International Journal of Computational Intelligence Systems

Volume 13, Issue 1, 2020, Pages 1218 - 1226

Tolerance Rough Set-Based Bag-of-Words Model for Document Representation

Authors
Dong Qiu1, *, Haihuan Jiang1, Ruiteng Yan2
1College of Science, Chongqing University of Posts and Telecommunications, Nanan, Chongqing, 400065, P.R. China
2School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Nanan, Chongqing, 400065, P.R. China
*Corresponding author. Email: dongqiumath@163.com
Corresponding Author
Dong Qiu
Received 12 May 2020, Accepted 4 August 2020, Available Online 19 August 2020.
DOI
10.2991/ijcis.d.200808.001How to use a DOI?
Keywords
Document representation; Tolerance rough set; Bag-of-Words
Abstract

Document representation is one of the foundations of natural language processing. The bag-of-words (BoW) model, as the representative of document representation models, is a method with the properties of simplicity and validity. However, the traditional BoW model has the drawbacks of sparsity and lacking of latent semantic relations. In this paper, to solve these mentioned problems, we propose two tolerance rough set-based BOW models, called as TRBoW1 and TRBoW2 according to different weight calculation methods. Different from the popular representation methods of supervision, they are unsupervised and no prior knowledge required. Extending each document to its upper approximation with TRBoW1 or TRBoW2, the semantic relations among documents are mined and document vectors become denser. Comparative experiments on various document representation methods for text classification on different datasets have verified optimal performance of our methods.

Copyright
© 2020 The Authors. Published by Atlantis Press B.V.
Open Access
This is an open access article distributed under the CC BY-NC 4.0 license (http://creativecommons.org/licenses/by-nc/4.0/).

1. INTRODUCTION

With the explosive growth of the Internet, countless text data are accumulated constantly. In addition, unlike numerical data belonging to the structured data type, document or text data are unstructured data. Unstructured data are not appropriate to be directly applied in machine learning or deep learning algorithms. As the basis of natural language processing (NLP) and text mining tasks, efficient text or document representation is particularly important. The main challenges on document representation are the ways of transforming unstructured text data into structured data. For a good document representation, on the one hand, it should be able to truly reflect the content of the document, on the other hand, it should have the ability to distinguish different documents. Additional, it has optimal performance in some NLP applications such as text classification, information retrieval and text clustering.

The bag-of-words (BoW) model [1] is a representative document representation method, which has been widely used in information retrieval, text classification [2] and sentiment analysis [3]. In the BoW model, the word order, grammar and syntax of the document are ignored and every document is only regarded as a set of words, or a combination of words. The emergence of each word in the text is independent and does not depend on the appearances of other words. The BoW model takes all the non-repetitive words of all the documents in the dataset as basis terms. Each document is denoted by a fixed dimension vector, the length of which is equal to the number of the basis terms. Each component of the vector corresponds to the frequency of the basis word that appears in the document. The BoW model has the virtue of briefness and efficiency. However, it suffers from the defects of sparsity and lack of latent semantic relevance. These issues have been addressed many times in earlier research, but we are committed to solving this problem without any prior knowledge and without supervision. Hence, we apply tolerance rough set theory [4] to improve the traditional BoW model, and thus propose two new tolerance rough set-based BOW models, TRBoW1 and TRBoW2, which can expand the semantic space of documents. And they are unsupervised and not need any prior knowledge.

