intelligence,11(2):246–267, 1995.
[13] Nourine, L. & Raynaud, O. (1999) « A Fast Algorithm for Building Lattices ». Information
Processing Letters, 71:199–204.
[14] Valtchev, P. (1999) « An algorithm for minimal insertion in a type lattice ». Computational
Intelligence, 15(1):63–78.
[15] Valtchev, P. & Missaoui, R. (2001) « Building concept (Galois) lattices from parts: generalizing the
incremental methods ». In H. Delugach and G. Stumme, editors, Proceedings of the ICCS’01, volume
2120 of Lecture Notes in Computer Science, pages 290–303.
[16] Norris, E. M. (1978) « An algorithm for computing the maximal rectangles in a binary relation ».
Revue Roumaine de Mathématiques Pures et Appliquées,23(2):243–250.
[17] Sebastiani, F. (2005) « Text Categorization, in Zanasi (dir.), Text Mining and its Applications to
Intelligence », CRM and Knowledge Management, WIT Press, 109-129.
[18] Atmani B. and Beldjilali B. (2007) « Knowledge Discovery in Databases: Induction Graph and
Cellular Automaton », Computing and Informatics Journal, Vol.26, N°2 P: 171-197.
[19] Barigou, F., Atmani, B. & Beldjilali, B. (2012) « Using a Cellular Automaton to Extract Medical
Information from Clinical Documents ». Journal of Information Processing Systems, Vol.8, No.1.

LATTICE-CELL : HYBRID APPROACH FOR TEXT CATEGORIZATION

More Related Content

What's hot (18)

Similar to LATTICE-CELL : HYBRID APPROACH FOR TEXT CATEGORIZATION (20)

More from cscpconf (20)

Recently uploaded (20)

LATTICE-CELL : HYBRID APPROACH FOR TEXT CATEGORIZATION