Combining semantic and term frequency similarities for text clustering

A key challenge for document clustering consists in finding a proper similarity measure for text documents that enables the generation of cohesive groups. Measures based on the classic bag-of-words model take into account solely the presence (and frequency) of words in documents. In doing so, semantically similar documents which use different vocabularies may end up in different clusters. For this reason, semantic similarity measures that use external knowledge, such as word n-gram corpora or thesauri, have been proposed in the literature. In this paper, the Frequency Google Tri-gram Measure is proposed to assess similarity between documents based on the frequencies of terms in the compared documents as well as the Google n-gram corpus as an additional semantic similarity source. Clustering algorithms are applied to several real datasets in order to experimentally evaluate the quality of the clusters obtained with the proposed measure and compare it with a number of state-of-the-art measures from the literature. The experimental results demonstrate that the proposed measure improves significantly the quality of document clustering, based on statistical tests. We further demonstrate that clustering results combining bag-of-words and semantic similarity are superior to those obtained with either approach independently.

1. For the sake of compactness, Table 2 assumes the relaxed implementation of FGTM (Method 2). If the greedy implementation (Method 1) is used instead, just replace terms \(T'^2\) with \(T'^2\log (T')\) in the respective complexities.

2. As a practical example, it is worth reporting that the WR method [39] took about 18 days to compute all the pairwise document similarity values for the largest dataset used in our experiments (20ng-whole, with \(\approx \)19K documents), on a DELL machine i7 3.60GHz with 16GB RAM. This is about 300x as much time as required by our proposed measure, FGTM.

3. https://radimrehurek.com/gensim/ (accessed on December 18, 2017).

4. available at https://github.com/jhlau/doc2vec (on December 18, 2017).

5. available at https://github.com/PrincetonML/SIF (on December 18, 2017).

6. http://qwone.com/~jason/20Newsgroups/(on August 23, 2016).

7. https://vhasoares.github.io/downloads.html (on January 08, 2018).

8. http://vis.icmc.usp.br/vicg/tool/7/data-sets-from-papers (on January 08, 2018).

9. http://www.ncbi.nlm.nih.gov/pubmed (on January 08, 2018).

10. http://www.daviddlewis.com/resources/testcollections/reuters21578/  (on August 23, 2016).

11. http://www.scopus.com/.

12. http://www.dt.fee.unicamp.br/~tiago/smsspamcollection/ (on August 23, 2016).

13. http://www.cs.cmu.edu/afs/cs.cmu.edu/project/theo-20/www/data/ (on August 23, 2016).

14. For example, the stopword list used in this work contains 318 words.

15. The number of fuzzy c-means iterations is defined as 50 in this work.

16. Although the consensus method is not part of the original LDC algorithm, it may be implemented as an additional step. This is done in the present work.

The authors acknowledge the Brazilian Research Agencies Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES)—Finance Code 001, CNPq, FAPEMIG and FAPESP, the Natural Sciences and Engineering Research Council of Canada, the Boeing Company, CALDO, and the International Development Research Centre, Ottawa, Canada, for their financial support to this work.

Authors and Affiliations

1. Department of Informatics, Federal University of Viçosa (UFV), Viçosa, MG, 36570-900, Brazil

   Victor Hugo Andrade Soares & Murilo Coelho Naldi

2. School of Mathematical and Physical Sciences, The University of Newcastle, University Drive, Callaghan, NSW, 2308, Australia

   Ricardo J. G. B. Campello

3. Institute of Mathematics and Computer Sciences, University of São Paulo (USP), Av. Trabalhador São-Carlense, 400, São Carlos, SP, 13560-970, Brazil

   Victor Hugo Andrade Soares & Ricardo J. G. B. Campello

4. Faculty of Computer Science, Dalhousie University, 6050 University Ave., Halifax, NS, B3H1W5, Canada

   Seyednaser Nourashrafeddin & Evangelos Milios

5. Department of Computer Science, Federal University of São Carlos (UFSCar), São Carlos, SP, 13565-905, Brazil

   Murilo Coelho Naldi

Corresponding author

Correspondence to Victor Hugo Andrade Soares.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendix A: Proofs: Propositions

