Literature search is a critical step in scientific research. Most of the current literature search tools present the search results as a list of documents. These tools fail to show the structure of the search results. To address this issue, we propose an interactive visual tool for searching scientific literature. This tool creates, labels and visualizes clusters of documents that may be of relevance to the user. In this way, it provides the user with an overview of the structure of the search results. This overview is intended to be understandable even to a user who has only a limited familiarity with the scientific domain of interest. We present the concept of our tool, show a case study of its use and describe the technical specifications of the tool. In particular, we provide a detailed specification of the algorithm that we use to visualize clusters of documents.
Literature search is an essential part of any research project. Many of the current
literature search tools (e.g. Google Scholar [
There is some literature studying the idea of showing the structure of search
results. An example is the recent work on a tool called PaperPoles [
While these tools are helpful, some of them (e.g. CiteSpace, VOSviewer) were
developed primarily for bibliometric analysis, not for literature search. Others (e.g.
CitNetExplorer, Citation Gecko) have the limitation of showing search results only at the
level of individual papers, not at aggregate levels. To overcome the limitations of
currently available tools, we propose a new tool for literature search. This tool uses an
interactive visual interface to show the structure of the search results. We make use of
ideas and techniques that we also used in the development of other tools (i.e.,
VOSviewer and CitNetExplorer), but we now focus specifically on literature search
rather than on bibliometric analysis. To some degree, the proposed tool resembles
Open Knowledge Maps. However, by relying on the scatter/gather approach [
This paper is divided into three parts: We first provide a description of the proposed tool (Section 2), we then present a case study demonstrating the use of the tool (Section 3) and finally we give a technical specification of the algorithms included in the tool (Section 4). 2
Our proposed tool is based on the scatter/gather approach [
Our tool scatters a set of papers into clusters. The clustering uses the citation links
between papers. Each cluster is given a label. The label of a cluster consists of the ten
noun phrases with the highest weighted frequency in the titles and abstracts of the
papers in the cluster. The weighting considers the frequency of occurrence of the
noun phrases in the focal cluster relative to other clusters. This clustering and labeling
method is based on Waltman and Van Eck [
Our tool also visualizes the clusters to complement the labels. It visualizes the clusters as bubbles in a packed bubble chart. The size of the bubbles reflects the number of papers in the clusters and the distance between the bubbles approximately reflects the number of citation links between the clusters.
Our tool supports multiple iterations of scattering and gathering. The user can load the initial set of papers, choose the clusters to gather, choose the number of clusters to scatter, retrieve the papers in the clusters, and so on. 3 3.1
Set up First, let us consider a user working with a traditional literature search engine for scientific literature, like Google Scholar. She has to come up with several search queries. She does not have a background in the academic field that she is looking into, so probably she will not come up with good queries. Also, she has no way to know if she is missing important papers or even entire subfields!
Second, let us assume instead that she uses a literature search engine that offers some very basic features for exploring the structure of the search results, like Web of Science. She can now see to which academic fields her search results belong. Despite of this, she still has basically the same problems as with Google Scholar.
Third, now let us assume that she uses our proposed tool for her literature search.
For this example, we will follow her through all the steps of the search process. We
will assume that she is interested in getting to know the scientific literature about the
review process of grant proposals. For the initial set of papers, we will use the set of
the cluster of scientometrics papers obtained using the algorithmic methodology
employed at CWTS [
The researcher retrieves the set of papers and chooses a value of 10 for the number of clusters in the first scattering. Then she sees the visualization (Fig. 2A) and the labels (Table 1) of the clusters. From the labels, she sees that her topic of interest is in cluster 6. She also checks the labels of the clusters close to cluster 6 (clusters 0, 3, 5, 8 and 9). Their labels indicate that they do not relate to her topic of interest, so she only gathers cluster 6.