There are many researches have been done on document representation. Among them, the BoW model is one of the most classical method. Owing to that the BoW model is sparse, high-dimensional and lack of latent semantics, some other models emerge as the times require from distinctive perspectives, such as feature selection algorithms [58], weight calculation algorithms [912] and dimensionality reduction algorithms [1318]. Text feature selection algorithms contain document frequency [7], Chi square (χ2) [6], information gain (IG) [5] and mutual information (MI) [8]. To measure the importance of a feature item in document representation better, some works on feature weight calculation have presented, including term frequency inverse document frequency (TF-IDF) [12], entropy [10] and term frequency inverse word frequency (TF-IWF) [9]. TF-IDF is a commonly used one [11]. In order to mine the correlations among words and alleviate the problem of high dimensionality, many scholars have developed a series of latent topic models, the earliest of which is the latent semantic analysis (LSA) model [14]. LSA projects documents into low-dimensional latent semantic space by singular value decomposition (SVD) of word frequency-document matrix [19]. Hofmann et al. established a probabilistic latent semantic analysis (PLSA) model to solve the polysemy problem of LSA [15,16]. On the basis of PLSA, Blei et al. introduced the Dirichlet prior distribution and proposed the Latent Dirichlet Allocation (LDA) [13]. Some other scholars generalized the LDA model and put forward Gaussian LDA model [20], Latent Feature Topic Modeling (LFTM) model [21] and Topical Word Embedding (TWE) model [22]. Although the algorithms above have been successfully used for document representation and played the role of dimensionality reduction, they are still unable to capture the real semantics among words.

Methods based on neural network and deep learning arose at the historic moment. However, majority of them are supervised or need prior knowledge. A sentence encoder-decoder model was proposed in [23], called Skip-Thought model, which learned the representation through predicting the following sentence. But it need a lot of training. Wu et al. [24] represented documents based on phrase embeddings by parsing, in which three Phrase2Vec methods are constructed on the basis of Word2Vec. To solve the problem of sparsity, Yao et al. [25] utilized word sematic similarity by constructing a neural probabilistic language model. Gao et al. [26] proposed the CCTSenEmb model by obtaining the association between adjacent sentences in the process of sentence prediction. The Hybrid-WikiBoC approach was proposed to improve the performance of BoW in [27], taking Wikipedia as the background knowledge. To overcome the limitation of sparsity and inability to dig out semantic significance behind words of the BoW model, Zhao et al. combined the fuzzy system with BoW, and proposed a novel fuzzy bag-of-words (FBoW) model [28]. FBoW model converted the original hard mapping into fuzzy mapping, representing the component of document representation vector as the similarity between words and basis terms. Furthermore, the fuzzy bag-of-word clusters (FBoWC) model was developed to solve the problem of high dimension by clustering the basis terms in [28]. However, it is based on the prior knowledge, which needs the word embedding to capture the semantic relevance.

This paper is organized as follows: The tolerance rough set model is reviewed in Section 2. Section 3 presents our proposed tolerance rough set-based BOW models detailedly. Section 4 demonstrates the experimental results. The discussion and analysis of the experimental results is given in section 5. In Section 6, some conclusions are came to.

2. TOLERANCE ROUGH SET

In this section, the tolerance rough set model is reviewed in brief.

Rough set theory, put forward by Polish scholar Pawlak in 1982 [29], is applied to process the problem of uncertainty and fuzziness. It provides both theories and technologies for researchers in a broad variety of fields of artificial intelligence such as data mining, machine learning and NLP. Rough set, which is based on the equivalence relation, uses a pair of concepts, upper approximation and lower approximation, to measure the relationships of one certain object and a set X. The relationships can be represented as belonging to, possibly belonging to and not belonging to. For the shortcoming of traditional rough set model, on different circumstances and needs, researchers have developed plenty of extended rough set models, like probabilistic rough set model [30], decision rough set model [31] and tolerance rough set model [4]. For the reason that equivalence relation contains three properties of reflexivity, symmetry and transitivity, in which the limitation of transitivity leads to the inapplicability in some cases, Skowron et al. modified the original equivalence relation to tolerance relation, and the corresponding approximation space to tolerance approximation space [4].