The proofs related to the Proposition 1 (\(S(P, R) \in [0, 1]\)) and Proposition 2 (the Greedy and Relaxed methods preserve \(S(P, R) \in [0, 1]\)), previously described in Sect. 3.2, are presented here. To do so, we need to show that the statement \(0 \le S(P, R) \le 1\) is true. Based on the definition of Eq. 6, we must prove that:

In order to prove what is stated at Inequality 10, we divided it into Inequalities 11 and 13 and developed them separately, for the sake of clarity. Starting with Inequality 11:

We must assume that both documents are not empty and contain terms, i.e., \(p_{\mathrm{Total}} > 0\) and \(r_{\mathrm{Total}} > 0\) are true. Thus, demonstrate Eq. 11 is equivalent to demonstrate that \(S_{\mathrm{SF}}(P,R) \ge 0\) is true. That is equivalent to prove that the following statement is true, based on the definition of \(S_{\mathrm{SF}}(P,R)\) in Eq. 4:

where the similarity \(\hbox {Sim}(p_i, r_j)\) is in the range [0, 1] (as defined in Eq. (1)), the frequency of a given document term is not negative (\(\hbox {freq}(\cdot ) \ge 0\)) and \(\delta (i,j) = 1\) if the term \(p_i\) is associated with the term \(r_j\) and \(\delta (i,j) = 0\) otherwise. Thus, every component of the left side of Inequality 12 is not negative, which makes \(S_{\mathrm{SF}}(P,R) \ge 0\) true and, consequentially, \(S(P, R) \ge 0\) is also true. The lower bound occurs for \(S_{\mathrm{SF}}(P,R) = 0\) (when there are no semantic relation between the terms of the documents), which results into \(S(P, R) = 0\). This holds for Propositions 1 and 2 of Sect. 3.2.

Using the definition in Eq. 6, the proof for \(S(P, R) \le 1\) is equivalent to demonstrate that:

which is equivalent to:

Inequality 13 is true if \(S_{\mathrm{SF}}(P,R) \le \hbox {min}(p_{\mathrm{Total}}, r_{\mathrm{Total}})\). Thus, replacing \(S_{\mathrm{SF}}(P,R)\) by its definition in Eq. 4, we have to prove that:

Once more, it is important to note that \(\hbox {Sim}(p_i, r_j)\) is in the range [0, 1] [as defined in Eq. (1)], and thus the following inequality is true:

Inequality 14 can be simplified based on Inequality 15, which results in Inequality 16.

Replacing \(p_{\mathrm{Total}}\) and \(r_{\mathrm{Total}}\) in Inequality 16 by their calculations results in Inequality 17.

The \(\delta (i,j)\) values result from the assignment Problem (2), described in Sect. 3.2, i.e., \(\delta (i, j)= 1\) if \(p_i\) is optimally associated with \(r_j\), or \(\delta (i, j) = 0\) otherwise. The FGTM assignment defines that each term \(p_i \in P\) is associated with at most one term \(r_j \in R\) and vice versa, which is a constrain of the assignment problem. As in the exact solution, the Greedy Strategy, proposed in Sect. 3.2, satisfies that constrain. Thus, \(\sum _{i=1}^m \sum _{j=1}^n \delta (i,j) \le \hbox {min}(m,n)\) is true, which means that the maximum number of associated pairs for FGTM must be the number of terms in the document with fewer terms. Only associated terms, \(\delta (i,j) = 1\), contribute to the sum in the left side of Inequality 17. Thus, not associated terms, \(\delta (i,j) = 0\), can be removed from the sum of Inequality 17 without loss, which result in Inequality 18.

Analyzing Inequality 18, we have the sum of minor frequencies of the \(\hbox {min}(m,n)\) best associated terms at the left side, while there is the minor sum of the frequencies of the terms of the documents in the right side. By definition, for each assigned term frequency value added at the sum of left side of the inequality, frequencies of equal or superior values (the term and its associated term) are added at the sums of the right side. Thus, the Inequality 18 is true, which means that Inequality 13 is also true.

