The successful use of the Princeton WordNet for Text Categorization has prompted the creation of similar WordNets in other languages as well. This paper focuses on a comparative study between two WordNet based approaches for Multilingual Text Categorization. The first relates on using machine translation to access directly the princeton WordNet while the second avoids machine translation by using the WordNet associated for each language.
Multilingual Text Categorization(MTC) is a new area in Text categorization in which we have to cope with two or more languages (e.g English, Spanish and Italian).
MTC is a relatively new research topic, about which not much previous work
in the literature appears to be available. Most approaches have mainly addressed
different translation issues to solve the problem. R.Jalam et al. presented in [
In order to overcome the disadvantage of using machine translation systems,
many researches have been working on using linguistic resources such as bilingual
dictionaries and comparable corpora to induce correspondences between two
languages. A.Gliozzo and C.Strapparava propose in [
These last years, researches showed that using ontologies in monolingual
text categorization is a promising track. J.Guyot proposed in [
As shown in figure 1, the two approaches are composed of three phases: – Knowledge representation step; – Training step; – Predicting step.
For our experiments, the two approaches have the same training and prediction phases. The only difference is on the knowledge representation phase. 3.1
First approach The first approach consist on representing knowledge with the use of the Princeton WordNet. The labelled documents are mapped directly into the synsets of the princeton WordNet since they are expressed in English language. The unlabelled documents needs to be translated into the English language in order to be able to be mapped to the Princeton WordNet. The mapping into the princeton WordNet consists in replacing each term in a document by its most common meaning from the Princeton WordNet. We used a simple disambiguation strategy that consists of considering only the most common meaning of the term (first ranked element) as the most appropriate. Thus the synset frequency is calculated as indicated in the following equation: sf (ci, s) = tf (ci, {t ∈ T | f irst(Ref (t)) = s}) (1) where: – tf (ci, T ): the sum of the frequencies of all terms t ∈ T in the train documents of category ci.
– Ref (t): the set of all synsets assigned to term t in WordNet.
Second approach The second approach excludes the direct use of machine translation techniques by incorporating the WordNet associated for document languages. Indeed, each term document will be firstly mapped to the WordNet synsets of the language in which the document is expressed. As result, the labelled documents and the unlabelled documents will be mapped on different taxonomies. The labelled documents will be mapped to the Princeton WordNet, and the unlabelled documents will be mapped to the WordNets associated to unlabelled documents languages. It is necessary to match the taxonomies of all the used WordNets to a common taxonomy in order to unify document representations. Since the Princeton WordNet is the richest taxonomy, we have chosen it to be the common taxonomy. This matching offers the following advantages: – Avoiding the direct use of machine translation techniques which eliminate the problem of translation disambiguation. – Interconnecting the different WordNets to the most rich WordNet (Princeton
WordNet) which resolves the richness of some WordNets.
Formally,the synset frequency is calculated as indicated in the following equation: sf (d, s) = tf (d, {t ∈ T | match(f irst(Ref (t, L))) = s}) where: – tf (d, T ): the sum of the frequencies of all terms t ∈ T in the unlabelled document d . – L: The language of the unlabelled document d. – Ref (t, L): the set of all synsets assigned to term t in WordNet associated to language L. – match(s): the corresponding synset of the synset s on the Princeton WordNet. Capturing relationships After mapping terms into Princeton WordNet synsets, this step consists in using the WordNet hierarchy to capture some useful relationships between synsets (hypernymy in our case).The synset frequencies will be updated as indicated in the following equation: sf (ci, s) =
b∈H(s) sf (ci, b)
– ci: the ith category. – b and s are synsets.
– H(s): the hyponyms set of synset s 3.2
The training phase consists in using the labelled documents to create conceptual
categories profiles. Formally, each category will be represented by a conceptual
profile which contains the K better synsets (our features) characterizing best
the category compared to the others. For this purpose we used the χ2
multivariate statistic for feature selection. The χ2 multivariate [
Cjχk2 = N (fjkf−j.ffj..kf.k)2 × sign(fjk − fj.f.k) (4)
– fjk = NNjk : the relative occurrence frequency. – N : The total sum of the occurrences.
– Njk: The frequency of the synset sj in the category ck.
Once the contributions of synsets are calculated and ordered for each category, the conceptual profile of each category contains the k first sorted synsets. 3.3
The Prediction phase consists on using the conceptual categories profiles in
classifying unlabelled documents. Our Prediction phase consists of:
– Weighting the conceptual categories profiles and the conceptual vector of the
unlabelled document. In our experiments, we used the standard tf idf (term
frequency - inverse document frequency) function [
Si,j = √
s∈i j tfidf(s,i)×tfidf(s,j) s∈i tfidf2(s,i)× s∈j tfidf2(s,j) With: s: a synset, i and j: the two vectors (profiles) to be compared. tf idf (s, i): the weight of the synset s in i.
tf idf (s, j): the weight of the synset s in j. 4 4.1
For our experimentations, we extracted a bilingual dataset from Reuters Corpus Vol. 1 and 2 (RCV1, RCV2) using English training (RCV1) and Spanish test documents (RCV2). Our dataset is based on topic (category) codes with a rather varying number of documents per category as shown in Table1 For comparison, we have tested the two approaches on our multilingual dataset. Experimental results reported in this section are based on the so-called ”F1 measure”, which is the harmonic mean of precision and recall.
F1(i) = 2×precision×recall
precision+recall
The results of the experimentations are presented in Table2, Concerning the profiles size, the best performances are obtained with size profile k = 900 for the two approaches. Indeed, the performances improve more and more by increasing the size of profiles.
Comparing the results of the two approaches, the first approach largely outperform the second approach. (7) In this paper, we have compared two approches for using WordNets for MTC, The first approach is based on using machine translation to use the Princeton WordNet while the second approach is based on replacing the use of machine translation by incorporating a WordNet for each language. The results of the experimentations show that the use of WordNets does not guarantee good results rather than those obtained by the Princeton WordNet. Future works will concern the experimentation of the second approach with differents WordNets in order to be able to confirm the obtained results.