The paper contains a brief overview of full-text clustering methods for current research directions detection. A novel full-text clustering method is proposed. A dataset is created and experimental results are verified by problem domain “Regenerative medicine” experts with PhD degrees. The proposed method is well applicable for research directions detection according to experimental results. Finally, prospects and drawbacks of the proposed method are discussed. *The research is supported by Russian Foundation for Basic Research, project 14-29-05008-ofi_m
Commonly methods for research directions detection
are based on different clustering methods [
The primary goal of this paper is to present an
improved method for research directions detection,
which can process datasets semi-automatically. Besides,
we present results of comparison between proposed
method and the state-of-the-art clustering methods such
as Birch [19], affinity propagation [
Consider some methods for science directions
detection. There are three groups of methods, used for
research direction detection. All of them use clustering
methods, but differ in the applied measure of similarity
detection: the co-citation measure [
In [
In this work we will use full-text clustering approaches because of their universality: they do not need citation databases for proper results. Let us discuss several clustering methods.
For example, affinity propagation is one of common
clustering methods. In this method a set of papers is
considered as a network of connected nodes, and the
similarity of papers corresponds to the weight of edges
in this network. The method is based on mechanism of
“passing messages” between nodes of the network [
Specialized methods for text clustering are worth
mentioning. In [
The methods of topic modelling based on the
generating models are often applied for full-text
clustering. For example, Latent Dirichlet Allocation
(LDA) [
3.1
The cornerstone of proposed full-text clustering
method is similar document search, previously
discussed in [
tf (w, d ) ni nk paper d , in d .
k k Where ni is number of times term w occurs in the nk is a number of all terms occurrences idf (w, D) log
| D | | di w | | di
Where | D | is the number of papers in corpus D, w | is the number of papers from D that have term w and the logarithm base does not matter.
Define direct index of papers as a function that returns a set of different terms and their weights from the paper d .
Didx(d) { w1,t1 , w2 ,t2 ,.., wn ,tn } Where n is the number of different terms in the document.
Define inverted index of papers as a function that returns a set of different papers, in which given term w is occurred and weights of this term in these papers: Iidx(w) { d1,t1 , d2 ,t2 ,.., dn ,tn } . The algorithm of the fast similar document search is the following. 1. Get terms of the paper d from its direct index Didx(d ) .
Didx(d ). 2. Retrieve list of papers containing the terms got in step 1 from inverted index Iidx(w) . 3. Filter papers that are weakly intersected by the terms of the input paper. 4. Get lists of terms for retrieved papers from
Calculate Manhattan [
In step 3 of this algorithm we cut weakly similar papers without calculating a distance measure that significantly improves the algorithm. Steps 1-4 can be performed in parallel, which improves the performance. 3.2
Suppose a descriptor of a cluster is a set of terms, relevant to this cluster. Let an initial core of a cluster is a subset of highly similar papers, which is used for generation of a descriptor. A descriptor is built based on full-text of these papers. Let Core(d ) is a function that returns a tuple c, m
for given paper d , where c is an identifier for initial core of cluster and m is a numeric characteristic of paper d in initial core c. Characteristic m is used for decreasing of distance threshold H during building of initial core, that prevents merging of initial cores. If paper d is not presented in any initial core, the function Core(d ) returns empty tuple. Direct index of descriptors is a function that returns a set of different terms and their weights from the descriptor of cluster
{ w1,t1 , w2 ,t2 ,.., wn ,tn } where n is count of different terms in the descriptor. Inverted index of descriptors is a function that returns a set of descriptors in which given term w is occurred and weights of this term in these descriptors:
{ c1,t1 , c2 ,t2 ,.., cn ,tn }.
We use topic importance measure I for weighting terms of descriptors. This measure for term w from corpus D and cluster c is calculated as follows:
(w, c, D) idf (w, D) idf (w, Dc )
I (w,c, D) (w,c, D) X ( (w,c, D)) , where Dc is a set of papers from cluster c and X () is a Heaviside step function. Due to the topic importance measure, descriptor of cluster consists mostly of terms specific for papers of the cluster.
Let H a and are clustering thresholds, affecting on insularity of clusters.
The proposed full-text clustering method contains two parts. The first part could be performed independently on different nodes of computer network. This part conventionally can be called “initial cores of clusters detection”. The steps are the following. 1. Index the input corpus D . 2. While D is not empty, exclude a paper d from D . If D is empty, turn to step 6. If Core(d ) returns empty tuple, then turn to step 3, else turn to step 4. 3. Generate new cluster core index c. Add new tuple c,1 with m 1 to Core . Then turn to step 5. 4. Get identifier c for current initial core and m for
current paper. Reduce m m . 5. Search similar papers by the fast algorithm with threshold H H a current value of m. Then turn to step 2. 6. Build descriptors CDidx (c) and m . Add it to Core with
each initial core c .
