VisKoP: Visual Knowledge oriented Programming for Interactive Knowledge Base Question Answering

• Users can drag to add operators, connect them to express dependencies, and fill slots to specify operator arguments.Source: §1 · p. 2
• Intermediate operator results support debugging, and KB-schema autocomplete helps users unfamiliar with the schema.Source: §1 · p. 2
• The engine executes KoPL on a million-entity-level KB in less than 200 milliseconds.Source: §1 · p. 2
• Users found correct answers in an average 110.6 seconds; the engine was 16× faster than the original KoPL engine and 5× faster than Virtuoso.Source: §1 · p. 2
• VisKoP couples human-in-the-loop query editing and debugging with a KoPL execution engine that scales to a million-entity-level KB.Source: §1 · p. 2
• The authors report that allowing questioners to monitor query errors validates the approach and may improve system robustness.Source: §1 · p. 2

2.1 Knowledge Base

• KB represents entities, concepts, relations, and attributes, with relational, literal, and qualifier knowledge encoded as structured triplets.Source: §2.1 · p. 2

2.2 Knowledge Base Question Answering

• KBQA takes a natural-language question and returns either KB knowledge elements or results computed by logical or algebraic operations.Source: §2.2 · p. 3

2.3 KoPL

• KoPL comprises 26 connected operators with argument and dependency inputs; execution yields answers, while preserved intermediate results can be inspected and debugged.Source: §2.3 · p. 3

3 The VisKoP Platform

• The platform follows four design principles: graphical visualization, interactive editing, efficient execution, and transparent execution, split between the front end and back end.The interface handles visualization and editing; the execution engine handles efficiency and preserves operator outputs for inspection.Source: §3 · p. 3

3.1 Interactive Programming Interface

• A BART-based neural program induction module fine-tuned on KQA Pro converts natural-language questions into KoPL programs for graphical display.The generated KoPL program is output in depth-first-search order.Source: §3.1 · p. 3
• Users edit the tree-structured KoPL program by dragging in operators, linking dependencies, and slot-filling arguments; schema autocomplete, link checks, and operator outputs support correction and debugging.The interface presents intermediate execution results so users can inspect and debug their programs.Source: §3.1 · pp. 3–4, §3.2 · p. 4

3.2 Highly Efficient KoPL Engine

• The KoPL engine uses inverted indices, merges consecutive FindAll and filtering operators, and selects key-value storage structures by measured runtime on KQA Pro questions.The storage search considers hash maps, red-black trees, tries, and ternary search trees.Source: §3.2 · p. 4

4.1 Interactive Programming Interface

• VisKoP parses the border-count question into a KoPL program, but the example cannot be correctly parsed automatically and requires user correction.Users run the generated program to see its predicted answer.Source: §4.1 · p. 4
• Users replace the incorrect Or with And and correct the Relate argument; schema autocomplete and the And operator’s intermediate results help check the corrected query.The intermediate result shows which countries satisfy the And operator.Source: §4.1 · pp. 4–5, Figure 3 · p. 5, Figure 4 · p. 5

4.2 KoPL Engine

• The back-end engine loads the KB from JSON, builds in-memory data structures and indices, parses the KoPL program into its internal representation, and returns execution results.The engine is provided as a Python extension.Source: §4.2 · p. 5

5.1 Efficiency

• The efficiency study compares wall time over 94,376 KQA Pro queries on one Intel Xeon 5218R CPU with 1.0TB RAM.The platform is deployed on a million-entity-level KB extracted from Wikidata.Source: §5.1 · pp. 5–6
• 16× faster than the original KoPL engine and 5× faster than Virtuoso on equivalent SPARQL queries, VisKoP also has a smaller runtime-distribution peak and maximum than Virtuoso.Table 2 reports average running time over all queries; Figure 5 shows the running-time distribution.Source: Table 2 · p. 6, §5.1 · p. 6, Figure 5 · p. 6

5.2 User Study and Case Study

• 65.8% accuracy followed user correction of 20 questions selected as unanswerable without correction, with six users averaging 110.7 seconds per question and a 68.0-second median.The evaluation used manually annotated questions.Source: §5.2 · p. 6
• Questions requiring obscure or similarly named entities remain difficult; testers struggled to identify Sony Interactive Entertainment rather than Sony.The authors also note that some needed knowledge elements are not included in the KB.Source: §5.2 · p. 6

6 Related Works

• VisKoP follows semantic parsing and neural program induction, translating questions into multi-step programs rather than ranking entities from a retrieved sub-KB.The passage contrasts semantic parsing with information retrieval as two KBQA categories.Source: §6 · p. 6
• VisKoP integrates humans into program induction to provide external guidance and potentially improve robustness.The cited challenge is the expense of question-program parallel data and the large program search space.Source: §6 · p. 6
• VisKoP is the first among the compared systems to enable visual interaction and intermediate-result checking for users.The comparison names ReTraCk, SEMPRE, TRANX, and DTQA.Source: §6 · p. 6

7 Conclusion and Future Work

• VisKoP lets users monitor, edit, and debug KB queries.Its accompanying engine scales KoPL execution to a million-entity-level KB.Source: §7 · p. 7
• Future work could allow users to customize the KB.Source: §7 · p. 7
• The authors also identify guidance for recognizing true knowledge elements in a large-scale KB as important.Source: §7 · p. 7

Limitations

• VisKoP depends on the background KB being correct and broadly covering relevant knowledge.Finding the correct answer is extremely difficult when relevant knowledge elements are missing or incorrect.Source: Limitations · p. 7

Ethics Statement

• VisKoP is designed for users to edit KB queries with graphical elements.Source: Ethics Statement · p. 7
• The authors report 339,531 female and 1,458,903 male human entities in Wikidata and warn that missing female entities can lead to gender-biased answers.They attribute the imbalance to Wikidata data and identify KB debugging as important future work.Source: Ethics Statement · p. 7
• VisKoP is built on a high-quality subset of Wikidata.Source: Ethics Statement · p. 7
• The user-study participants were engineering graduate-student volunteers who received written and oral guidance; usernames were anonymized.Source: Ethics Statement · p. 7

A KoPL Definition

• KoPL operators retrieve, filter, verify, select, and combine knowledge elements according to names, conditions, values, or relationships.Query Operators find knowledge elements by name; Filter Operators keep elements satisfying conditions; Verification Operators test relationships to values; Selection Operators choose elements under a condition; and Set Operators perform inter-set operations.Source: A KoPL Definition · p. 10
• The operator descriptions are brief because the functions match those described in Cao et al. (2022a).The paper directs readers to that work or its website for each function's specific meaning.Source: A KoPL Definition · p. 10

B KoPL Program Format

• In Python, each knowledge operator is represented by a dictionary with function, inputs, and dependencies keys.The keys identify the operator name, its argument inputs, and its dependency inputs.Source: B KoPL Program Format · p. 10

C Questions for User Study

• The questions are accompanied by their corresponding answers in Table 3.Source: C Questions for User Study · p. 10, Table 3 · p. 11
• Table 3 presents the 20 questions used in the user study.Source: C Questions for User Study · p. 10, Table 3 · p. 11
• The table is identified as the user-study question set.Source: Table 3 · p. 11