A tolerance space is represented by a quadruple =(U,I,ν,P) in [32]. Considering that the structural function P has no practical meaning in the definition and has no impact on the whole content, we simplify the quadruple to a triple =(U,I,ν). In the triple, U={x1,x2,,xn} is the universe of all the objects, I:U2U is an uncertainty function, and I(x) is a tolerance class on x. If an object shares similar information with x, it belongs to I(x). ν:2U×2U[0,1] is a vague inclusion. Any function having the properties of reflexivity and symmetry can be an uncertainty function I(x), which can be explained that for any x,yU,xI(x) iff xI(y). The vague inclusion ν has the property of monotonicity, i.e., for any X,Y,ZU and YZ, ν(X,Y)ν(X,Z). It measures the degree of inclusion of sets, whether a set X contains the tolerance class I(x) of an object xU [32]. The upper approximation U(,X) and the lower approximation (,X) of any XU are defined as follows:

U(,X)={xUP(I(x))=1&ν(I(x),X)>0},(1)
(,X)={xUP(I(x))=1&ν(I(x),X)=1}.(2)

3. TOLERANCE ROUGH SET-BASED BOW MODELS

In this section, we describe the proposed TRBoW models in detail. Now we introduce the definition of the triple of tolerance rough set in document representation. The notations in this section are listed in Table 1, where the vectors and matrixes are written as bold formatting.

Notation Description Notation Description
D The document corpus W The universe
v Document size n Basis terms size
di The vector representation of document di cij The co-occurrence times of words wi and wj
θ Co-occurrence threshold I n-dimensional unit vector
The tolerance space X The document representation matrix by BoW
Iθ(W) The uncertainty matrix of the basis terms μ(W,D) The fuzzy membership matrix of the words and documents
U(,di) The upper approximation set of the document di (,di) The upper approximation set of the document di
U(,D) The upper approximation matrix of the documents L(,D) The lower approximation matrix of the documents
M1(D) The document representation matrix by TRBoW1 M2(D) The document representation matrix by TRBoW2
Table 1

Table of notations.

Suppose that D={d1,d2,,dv} is the collection of all the documents in the corpus, where v is the document size. W={w1,w2,,wn} denotes all the non-repetitive words in the document corpus, also called basis terms, where n is the vocabulary size. Take the universe as W. For any document diD, let di be the representation vector of di, which is n-dimensional and defined as di=[woi1,woi2,,woij,,woin], where

woij=1ifwjdi0otherwise.(3)

Let cij denote the times of words wi and wj occurring in the same document. For a positive threshold θ, the uncertainty function Iθ projects each word into a vector as

Iθ(wi)=A(ci1)A(ci2)A(cin),(4)
where A(cij) (1in,1jn) is defined as
A(cij)=1ifcijθ,ori=j0otherwise.(5)

Then the uncertainty matrix Iθ(W) is defined as follows:

Iθ(W)=1A(c12)A(c1n)A(c21)1A(c2n)A(cn1)A(cn2)1.(6)

For any two documents di and dj, the vague inclusion function is defined as

ν(di,dj)=didjdiI,(7)
where I is an n-dimensional unit vector. Then we define the fuzzy membership function μ for wjW,diD as
μij=ν(Iθ(wj),di)=Iθ(wj)diIθ(wj)I.(8)

And the fuzzy membership matrix of the whole corpus is represented by

μ(W,D)=μ11μ12μ1nμ21μ22μ2nμv1μv2μvn.(9)

Then the upper approximation U(,di) and the lower approximation (,di) of any diD are expressed as

U(,di)={wjWμij>0},(10)
(,di)={wjWμij=1}.(11)

Hence the upper approximation matrix U(,D) and lower approximation matrix L(,D) of the set of documents D are respectively as follows:

U(,D)=u11u12u1nu21u22u2nuv1uv2uvn,(12)
L(,D)=l11l12l1nl21l22l2nlv1lv2lvn,(13)
where
uij=1ifμij>00otherwise,(14)
lij=1ifμij=10otherwise.(15)

As for the original BoW model, the representation matrix is defined as

X=x11x12x1nx21x22x2nxv1xv2xvn,(16)
where xij denotes the number of occurrence times of the word wj in the document di.

