Managing text documents of a large size can be particularly challenging. For this reason, the indexation process is a crucial and decisive step for retrieving relevant textual features. Therefore, it is essential to design methods that effectively take into account complex data. In this work, we define a new contextual automatic indexation approach. Thus, we present ConIText System, a context-based approach for document indexation and its application. Also, we study its tolerance to large amounts of documents. In other words, we propose a new large corpus of texts and assess the performance gradually from 1.000 to 20.000 documents. It is to observe the behavior of the indexation system while data is getting big. To compare classification results, we used KNN and SVM classifiers. ConIText system outperforms the conventional statistical indexation, based on TFIDF method. Nevertheless, our contextualization system is generic, is not based on external resources. Although we have tested ConIText System on an Arabic dataset, it is not limited to one and unique language.
The indexation of texts is a crucial step in text processing. It allows to represent documents by their most relevant features. Several approaches are used for this purpose. However, extracting knowledge from textual data is an important issue, especially for large amounts of data.
Consequently, we proposed a contextual approach for the automatic indexation of texts. Indeed, in order to explore big data and to disclose hidden semantic information in unstructured documents, such as texts, an efficient indexation system is required. Consequently, we propose a new approach for text indexation based on semantic proximity and taking into account the contexts contained in each document.
This lead to our second proposition of a new approach for document
modeling. To test the performance of our approach, a large corpus was needed.
Therefore, we built our dataset from Arabic online Encyclopedias. It contains 20,000
documents labeled and categorized into 7 classes. The tests will be done gradually
in 1000, 5000, 10000 and 20000 documents to study the robustness of ConIText
system. Once the input is integrated into the system, a certain level of syntax
processing is required. After the preprocessing step, ConIText System identifies
common sentences of each document, and classify them according to their
semantic proximity. Then, the system identify contexts and model the document,
for classification aim [
We propose, in this work, a new approach for context discovery, based on sentence clustering techniques. Moreover, we introduce an efficient document modeling method. This model illustrates the most dominant context in a given document. Nevertheless, to assess the robustness of our proposed system, we have conducted experimental tests to compare its results to conventional statistical methods, as TF IDF.
The organization of this paper is as follows. In part 2, we introduce related works. Part 3 details our proposed ConIText System. In part 4, we highlight the experiments and results. Finally, we conclude by synthesizing the contributions of this work. 2
Most of the researches in the field of unsupervised information extraction focuses on keyword extraction Few of them offer methods to extract the contextual relations present in a document.
The contextual approaches aim, on the one hand, to remove the ambiguity of the meaning of texts. On the other hand, they highlight the semantic relations between these words. Semantic relationships can also be calculated using methods that evaluate the quantity of information shared between n-to-n words.
A study of Named Entity Recognition (NER) is presented in [
A survey of documents clustering using semantic approaches is introduced
in [
The authors of [
Mohamed and al. [
The authors of [
Same for Herskovic [
In [
Multivariate Relative Discriminative Criterion (MRDC) is proposed in [
The objective of the authors in [
The authors of [
Authors in [
In [
TF-IDF with dimensionality reduction can improve the precision in the
process of lexical matching, for identification of domain categories referencing to
a document, as proposed in [
The study in [
The authors of [
In [
Authors in [
Each of the presented methods highlight certain criteria. The approach we propose takes advantage of the existing advances and introduces a new concept of contextualization for a more refined indexation process. 3
In this part, we introduce the architecture of the automatic system for contextual indexation. It is a set of complex text mining methods, which forms an autonomous process of extracting contexts, then document features, for a context-based indexation (Fig. 1).
In fact, a dataset may contain different categories of documents that are either homogeneous or heterogeneous. Thus, the categories of politics and economics can be homogeneous. In contrast, the categories of new technologies and literature can be heterogeneous. Moreover, a document can express several contexts. For example, a document that describes a political decision and its impact on the economy will be difficult to be classed within the appropriate category. This is the kind of ambiguity that the ConIText System overcomes, by selecting the most appropriate context for each document.
