In this paper, we present a general-purpose interactive visual query language, GPVQL, to improve the e ciency of end-users' understanding and querying of knowledge graphs. Furthermore, GPVQL realizes the novel capability of exible bidirectional transformations between query patterns and graph results, therefore signi cantly assists end-users in formulating queries over large and unfamiliar knowledge graphs in an incremental way. We present the syntax and semantics of GPVQL, discuss our design rationale behind this interactive visual query language, and evaluate the e ectiveness of a visual query system based on GPVQL against a number of textual and visual query environments over a large knowledge graph, DBpedia. Our evaluation demonstrates the GPVQL's superiority in e ectiveness and accurateness.
Arti cial intelligence has become a powerful tool to meet practical requirements in various domains. Knowledge graphs have been identi ed as a critical component in diverse AI-based applications. Thus, designing query languages to support the e ective and e cient exploration and query answering over knowledge graphs has become a key research problem under insentive investigation. However, a number of key challenges still remain on the generalizability and ease of use of these query languages.
An illustrating example of the process from a question to the corresponding query results is shown in Fig. 1. As can be seen from the textual query (in SPARQL) at the top middle in Fig. 1, it may be di cult for an end-users to quickly learn and use such a query language, or to be familiar with a knowledge graph. Visual query languages can help make it easier for users to construct query Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Question Find philosophers who were influenced by people that influenced Bob Black.
Edit Draw
Textual language SELECT * WHERE { Bob Black influencedBy ?x. ?y influencedBy ?x.}
Unidirectional
query Bob Black ?m
influencedBy ?x influbeirnthcYeedaBry ?z influenced ?y influencedBy
Visual language
Bidirectional deathYear ?n transformation ?y ?z ?m ?x ?x Graphical result ?z ?n patterns, as shown in the bottom middle of Fig. 1. However, end-users may not understand the correspondence between a query pattern and its results (such as those nodes labeled by the same variable name at the bottom middle/right of Fig. 1) when they are exploring a knowledge graph. Such correspondences allow a user to easily modify and expand the current query to build more complex ones and are supported by the bidirectional transformation functionality provided by our GPVQL visual language. In this paper, we introduce the visual syntax and semantics of GPVQL, discuss our design rationale, and demonstrate its accurateness and e ectiveness in a user study. 2
In GPVQL, (
Based on our proposed GPVQL language, we have developed a visual interface that addresses these challenges to achieve our goal of enabling end-users to query and explore knowledge graphs. The interface 3 of GPVQL is composed of six main components that are displayed in four panes in Fig. 2. Fig. 2 shows three components: an interactive visual query editor (left pane), keyword and type search panel (right top pane), and query code display panel (right bottom pane). In the top left are buttons for additional functionality: adding new nodes, showing user manual, user study documents, using force-directed layout, etc. In GPVQL, the construction of a query pattern is mainly done in a drag-and-drop manner. The operations (i.e., Expand, Collapse, Filter, Lock, Optional, and Union) are in the right-click context menu. As shown in Fig. 2, GPVQL not only supports keyword queries, but also type-based queries as the starting point. The process starts with a blank visual query editor. A user can choose keyword or type as the starting point for queries. For example, as shown in Fig. 2, after the user adds a new node 1 , and input \Bob" as a keyword 2 , the related entities will be automatically displayed. Once \Bob Black" is chosen, the thumbnail 3 will be displayed automatically. Further, the user can nd people who have in
uenced both Bob Black and people in uenced by Bob Black (the corresponding query pattern is marked with red arrows). Users can select the Expand operation in the context menu to expand a result. The expanded results are pointed to with blue arrows. When users are interested in some query result, deep exploration to query additional information is supported. For example, users can nd the birth year and death year of Peter Kropotkin, or people who in uenced Marshall Sahlins (pointed to with green arrows).
We conducted a user study to assess how accurate and e ective our new visual query language GPVQL 3 is in comparison to the current semantic query language SPARQL, a visual query language QueryVowl [1], and RDF Explorer [3]. We used the same SPARQL endpoint 4 of the DBpedia knowledge graph to ensure the fairness of the study. We created ve tasks, shown in Table 2, based on an informal survey of interesting patterns that people formed when exploring prior graph query research works. We recruited 20 participants in total and divided them into four groups evenly. Each group uses a di erent tool, e.g., Group A + GPVQL, Group B + SPARQL, Group C + QueryVOWL, and Group D + RDF Explorer.
We ranked the di culty of each task based on the number of nodes, edges, and constraints needed. As can be seen in Table 2, the tasks are related to each other, and that later task can be built incrementally from previous tasks. This design is intended to simulate the actual work ow in a real-world setting, where an end-user typically writes related and incrementally more complex queries, but not unrelated random ones.
From Fig. 3, we can obtain the following conclusions: (
GPVQL SPARQL QueryVOWL RDFExplorer 40 20 0 Task1
Task2 Task3 Task4 Task5 (a) Mean completion time for each tool per task. (b) Mean accuracy for each tool per task. with RDF Explorer, GPVQL reduces the completion time by approx. 15%, and increases the accuracy by about 10%. In summary, our evaluation demonstrates the superiority of GPVQL in both aspects of completion time and accuracy. We believe that the main reasons of GPVQL's superiority are bidirectional transformation and the deep exploration it facilitates. For changes of the query tasks, traditional query languages requires users to construct new queries from scratch, which reduces the speed, usability, and user-friendliness. With the bidirectional transformation in GPVQL, users can use query results of the previous query pattern as the input of the next query pattern, which reduces the di culty and time used of creating queries. 4
In this paper, we propose a general-purpose visual query language. The main advantage of GPVQL is the exible bidirectional transformation between query patterns and graph results, which is a useful method for end-users to gain insights of large-scale knowledge graphs, and eliminates the boundary between query patterns and graph results. We experimentally validated the accurateness and e ectiveness of GPVQL and its associated interface, showing its superiority over other visual query languages.