Now, we propose two different tolerance rough set-based BOW models, called as TRBoW1 and TRBoW2 according to different weight calculation methods. In the TRBoW1 model we strengthen the weights of words exactly contained in the document. Thus the representation matrix is represented by

M1(D)=m11m12m1nm21m22m2nmv1mv2mvn,(17)
where
mij=xij×μijifwjdiμijifwjU(,di)di0ifwjU(,di).(18)

In the TRBoW2 model we take the membership matrix as the document representation matrix directly

M2(D)=μ(W,D)=μ11μ12μ1nμ21μ22μ2nμv1μv2μvn.(19)

Algorithm 1 is the detailed procedure of the proposed methods.

Algorithm 1: Tolerance Rough Set-Based Bag-of-Words Models

Input: A set of text data including m classes K={k1,k2,,km} and in total v documents D={d1,d2,,dv}; the class label ki of each document di.

Parameters: The co-occurrence threshold: θ; the dimensionality of vectors for text representation: h.

Output: Document representation matrix: M.

1: Count the term frequencies of all the words occurring in the corpus, and choose the most frequent l words as the basis terms;

2: Calculate the uncertainty function Iθ(wi) of each basis term according to (4) and get the uncertainty matrix according to (6);

3: The fuzzy membership degree μij of each basis word in document di can be obtained according to (8), 1iv. And the fuzzy membership matrix μ(W,D) is constructed according to μij;

4: Acquire the upper approximation U(,di) of each document diD according to (10). And obtain the upper approximation vector U(,di) of each document in the text corpus;

5: if the BoW model is performed then

6: Let M=X, which is the document representation using the BoW model;

7: else if TRBoW1 model is performed then

8: Apply formula (17) and (18) to compute the representation matrix M1(D), let M=M1(D);

9: else if TRBoW2 model is performed then

10: Let M=M2(D)=μ(W,D), using the membership degree as the weight directly, as shown in formula (19);

11: end

12: return M

Example 1. Assume that the corpus is consisted of the following four documents.

  1. Itwasthebestoftimes.

  2. Itwastheworstoftimes.

  3. Itwastheageofwisdom.

  4. Itwastheageoffoolishness.

The universe is the set of all the basis terms, W={it, was, the, best, of, times, age, worst, wisdom, foolishness} and the vocabulary size is 10. If the BoW model is performed, the document representation matrix is

X=1111110000111011010011101010101110101001.

If the TF-IDF model is performed, the document representation matrix is

T=0.320.320.320.610.320.4800000.320.320.3200.320.4800.61000.320.320.3200.3200.4800.6100.320.320.3200.3200.480.00.61.

If the TRBoW1 model is performed, setting θ=2, we have

I2(W)=1110111000111011100011101110000001000000111011100011101110001110111000000000010000000000100000000001.

Then the membership matrix is

μ(W,D)=565656156145000565656056145100565656056451010565656056451001.

And the upper approximation matrix is

U(,D)=1111111000111011110011101110101110111001.

So the document representation by the TRBoW1 model is

M1(D)=565656156145000565656056145100565656056451010565656056451001.

If the TRBoW2 model is performed, setting θ=2, the document representation matrix is

M2(D)=μ(W,D)=565656156145000565656056145100565656056451010565656056451001.

Because each basis term appears no more than once in each document, the representation matrixes are the same by the TRBoW1 and TRBoW2 model in Example 1. For the practical corpus is large enough, a word often occurs many times in the same document, the representation matrixes by the two models will be different.

As can be indicated from Example 1, the TF-IDF model emphasizes the weight of each word more than the BoW model. Besides, by applying the rough set theory, our methods capture some latent semantic information. For example, the first two sentences are represented by the first two row vectors of the document representation matrixes, respectively. We can see that X17=0 and T17=0 but M17=45. In other words, the membership degree of the word “age” belonging to the first two documents is 0 in the BoW and TF-IDF model, however the membership degree of the word “age” belonging to the first two documents is 45 in the TRBoW1 and TRBoW2 model. Hence our models mine the meaning of “age” that may exist in the first two documents potentially.