We define context as the linguistic environment of a textual element (word, sequence of words, etc.) within the utterance in which it appears. That is to say the series of text units that precede and follow it. The term context refers to all the circumstances in which an act of enunciation takes place, as cultural and psychological situations, experiences and knowledge of the world, trade and promotion in economics, etc.
Furthermore, we define a sentence as the minimal element, which can express a context. A sentence is a set of words giving a complete meaning. Therefore, sentences is the first unit detected by ConIText System. Then, sentences close semantically are gathered to form a context. From each context of a document, we extract relevant features to form contexts vectors. Finally, the vector corresponding to the most dominant context models the whole document. To perform ConIText System, three steps are essential. The first step is the segmentation of texts. The second step is building context from which the features will be extracted. Finally, we model documents with our proposed principle of dominance. 3.1
The works that study the semantic grouping of sentences to extract relevant information inspire this proposition. As matter of fact, a sentence tends to express an idea, a context in our case, in an affine way. To go further in our data processing, the phase of sentences splitting is essential. In fact, words are organized into sequences, sentences or paragraphs, to define the meaning of the document. Therefore, the exploration of the relationship between the different components of the document is important to understand the document in depth.
Hence, this step consists on defining units of a text that will form the contexts. Segmentation of the document is the process of dividing the textual documents into meaningful sentences. Humans naturally understand the sentence when reading the text. Intuitively, we instill this power of understanding to our algorithm.
Texts have markers of explicit sentence boundaries. We use punctuation marks to delineate a sentence. In this work, we test ConIText system on an Arabic corpus. Since the Arabic language does not have a capital letters, our sentence segmentation is based on points, exclamation points and question marks (".", "!", "?"). 3.2
This phase consists on grouping the sentences obtained in clusters. Each cluster represents a context. Obviously, each document will have at least one context. To perform sentence clustering, we used Iterative K-means with a mean square error metric. This method makes possible to find, for each document, the optimal k number of the clustering method. Thus, each document is subdivided into k clusters. Therefore, we group together all the sentences to obtain k contexts in each document in the corpus (Fig. 2).
To conceive an efficient clustering process, the weights of words must be standardized based on their apparition in the document and their distribution in the entire corpus. In general, a common representation used for text processing is the TF-IDF representation.
The TFIDF weighting method is widely used by researchers. It is a frequency associated to the Vector Space Model (VSM), which involves associating a weight vector to each document. TF represents the number of occurrences of a word in the document and IDF is the absolute inverse frequency of the word in the corpus.
This method reduces the importance of common terms in the collection while ensuring that the matching of documents is more influenced by most discriminating words, which have a relatively high frequency in the document and low frequencies in the corpus.
In this work, since the form of the documents has changed, considering the generated contexts, we have introduced a slight modification for the TFIDF formula. Named TF-ICF (Term Frequency – Inverse Contexts Frequency), it is expressed as follows:
T F
ICF context(i) (t) = T F context(i)(t)x log( tf (t)context (i)
ICF (t) )
Where t is a term of the context i, TF context(i) (t) is the frequency of t within the context(i) and ICF(t) is the occurrence of t in all contexts of the corpus.
The obvious advantage of using this method is to calculate the relevance of a term according to all contexts of the corpus. This induces to express the value of the terms judged irrelevant in the conventional TFIDF system, whereas they have a powerful discrimination role in the document. 3.3
The dilemma in text mining is to select the appropriate representation of the textual information that will be able to represent the semantic content of the text. To model the document, we use the VSM representation. After building contexts, we calculate the score of each word in the context using the TF-ICF method, in order to model each context by a corresponding vector of words weights. Thus, each document is represented by a set of vectors (Fig. 3).
A constraint occurs, it is to represent each document by a single vector of weight. To perform this step, we define the principle of dominate context. After the contextualization, each document is divided into one or more contexts. Each context is modeled by one vector of weights. The Dominant Context is the vector of the strongest weights. Formally, n vectors model contexts of a given document D, as:
V (D) = fM axV i (D); i2[1; k]g
Where V(D) is the unique vector that models the document D, Vi is the vector modelling the context i, and k the number of contexts discovered in D. 4
During our research, we often face the problem of a lack of significant corpus. To approve the efficiency of an indexation system, it is essential to test it on a large amount of data. The best structured corpus are often not open access (see Table 2).