She chooses to have 5 clusters for the second scattering and sees the visualization (Fig. 2 B) and the labels (Table 2) of the clusters. Now the labels are more ambiguous, so she will have to also read the titles of the papers inside clusters to understand what the clusters are about. She suspects that her topic of interest is in clusters 1 and 2. From the visualization and the labels, she also sees that her topic could be in cluster 4. She reads the titles of the top 5 most cited papers in these three clusters (Tables 3, 4 and 5). She finally decides that she should start reading paper 3 from cluster 1 and papers 2 and 4 from cluster 2.
In this example, we have illustrated how our tool could improve scientific literature search. The key advantage of the proposed tool is that the user is informed about the way in which the scientific literature is organized. For instance, the user is able to see how a field is divided into subfields or topics. As a result of this, the user is able to discard papers unrelated to the topic of interest without the need to skim the titles of large numbers of individual papers. Instead, the user examines the labels of clusters and then decides to discard entire clusters that appear to be of no relevance. Also, the user does not need to try to come up with a detailed keyword query that identifies exactly the right papers. It is sufficient to be able to identify a broad set of papers that could potentially be of relevance. Within this broad set of papers, the papers of interest can then be found by drilling down into the right clusters.
Papers 4344 3154 1652 1651 1231 1230 932 816 810 492 Papers 387 270 104 104 67 11 2013
JOURNAL OF INFORMETRICS
Cit. 23 12
Year 2013 2011 11
2012
10
2011
We cluster the papers by applying the Leiden algorithm to their citations links [
We visualize clusters using a packed bubble chart. We developed an algorithm to create these charts (see below). The input of our algorithm is an undirected network. In this network, nodes represent clusters of papers, the weight of a node indicates the number of papers in a cluster, and the weight of an edge between two nodes indicates the relatedness of two clusters in terms of citation links.
Our bubble chart algorithm determines the coordinates of the bubbles, where each
bubble is a node in a network. The objective of our bubble chart algorithm is to obtain
a visualization in which the bubbles do not overlap, the empty space is minimized,
and the positions of the nodes relative to each other reflect their relatedness as
accurately as possible. We base our algorithm on the VOS layout algorithm [
The area of a node is proportional to the weight of the node. Therefore, the radius
of a node is the square root of w, where w is the weight of the node. Nodes connected
by edges with a high weight should be close together. To achieve this, we minimize a
weighted sum of the squared Euclidean distances between all pairs of nodes, which is
similar to the VOS layout algorithm [
The best strategy to minimize Eq. 1 while satisfying Eq. 2 in a network of two nodes (nodes 1 and 2) is to place the nodes adjacent to each other. When we fix the coordinates of node 1, the coordinates where node 2 can be placed form a circle c(1,2) around node 1 (Fig. 3A). This circle has a radius equal to the sum of the radius of node 1 and the radius of node 2. Now, we also fix the coordinates of node 2 and add node 3 to the network layout. We can use the same strategy to get its coordinates. The adjacent coordinates for node 3 form the circles c(1,3) and c(2,3) (Fig. 3B). Therefore, the available coordinates to place node 3 are the intersection points of c(1,3) and c(2,3) (Fig. 3C).
When we add node 4 to the network layout, the available coordinates for this node are no longer all the intersection points of the circles c(i,j), because some coordinates would cause nodes to overlap (Fig. 3D). Of the available coordinates, we select the ones that result in the lowest stress. We can find these coordinates by calculating the weighted sum of the squared Euclidean distances between node 4 and each node that has already been assigned to coordinates. We proceed in the same way for all other nodes.
Our minimization algorithm obtains the coordinates of the nodes by adding them one-by-one to the network layout. However, we found that the value of the stress at the end of an algorithm run is highly dependent on the order in which the nodes had been added. To improve our minimization algorithm, we added a step in which we create several lists of the nodes in a different order. For each list, we run the minimization procedure and in the end we return the network layout with the lowest stress.
We order the nodes in the lists as follows. For each node in the network, we create a list with that node as the first node. The next node in the list is the one that is most strongly related to the nodes already in the list. We repeat this process until all nodes have been added to the list.