Although the previous proofs are enough to show that Inequality 10 is true if the Exact Method (Proposition 1) and for the Greedy Strategy (Proposition 2) are used for the FGTM calculation, it is not enough for the Relaxed Strategy, as it relaxes the assignment constraint of the association problem (that a term cannot be associated with more than one term). As described in Sect. 3.2, the Relaxed Strategy computes the similarities between P and R and vice versa, selecting the lower result. The two steps of this strategy follow the Inequalities 19 and 20 when comparing \(S_{\mathrm{SF}}(P, R)\) and \(S_{\mathrm{SF}}(R, P)\), respectively:

where \(r_j = \mathrm{arg} \max _{r_j} \hbox {Sim}(p_i, r_j)\). By definition, \(\hbox {Sim}(p_i, r_j) \in [0,1]\) and \(p_{\mathrm{Total}} = \sum _{i=1}^m \hbox {freq}(p_i)\), thus the Inequality 19 is true. Likewise, \(S_{\mathrm{SF}}(R, P)\) is computed as follows:

where \(p_i = \mathrm{arg} \max _{p_i} \hbox {Sim}(p_i, r_j)\). Analogously, \(\hbox {Sim}(p_i, r_j) \in [0,1]\) and \(r_{\mathrm{Total}} = \sum _{j=1}^n \hbox {freq}(r_j)\), thus the Inequality 20 is also true.

The result of the Relaxed Strategy is the lower value between \(S_{\mathrm{SF}}(P, R)\) and \(S_{\mathrm{SF}}(R, P)\), i.e., \(\hbox {min}\{S(P,R), S(R,P)\}\). Therefore, as \(S_{\mathrm{SF}}(P, R) \le p_{\mathrm{Total}}\) and \(S_{\mathrm{SF}}(R, P) \le r_{\mathrm{Total}}\), the \(\hbox {min}\{S_{\mathrm{SF}}(P,R), S_{\mathrm{SF}}(R,P)\} \le \hbox {min}(p_{\mathrm{Total}}, r_{\mathrm{Total}})\), which satisfies the Inequality 13. Thus, the Proposition 2 is also true for Relaxed Strategy.

For all strategies, the upper bound of S(P, R) is assessed when \(S_{\mathrm{SF}}(P,R) = p_{\mathrm{Total}} = r_{\mathrm{Total}}\), which means that all the terms of both documents have maximum similarity, i.e., \(\hbox {Sim}(p_i, r_j) = 1\) for each associated term and all terms in both documents are associated. This scenario happens when both documents are equivalent, syntactically or semantically. In this case, we can assume that \(x = S_{\mathrm{SF}}(P,R) = p_{\mathrm{Total}} = r_{\mathrm{Total}}\) and replace them at Eq. 6, which results:

Appendix B: Detailed computational complexity analysis

Here is presented the estimation of the asymptotic complexity of the algorithms in Table 2.

1.1 B.1 Preprocessing complexity analysis

Two preprocessing methods considered in this present paper are BOW and BOC. In BOW, the preprocessing performs \(\sum _i^N{T_i}\) operations to separate and count the terms, where N is the number of documents and \(T_i\) is the number of terms (including repetitions) contained in the \(i^{th}\) document. To remove the stopwords, \(S \cdot \log {T'}\) operations are performed, where S is the stopword list length^{Footnote 14} and \(T'\) is the total number of terms, without repetitions, in the document set. The computational complexity of TF–IDF normalization is \(O(NT')\). The VAR-TFIDF feature selection [26] also has complexity \(O(NT')\). Thus, the total complexity of the BOW preprocessing is \(O((T_1 + T_2 + \cdots + T_N) + (S \cdot \log {T'}) + (NT') + (NT'))\), which is equivalent to \(O(N\bar{T} + NT')\), where \(\bar{T}\) is the average number of terms in the document set. As the search and removal of stopwords are performed in logarithmic time and the number N of documents in a collection is usually greater than the list with S stopwords, the cost \(S \cdot \log {T'}\) is omitted in the notation O. Since the values of \(\bar{T}\) and \(T'\) depend on the characteristics of the document set, it is not possible to affirm that one is greater than the other. In a set of short documents, but with high diversity of subjects and vocabulary, the number of terms \(T'\) is usually much greater than the average size of the documents, \(\bar{T}\). On the other hand, in a set of large documents, but with little diversity of subjects and small vocabulary, the average size of the documents, \(\bar{T}\), may be greater than the number of terms, \(T'\).