In lots of works on text mining, the authors build their own dataset. They choose the number of categories and themes to use. For each category, the documents are collected manually and those belonging to several categories are eliminated. Nonetheless, the size of datasets is relatively small to assess a system power, and the areas covered are geared towards specific issues.
This problem led us to create a new labeled corpus, in Arabic language, of 20,000 texts, with 27,605,263 words after document pretreatment (stemming and stopwords removing), and 7 classes labeled as presented in Table 3:
This corpus is collected from Arabic encyclopedias, and have the particularity of containing homogeneous themes and other heterogeneous to better assess the precision of systems. We make this data freely available to researchers. 5
In this work, we proposed ConIText System, a context-based system for automatic text indexation. To test our approach, we opted for a large dataset to evaluate its robustness and reliability. We have tested the proposed system gradually on a large amount of data (Table 3).
For comparative reasons, we conducted the tests using two classifiers, KNN and SVM. The different models of these classifiers will show the tolerance of our indexation approach to classification systems.
The experimental evaluation of the classifier is the final step in the classification process. It usually tries to evaluate the effectiveness of a classifier, namely its ability to make categorization decisions.
Fig. 4 and Fig. 5 show the results of KNN and SVM classification of documents using TFIDF and ConIText systems. These results are expressed by the f measure. The performance of our system is obvious. This is due to the complexity of the techniques used for the context-based indexation. However, the classification parameters are stationary for both classifiers.
The strong point that can be drawn from these experiments is the behavior of ConIText on a wide range of documents including more than 10000 documents. This advantage is more visible in the following figures.
The first test was performed on 1000 documents, which is a proportion widely used in the literature. Then, we have passed the test on 5000 documents. This number represents the maximum of the documents used in similar works (Table 2). We have pushed the tests on 10,000 documents to see how the system will react with such a mass of documents. Finally, we have performed a test on 20000 documents to study the stability of the system.
Table 4 and 5 presents the results of the comparison between TFIDF system and ConIText system, expressed by recall, precision and F-measure. Table 4 presents the result using KNN classifier and Table 5, SVM Classifier. In particular, those results show the relevance of using contextual indexing that effectively improves classification performance.
The results are illustrated in Fig. 6 and Fig. 7 in terms of precision and recall of classifiers KNN and SVM. The curves indicate that the performance of the TFIDF method drops dramatically as soon as the database takes more and more documents. However, ConITexte’s results are not only more satisfying but also tolerable to large datasets. We can see clearly that the performances are almost constant between 10,000 and 20,000 documents. This enhances the effectiveness of the context-based indexation system and confirms our theory.
We can deduce many conclusions from our experimental results. First, the contextual model showed its performance to be the appropriate representation for large datasets. Indeed, the context has several advantages over which it is possible to act to refine the extraction of the keywords. Second, the relations between words are expressed by maintaining the shared information of the context. This will certainly lead to better results. 7
In this paper, we have introduced ConIText, a contextual indexation System for texts. The integration of a semantic measure between sentences in this approach is necessary. For this reason, we have introduced our sentence grouping contribution to formalize the adaptation of the model to the semantic proximity. The advantage of this approach is that it does not need any preliminary specific knowledge to identify terms in order to assign them weights since the identification of terms is done from an automatic document processing.
We also proposed contextual modeling for document to increase the accuracy of indexation. In fact, the semantic proximity between words must be emphasized when we are dealing with complex and unstructured documents such as texts. For this reason, it is essential to broaden our thinking to models of representation adapted to the nature of our resources. To this end, we have introduced a contextual modeling for documents based on the principle of dominance. The advantage of this model is that it reduces the space representation of features and reduce the whole modeling of a document to its most significant context. 8
This work was funded by LITIS laboratory, and the University of Rouen Normandy, France.