In the BoW model, the document representation of the first document is expressed as (1111110000); in the TRBoW1 and TRBoW2 model, the document representation of the first document is expressed as (565656156145000). It can be seen that the sparsity of BoW model and TF-IDF model is slightly alleviated and the document matrixes become denser. Furthermore, if we set θ as 1, the document representation matrix of our models is

353535135147562323353535563514712323353535233547123156353535233547123561.

It is evident from the matrix above that the sparsity of BoW model is improved considerably. In fact, the sparsities of document representation matrixes in the TRBoW1 and TRBoW2 model are depended on the value of θ, that is, the larger the threshold is, the sparser the representation matrixes are.

4. EXPERIMENTS

In this section, in order to evaluate the performance of our methods, we compare our proposed models with the BoW model, TF-IDF model and average embeddings (AEs) by conducting some popular text classification tasks [1,12]. For better comparison, the dimension of BoW and TF-IDF is also set as 1000. AE is a neural network based method, which represents a document by averaging out the word embedding of each word included in the document. The pretrained word2vec word embeddings are used in the experiments [33], which are 300 dimensional. The other parameter settings are the same.

4.1. Datasets and Preprocessing

In order to verify the validity of our methods and experiments, we will use the commonly used benchmark datasets, Reuters, BBC and BBCsport in text classifications.

Rueters: Reuters is a dataset of news documents generated from the Reuters news website, including 46 categories [34]. Due to the serious imbalance of the document size of per category, we choose the five top categories in our study, a total of 8157 documents. The dataset is beforehand cut into training set with 6533 articles and testing set with 1624 articles. Since we download the data from the toolkit of Keras, and the data is represented in the form of arrays, we firstly converted it into the form of original text.

BBC: BBC dataset is a corpus of 2225 documents containing 5 different categories, which is crawled from the BBC news website [35]. Business, entertainment, politics, sport and tech are respectively the labels of each category. Since the training set and testing set have not been given, we split the corpus into training set with 1225 documents and testing document with 1000 documents randomly.

BBCsport: BBCsport is a collection of 737 documents from the BBCsport website including athletics, cricket, football, rugby and tennis in total of 5 categories [35]. Since the training set and testing set have not been given, we split the corpus into training set with 400 documents and testing set with 337 documents at random.

Table 2 shows the statistical information of the three datasets above. As for the preprocessing of the datasets, the headers and footers of every document have been removed. Then the punctuation and stop words with no practical sense in the document, have also been deleted. The stop words list is obtained from the library sklearn [34]. In addition, we convert all the letters into lower case.

Statistics Reuters BBC BBCsport
Document number 8157 2225 737
Class number 5 5 5
Training/Testing splits 6533/1624 1225/1000 400/337
Table 2

Statistical information of the three datasets.

4.2. Experiment Setup

We carry out the experiments to compare our proposed TRBoW1 model and TRBoW2 model with BoW model and TF-IDF model and AE method. In order to get the best result, we set the parameters of each classifier in a reasonable range, and the program automatically selects an optimal one. Since the most frequent based keyword extraction method (MF) was verified to be more efficient than other keywords extraction methods, such as term frequency inverse sentence frequency (TF-ISF) and co-occurrence statistical information (CSI) [36], we set the most 1000 frequent words in the corpus as the basis terms in all the experiments. Therefore, the representation vectors in this paper are set as 1000 dimensional.

After learning the document representation, the following four different popular machine learning classifiers are applied for classification tasks:

Support vector machine (SVM): In the SVM theory [37], a linear and a nonlinear kernel can be chose to classify linear data and nonlinear data, respectively. We use the linear one, which is called as linear SVM. We set the parameter C as {0.1,1,10} and the parameter gamma as {0.001,0.01,0.1,1,10,100,1000} additionally. Outcomes of the SVM classifier are chose the optimal one from these different parameters by the system.