The second part is performed in a selected “supervisor” node and consists of the following steps. 1. Search similar descriptors by the fast algorithm using CDidx and CIidx . Then merge these descriptors. In practice this step can be useful, if the first section executes on several computer nodes. 2. Classify all papers. For each retrieved descriptor use steps 2-6 from the fast similar search algorithm and resulting CDidx as a list of target terms. To prevent fuzzy clustering each paper is labelled by identifier of its cluster and by value of its similarity to descriptor of this cluster. Classifier uses these labels to include each paper in the most similar cluster.
The advantage of the proposed method is its distributed nature thereby full-text clustering of large amounts of papers can be implemented. That is necessary for detecting current research directions represented in wide research area. Another benefit consists in using indexes which have incremental structure, so adding a number of papers to the corpus D does not involve full re-clustering. 4 4.1
For experiment a dataset of papers of the research
area “Regenerative medicine” is created and verified by
experts with PhD degree. The dataset contains 112
wellcited publicly accessible papers from 2000 to 2014
distributed into 4 research directions. We apply widely
used linguistic analysis library Freeling [
Descriptor Brain stroke Chimeric antigen receptor cell therapy Induced pluripotent stem cells Wound healing burns Size Brain stroke Chimeric antigen receptor cell therapy Induced pluripotent stem cells Wound healing burns Terms stroke, brdu test, behavioral, endothelial cell, cell brdu, psa-ncam, cortical cell, reactive, regeneration neuronal, endogenous, neurogenesis, endostatin, cortical, expansion nonhematopoietic, ischemic neuron pbls, receptor chimeric, cd19-speci, immunity, antitumor, adoptive ifn, malignancy b-cell, aapcs, protein fusion, infusion cell gene signature, photoreceptor, epithelium retinal, retinal, suppressor, tumor, cell ips, technology ips, retinal cell, ips-derived, ipscs pigmented wound chronic, wound heal, keratinocyte, epidermal cell, fusenig keratin, epidermis region, We use the semi-automatic approach for tuning clustering methods that can avoid empirical parameters where the parameter setting precision priority over recall. In our experiment, 1 . We use a controlled Measure Precision
F-measure
Affinity propagation 0.85 0.28 0.42
Birch setup. Clustering parameters are calculated by maximization of the quality function with boundaries 0 Ha 1 and 0.7 0.99 (for the proposed method). For other methods we set boundaries according to constraints of these methods. We apply commonly used F-measure function for classification methods assessment. This function based on precision and recall.
Recall R is defined as the ratio of the number of correctly classified papers to size of class in the train set: where t is a number of correctly classified papers; c is a number of papers which the method carried to other classes.
Precision P is defined as the ratio of number of correctly classified papers to the total of classified papers: where t is a number of correctly classified papers; i is a number of incorrectly classified papers.
Then the f-measure calculates as follows:
R P
t t c
, t t i
, F ( 2
1)PR 2 P R , random search method [7] for optimization of quality function. 4.3
Thus, we show that proposed method can be tuned properly using a part of analyzed dataset. Results show that proposed method has sufficient quality of clustering (Table 2). Generated terms represent research directions from the dataset. Table 3 shows experiment results using cross-validation technique. Both qualitative and quantitative assessment shows that the proposed method produces relatively good results. In opposite, the affinity propagation method returns unsatisfactory results. Obviously this method is not suitable for the solution of research direction detection problem. Possibly it relates to incorrect exemplars selection due to the missing of the initial distribution for exemplars.
We suggest that our method overcomes other examined methods due to applying Manhattan distance for paper similarity estimation, to using of topic importance measure for determination of terms of clusters and due to implementation of linguistic analysis for terms extraction as well. That leads to more precise estimation of similarity measure between papers and, consequently, to better quality of research directions detection.
In this study we investigate various methods for detection of current research directions, which are based on clustering. We found that it is necessary to create a method that can be tuned successfully in a small labelled part of analyzed dataset.
In this paper we present the improved full-text clustering method for detection of research directions. Experiment results show that proposed method is more applicable for research directions detection, than the other widely used methods. The semi-automatic approach for tuning of clustering method demonstrates its performance also. Besides, it was shown that fulltext clustering methods work imprecisely for thematically heterogeneous research directions. This paper is considered as an initial study that will be extended in a further research. Moreover, the further development of proposed method consists in implementation of hybrid metrics for clustering. These metrics can be used not only texts similarity, but also co-citations and co-authorships of papers in dataset. It is necessary to continue work with experts for extension of our experimental dataset, because hybrid clustering methods cannot process small datasets properly.
A. //
Additive Machine [19] Zhang T., Ramakrishnan R., and Livny M.
BIRCH: an efficient data clustering method for very large databases. // ACM SIGMOD Record. ACM, 1996. – Vol. 25(2). – pp. 103-114.