In BOC method, the Wikipedia Miner toolkit [41] is used to identify concepts in the documents. Given as input the \(i^{th}\) document with length \(T_i\), the Wikipedia Miner toolkit builds a list with all possible n-grams in the document. A n-gram is a sequence of words found in the text which is not separated by punctuation mark or line break. For example, in the sentence “I love cookies,” the following n-grams are obtained:

• “I love cookies”

• “I love”

• “love cookies”

• “I”

• “love”

• “cookies”

As there is no maximum value of n, in the worst case, where in the document there is no punctuation mark or line break, the n-gram list will contain, approximately, the following number of n-grams:

After constructing the n-gram list, the Wikipedia Miner Toolkit searches for related concepts to each of the n-grams found. The tool uses a hash map structure to query the concepts. Assuming the complexity to query concepts approaches O(1) [28] and that the tool measures the relationship among all concepts found, the complexity for measuring the relationship among all concepts in the \(i^{th}\) document is \(O(C_i^2)\), where \(C_i\) is the number of concepts found in the \(i^{th}\) document. After computing the relationship, the tool iterates over and removes the concepts with lower weight using a user-defined threshold. The removal of concepts has complexity \(O(C_i)\).

After using the Wikipedia Miner Toolkit for all N documents, the document-concept table is constructed and normalized by TF–IDF method. This operation has complexity \(O(NC')\), where \(C'\) is the number of relevant concepts throughout the set document. Thus, the total complexity of the preprocessing by BOC method is estimated as \(O((T_1^2 + T_2^2 + \cdots + T_N^2) + (C_1^2 + C_2^2 + \cdots + C_N^2) + (C_1 + C_2 + \cdots + C_N) + (NC'))\). As the size of each document may vary, it is not possible to simplify the complexity. If we assume that all documents have the same size and number of concepts, the complexity can be represented as \(O((N\bar{T}^2) + (N\bar{C}^2) + (N\bar{C}) + (NC'))\), where \(\bar{T}\) and \(\bar{C}\) are the averages of the numbers of words and concepts, respectively, in the document set. As the number of concepts in a document is lower than the number of words in the same document, it follows that \(N\bar{T}^2 \ge N\bar{C}^2 \ge N\bar{C}\). Thus, \((N\bar{C}^2)\) and \((N\bar{C})\) are omitted in the notation O. As with the BOW model, it is not possible to say that \(\bar{T}^2 \ge C'\); therefore, the BOC preprocessing complexity is estimated as \(O(N\bar{T}^2 + NC')\).

1.2 B.2 Similarity measures complexity analysis

Considering that m and n are the number of terms in the documents P and R, respectively, and \(T = |P \cup R|\), that is, T is the number of terms of the union of P with R, the similarity between two documents using the Euclidean distance is calculated with \(2 \cdot T\) operations. Therefore, the asymptotic complexity is O(T). For the similarity calculation using cosine, \(3 \cdot T\) operations are performed, and thus, the asymptotic complexity is also O(T).

For both, GTM and FGTM measures, the first step is to construct a table with the similarity among all terms. For the calculation of similarity between terms, queries of uni-grams and tri-grams are made for each pair of words of the Google corpus. The corpus is structured into a hash map, and therefore, the computational complexity for queries approaches O(1) [28]. The complexity to construct the term-similarity table is \(O(T^2)\). Once the term-similarity matrix is created, the measures compute the document similarity between P and R. For the GTM measure, the worst case is when \(m = n = \frac{T}{2}\), which is when P and R have the same number of terms, all of them are different. In this case, the GTM calculation between documents is performed in \(4 \cdot \left( \frac{T}{2}\right) ^2 + 3 \cdot \frac{T}{2}\) operations. Thus, this asymptotic complexity can reduced to \(O(T^2)\). Unlike GTM, the FGTM’s worst case is when \(m = n = T\), that is, the documents are the same. When using the greedy implementation strategy (Method 1), FGTM performs \(T^2\log {T} + T^2\) operations; thus, the asymptotic complexity is \(O(T^2\log {T})\). When using the relaxed implementation strategy (Method 2), FGTM performs \(2 \cdot T^2 + T\) operations; thus, its asymptotic complexity is \(O(T^2)\).