K-nearest neighbor (KNN): In the KNN algorithm [38], one certain document belongs to the category that the most of the nearest k documents belong to. We search the best parameter K from 2 to 12.

Random forest (RF): RF algorithms [39] are a set of classification and regression trees derived from the guided samples of training data. The connection among trees and the efficiency of a single tree influence generation error of the classifiers. The depth of RF is searched from 6 to 12.

Ridge regression (RR): RR algorithm [40] is a nonlinear partial estimation method. And it is a biased estimator regression method specially used in the analysis of collinear data.

The other hyper-parameters are set as the default values of the system. To make a comprehensive comparison of the performance of our proposed TRBoW1 and TRBoW2 model for document representation, we use the five-fold cross validation in all the experiments to obtain the best result of every single experiment.

4.3. Evaluation Metrics

In this section, we exploit precision rate, recall rate and F-measure to measure the performance of text classification.

Precision rate refers to the proportion of the number of texts to the total number of texts correctly classified by the classifier, which is defined as

pr=aa+b×100%,(20)
where a denotes the number of texts that correctly classified into a category, and b means the number of texts that incorrectly classified into a category.

Recall rate refers to the proportion of the number of correct texts classified to all documents of the category, which is defined as

re=aa+c×100%,(21)
where c represents the quantity of texts that incorrectly excluded from a category.

Precision rate and recall rate are contradictory measures. In general, the improvement of precision rate will lead to the depression of recall rate, and vice versa. So F_measure, harmonic mean of precision and recall, is proposed to make the comprehensive assessment, given as

F_measure=2×pr×repr+re.(22)

4.4. Experimental Results

The experimental results are illustrated in this subsection. The experimental results of precision, recall rate and F_measure of document categorization on the BBCsport dataset obtained by the original BoW model, TF-IDF model, AE method and the proposed TRBoW1 and TRBoW2 are respectively described in Tables 35. Table 6 describes the classification precision of the four methods and four classifiers on the BBC dataset. Table 7 shows the recall and Table 8 shows the F_measure. Tables 911 present the study results on the dataset of Reuters. They are individually the precision, recall and F_measure.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 78.64 86.65 92.56 92.58 92.58
RR 96.44 96.71 97.02 97.33 97.63
RF 94.96 93.18 89.91 95.85 95.25
SVM 97.63 97.33 96.44 98.22 97.63
Table 3

Precision (%) of four methods on the BBCsport dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 80.16 86.84 93.65 93.59 94.04
RR 96.70 96.66 97.20 97.70 98.07
RF 95.28 95.18 90.33 96.34 95.60
SVM 97.79 96.37 97.51 98.37 98.18
Table 4

Recall (%) of four methods on the BBCsport dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 79.37 87.03 93.11 93.08 93.31
RR 96.57 96.68 97.11 97.52 97.85
RF 95.56 95.56 90.12 96.09 95.72
SVM 97.71 96.85 96.97 98.30 97.90
Table 5

F_measure (%) of four methods on the BBCsport dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 69.30 69.60 94.70 91.80 94.10
RR 94.00 88.60 95.60 95.20 95.80
RF 91.40 92.10 94.20 93.80 94.40
SVM 94.20 94.80 96.40 96.10 95.30
Table 6

Precision (%) of four methods on the BBC dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 66.66 66.82 92.64 90.98 93.66
RR 93.65 87.85 95.49 94.90 95.54
RF 91.43 91.47 93.87 94.08 93.96
SVM 93.82 94.40 96.22 95.85 95.34
Table 7

Recall (%) of four methods on the BBC dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 67.96 68.18 93.66 91.39 93.88
RR 93.82 88.22 95.55 95.05 95.67
RF 91.76 91.78 94.03 93.54 94.18
SVM 94.01 94.60 96.31 95.47 95.16
Table 8

F_measure(%) of four methods on the BBC dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 74.97 68.84 76.97 80.36 79.43
RR 86.51 86.58 86.45 86.02 87.99
RF 79.06 79.56 74.88 87.07 86.15
SVM 88.18 90.21 86.45 89.10 88.55
Table 9

Precision (%) of four methods on the Reuters dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 63.97 54.77 67.28 70.38 67.36
RR 78.15 80.06 76.07 77.78 81.92
RF 63.76 63.85 48.39 79.24 78.24
SVM 81.88 84.48 78.67 84.66 83.62
Table 10

Recall (%) of four methods on the Reuters dataset.