Our minimization algorithm is a heuristic approach to the minimization of Eq. 1 and does not guarantee that the global minimum of Eq. 1 will be found. The pseudocode of the algorithm is provided in the appendix. 5
We have proposed a tool for scientific literature search based on the scatter/gather approach. The tool visualizes the structure of the search results using a packed bubble chart. We have presented a case study demonstrating the use of the tool and we have provided a technical specification of the algorithms included in the tool, in particular the algorithm for creating packed bubble charts.
Compared to traditional literature search tools that present the search results as a list of documents (e.g. Google Scholar), we expect the advantage of our tool to be in the emphasis it puts on showing the structure of the search results. We expect this to be important especially when users are searching not for one specific paper but for a larger set of papers offering a broad understanding of a certain scientific domain. In future work, we plan to test the performance of the tool for different information retrieval tasks.
----INPUT: list INLIST containing nodes (x0,...,xn).
Each node possesses: A node identity id(x) A radius r(x) A list of edges E(x) containing (e0,....,en), with each edge e possessing a weight w(e) and an node identity id(e) of the node it connects to
A coordinate c(x) that contains nothing OUTPUT: list OUTLIST containing nodes (x0,...,xn) possessing non-empty coordinates c(x) ----Create list MASTERLIST containing nothing For each node xi in list INLIST (x0,...,xn):
Complete subroutine S_ORDER(xi,(x0,...,xn)) Create list Zi containing nothing Set coordinate c(xi0) of node xi0 as (0,0) Append node xi0 to list Zi Set coordinate c(xi1) of node xi1 as ((r(xi0)+r(xi1),0) Append node xi1 to list Zi Complete subroutine S_COOR(Zi,(xi2,...,xin))
Append list Zi to list MASTERLIST Return list OUTLIST in MASTERLIST (Z0,...,Zn), where OUTLIST is the list with lowest graph stress V as defined in the equation 1 V(OUTLIST) -----Subroutine S_ORDER creates an order of nodes S_ORDER(xi,(x0,...,xn)): Create list Xi containing nothing Append node xi to list Xi as node xi0 Create list Yi containing nodes (x0,...,xn) Remove node xi from list Yi While list Yi containing something:
For each node xj in Yi:
Declare twj is the total weight from xj to all the nodes in Xi Declare xtw is the node with greatest twj Append node xtw to list Xi as node xij
Remove node xtw from list Yi ----Subroutine S_COOR gets the coordinates of the nodes for nodes x>1 S_COOR(Zi,(xi2,...,xin)): For each node xij in (xi2,...,xin):
Create empty list TEMPij For each order-independent pair of nodes (xijm, xijn) in list Zi, where m > n:
Complete subroutine S_TEST(xij,xijm,xijn,Zi,TEMPij) Append node tempij to list Zi, where tempij is the temporary node with lowest node stress v in list TEMPij ----Subroutine S_TEST tests if the node xij can be adjacent to nodes (xijm, xijn), get the coordinates of center of these adjacent positions, test if the node xij on that coordinates overlaps with other nodes and get the stress of the node xij on that coordinates.
S_TEST(xij,xijm,xijn,Zi,TEMPij): Declare temporary node tempijm with coordinate c(xijm) and radius (r(xij)+r(xijm)) Declare temporary node tempijn with coordinate c(xijn) and radius (r(xij)+r(xijn)) If tempijm and tempijn DO overlap:
Declare coordinates coorijmn1 and coorijmn2 are the coordinates of the intersection between the borders of tempijm and tempijn For coorijmnk in list (coorijmn1, coorijmn2):
Declare temporary node tempijmnk is a node with the parameters of node xij, except that its coordinate c(tempijmnk) is coorijmnk If node tempijmnk DOES NOT overlaps with any node in Zi:
Declare node stress vijmnk is the total stress of the node tempijmnk with every node in the list Zi
Append tempijmnk to list TEMPij ----