1.3 B.3 Clustering algorithms complexity analysis

In the document clustering phase of this paper, we use the LDC and ELSDC algorithms. We also perform experiments using the consensus method in the LDC algorithm, in order to merge clusterings obtained by two different similarity measures.

In the LDC algorithm, the term-clustering and key-term selection phases have complexity \(O(NT'K^2I)\), where N and \(T'\) are the number of documents and terms in the dataset, respectively, K is the number of clusters and I is the number of iterations^{Footnote 15} of the fuzzy c-means algorithm [43]. Finding seed documents has complexity \(O(NT'K + NKI)\). Document clustering is performed in two steps: First, the document-centroids are computed, which has complexity \(O(NT'K)\); second, the distances between all N documents to all K document-centroids are measured. When Euclidean distance or cosine similarity is adopted, this step has complexity \(O(NT'K)\), which results in the total complexity of \(O(NT'K)\) for this phase [43]. Alternatively, when using the measures GTM or the relaxed implementation of the FGTM, the complexity of this phase rises to \(O(NT'^2K)\) in the worst case. For the greedy implementation of the FGTM, the complexity is \(O(NT'^2\log {(T')}K)\) in the worst case.

In the ELSDC algorithm, the term-clustering and key-term selection phases remain the same LDC and have complexity \(O(NT'K^2I)\). Then, for each extracted key-term, the algorithm queries and extracts relevant Wikipedia concepts. This step has complexity \(O(T')\). Once done, the seed documents of the BOW and BOC models are found. In this phase, the BOW model has complexity \(O(NT'K + NKI)\) and the BOC has complexity \(O(NC'K + NKI)\). In the next phase, the centroids for both models, BOW and BOC, are computed. The distances between all N documents to the K centroids from BOW and BOC are also measured. Considering the BOW model, the complexity is \(O(NT'K)\) with Euclidean or cosine measures and \(O(NT'^2K)\) for GTM and the relaxed implementation for the FGTM. For the greedy implementation of the FGTM, the complexity is \(O(NT'^2\log {(T')K}\). Considering the BOC model, the complexity is \(O(NC'K)\), where \(C'\) is the total number of concepts in the document set.

The consensus method is implemented as the last step of ELSDC.^{Footnote 16} Firstly, the consensus method finds the documents with the same label in both clusterings. This task has complexity O(N). After that, these documents are considered a training set for a Naive Bayes classifier, which is used to classify the remaining documents. The training and classification of the Naive Bayes have linear complexity, estimated as \(O(NT')\) [35]. Therefore, the asymptotic complexity of the consensus method is \(O(NT' + N)\), which is equivalent to \(O(NT')\).

Following these asymptotic complexities, the analyses of the variants implemented in this present work are made as follows:

1. 1.

   Euclidean this variant uses BOW preprocessing and clustering by Euclidean distance in the LDC algorithm. Complexity: \(N\bar{T} + NT' + NT'K^2I + 2NT'K + NKI = O(N\bar{T} + NT'K^2I)\), because \(NT'K^2I\) > \(NT'K\), NKI and \(NT'\).

2. 2.

   Cosine this variant uses BOW preprocessing and clustering by distance (cosine similarity) in the LDC algorithm. It has the same complexity of “Euclidean” variant, therefore: \(O(N\bar{T} + NT'K^2I)\).

3. 3.