BoW TF-IDF AE TRBoW1 TRBoW2
KNN 69.02 61.01 71.80 75.04 72.90
RR 82.12 83.19 80.93 81.69 84.85
RF 70.59 70.84 58.78 82.97 82
SVM 84.91 87.25 82.38 86.82 86.01
Table 11

F_measure (%) of four methods on the Reuters dataset.

5. DISCUSSIONS

In this section, we discuss the precision rate, recall rate and F-measure of the BoW model, TF-IDF model, AE method and our proposed TRBoW1 model and TRBoW2 for document representations on document categorization tasks by utilizing the KNN, RF, SVM and RR classifier on the BBCsport, BBC and Reuters corpus.

5.1. Performance on the BBCsport Dataset

As can be observed in all the tables of BBCsport dataset, in the comparison of these methods, the optimal results are labeled with bold fonts. And for the TRBoW1 and TRBoW2 method, the tables present the optimal results for the parameter θ from 4 to 20. For the reason that TF-IDF model takes the importance degree of each word in the documents into consideration, it has better performance compared to the BoW model, as shown in the tables. According to these tables, TRBoW1 model and TRBoW2 model both have better performance than BoW, TF-IDF and AE in all the four classifiers. Remarkably, the precision rate has increased by 13.94%, the recall rate has increased by 13.88%, and the precision rate has increased by 13.94% by using KNN. As a whole, the TRBoW1 model results in a greater performance when using RF and SVM as the classifier, and the TRBoW2 model results in a greater performance when using KNN and RR as the classifier.

5.2. Performance on the BBC Dataset

Data emerged in Tables 68 are searched the best results, when parameter θ ranges from 8 to 18. Compared with BoW, TF-IDF and AE, our methods achieves 96.10% in precision, 95.85% in recall and 95.67% in F_measure, especially the precision rate of TRBoW2 has increased by 24.91% than the BoW model, having significant improvements. It can be seen that our methods have captured more latent semantics than the BoW method.

5.3. Performance on the Reuters Dataset

Data in Tables 911 display that the TRBoW1 model outperforms the other models by the KNN, RF and SVM classifiers, and the TRBoW2 model outperforms the other models by the RR classifier. Because the theory of each classifier is different, it is acceptable that the experimental results of different classifiers are a little various. On account that we do not employ all the data of Reuters, the advantage of semantic mining may be not fully displayed. So compared with other datasets, performances of Reuters dataset do not have a more remarkable increment.

5.4. Co-occurrence Degree Threshold Value

In the tolerance rough set model, the co-occurrence threshold value θ determines the tolerance space. Hence the selection of the parameter θ has a tremendous impact on the performance of document representation. When the value of θ is too high, the upper approximation space may be too small, which will lead to the insufficient exploiting of potential relations behind documents. In the contrary, when the value of θ is too low, the upper approximation space may be too large, which will increase redundancy and noisy information among documents. Generally speaking, the selection of the value of θ relies on the corpus size and the vocabulary size. The larger the sizes, the higher the value is.

Figures 1 and 2 display the impacts of the value of θ on the precision and F_measure, which are the experimental results of our TRBoW1 and TRBoW2 model using the classifiers aforementioned on the BBCsport corpus in our study. The x label value of θ ranges from 4 to 16. Similarly, the influence on the BBC corpus are illustrated in Figures 3 and 4. The x label value of θ ranges from 8 to 18. As indicated in these figures, it is apparent that in majority cases, the trend of the curves generally satisfies the regular inferred above. On account that results predicted by RF have some randomness, it is explainable when some points are not satisfactory.