   GTM this variant uses BOW preprocessing and clustering by semantic (GTM) in the LDC algorithm. Complexity: \(O(N \bar{T} + NT' + NT'K^2I + NT'K + NKI + NT'^2K) = O(N \bar{T} + NT'^2K)\), because \(T'\) > K, thus, \(NT'^2K\) > \(NT'K^2I\), \(NT'K\), NKI and \(NT'\).

4. 4.

   FGTM (relaxed) this variant uses BOW preprocessing and clustering by semantic (FGTM relaxed implementation) in the LDC algorithm. It has the same complexity of “GTM,” therefore: \(O(N \bar{T} + NT'^2K)\).

5. 5.

   FGTM (greedy) this variant uses BOW preprocessing and clustering by semantic (FGTM greedy implementation) in the LDC algorithm. The complexity of the greedy compared to the relaxed implementation is just replace terms \(T'^2\) with \(T'^2\log {T'}\), therefore: \(O(N \bar{T} + NT'^2\log {T'}K)\).

6. 6.

   Euc \(+\) Wik this variant uses BOW and BOC preprocessing. The clustering is done by Euclidean distance in the ELSDC algorithm. Complexity: \(O(N\bar{T} + NT' + N\bar{T}^2 + NC'+ NT'K^2I + T' + 2NT'K + 2NKI + 2NC'K + NT') = O(N\bar{T}^2 + NT'K^2I)\), because usually \(T'\) > \(C'\) and I.

7. 7.

   Euc \(+\) FGTM this variant uses BOW preprocessing, clustering by Euclidean distance and semantic (FGTM) in the LDC algorithm. For the FGTM will be consider the relaxed implementation. The results are merged by the consensus method. Complexity: \(O(N\bar{T} + NT' + NT'K^2I + 2NT'K + NKI + NT'^2K + NT') = O(N\bar{T} + NT'^2K)\).

8. 8.

   Cos \(+\) FGTM this variant uses BOW preprocessing, clustering by distance (cosine similarity) and the semantic FGTM measure (considering the relaxed implementation) in the LDC algorithm. The results are merged by the consensus method. The complexity is the same of “Euc \(+\) FGTM,” therefore: \(O(N\bar{T} + NT'^2K)\).

9. 9.

   Wik \(+\) FGTM this variant uses BOW and BOC preprocessing. The clustering is done by the semantic FGTM measure (considering the relaxed implementation) in the ELSDC algorithm. Complexity: \(O(N\bar{T} + NT' + N\bar{T}^2 + NC' + NT'K^2I + T' + NT'K + 2NKI + 2NC'K + NT'^2K + NT') = O(N\bar{T}^2 + NT'^2K)\).

Comparison of the average NMI results obtained by the Greedy and Relaxed implementations of FGTM for all eighteen document sets

Full size image

Appendix C: Supplementary results: relaxed strategy

Experimental results using FGTM are done with two strategies described in Sect. 3.2: Greedy and Relaxed. Both methods have very similar results, as illustrated in Fig. 2, where the NMI mean values and standard deviations for stand-alone FGTM are presented. Considering these values for each document set, it is noticeable that the Greedy implementation outperforms the relaxed implementation only for the cbr-ilp-ir-son and cbr-ilp-ir-son-int sets. For all other 16 document sets, the results are not significantly different. Thus, for the sake of simplicity and compactness of the paper, we reported only the clustering results obtained by the Greedy Strategy (Method 1) in Sect. 6.

Here, the results obtained by the Relaxed Strategy (Method 2) are presented. Table 5 shows the NMI mean and standard deviation values for the relaxed stand-alone FGTM and its combinations with Euclidean, cosine and BOC (using Wikipedia concepts) by consensus method. The t-test was applied with \(95\%\) confidence to compare the results of the Relaxed Strategy (Table 5) and the results of the Greedy Strategy (Table 4). When the null hypothesis is rejected, there is statistical evidence to support that the compared results are different and the result is presented in bold in Table 5.

The absolute values of NMI are very similar for the vast majority of the datasets. Presented in bold in Table 5, there are four datasets for which the Relaxed Strategy had a result significantly worst than the Greedy Strategy, according to the results of the statistical test. Even so, the obtained values are close for both strategies.

Reprints and permissions