Figure 1

Precision using different classifiers with different threshold value θ.

Figure 2

F_measure using different classifiers with different threshold value θ.

Figure 3

Precision using different classifiers with different threshold value θ.

Figure 4

F_measure using different classifiers with different threshold value θ.

6. CONCLUSION

In this paper, we have proposed two novel document representation learning models, TRBoW1 model and TRBoW2 model, which adopt the tolerance rough set model to improve the traditional BoW model. The proposed TRBoW1 model and TRBoW2 model extend each document to its upper approximation. The extended upper approximation can mine the latent semantic relations behind documents, which solves the problems of lacking of latent semantics and the sparsity of the original BoW model. They can learn the document representation without any training or prior knowledge. The experiments have carried out on various document representation methods for text classification on different datasets using classifiers including KNN, RR, RF and SVM. The results of the experiments indicate that the performances have been improved remarkably and allow us to obtain the following conclusions: the highest F_measure of BBCsport is up to 98.30%; the proposed representation methods enrich the representation of BoW, making the improvement in performance up to 27%.

Except for text categorization tasks, the TRBoW1 and TRBoW2 model can be applied in other domains such as information retrieval and document clustering, which will be further studied in the future. Besides, we apply it in the sentence similarity calculation on the basis of this work.

CONFLICT OF INTEREST

The authors declare no conflict of interest.

AUTHORS' CONTRIBUTIONS

These authors contributed equally to this work.

ACKNOWLEDGMENTS

The authors would like to thank the referee for his valuable remarks. This work was supported by The National Natural Science Foundation of China (Grant no. 11671001).

REFERENCES

9.R. Basili, A. Moschitti, and M.T. Pazienza, A text classifier based on linguistic processing, Citeseer, in Proceedings of IJCAI99, Machine Learning for Information Filtering (Stockholm, Sweden), 1999, pp. 1-6.
10.C. Qiyue, Structure entropy weight method to confirm the weight of evaluating index, Syst. Eng. Theory Pract., Vol. 30, 2010, pp. 1225-1228.
22.Y. Liu, Z. Liu, T.S. Chua, and M. Sun, Topical word embeddings, AAAI Press, in Twenty-Ninth AAAI Conference on Artificial Intelligence (Austin, Texas, USA), 25–30 January 2015, pp. 2418-2424.
23.R. Kiros, Y. Zhu, R.R. Salakhutdinov, R. Zemel, R. Urtasun, A. Torralba, and S. Fidler, Skip-thought vectors, in Advances in Neural Information Processing Systems (Montreal, Canada), 2015, pp. 3294-3302.
39.A.S. More, D.P. Rana, I. Agarwal, and S. Vallabhbhai, Random forest classifier approach for imbalanced big data classification for smart city application domains, Int. J. Comput. Intell. Syst., Vol. 1, 2020, pp. 260-266.
Journal
International Journal of Computational Intelligence Systems
Volume-Issue
13 - 1
Pages
1218 - 1226
Publication Date
2020/08/19
ISSN (Online)
1875-6883
ISSN (Print)
1875-6891
DOI
10.2991/ijcis.d.200808.001How to use a DOI?
Copyright
© 2020 The Authors. Published by Atlantis Press B.V.
Open Access
This is an open access article distributed under the CC BY-NC 4.0 license (http://creativecommons.org/licenses/by-nc/4.0/).

Cite this article

TY  - JOUR
AU  - Dong Qiu
AU  - Haihuan Jiang
AU  - Ruiteng Yan
PY  - 2020
DA  - 2020/08/19
TI  - Tolerance Rough Set-Based Bag-of-Words Model for Document Representation
JO  - International Journal of Computational Intelligence Systems
SP  - 1218
EP  - 1226
VL  - 13
IS  - 1
SN  - 1875-6883
UR  - https://doi.org/10.2991/ijcis.d.200808.001
DO  - 10.2991/ijcis.d.200808.001
ID  - Qiu2020
ER  -