TL;DR
CounQER is a system designed to automatically discover and align "counting predicates" (e.g., staffSize) with their corresponding "enumerating predicates" (e.g., workInstitution⁻¹) within large knowledge bases. By linking these two different expressions of set data, CounQER improves question-answering robustness and helps identify data gaps and contradictions for knowledge base curation.
What Problem Does This Paper Address?
Large-scale Knowledge Bases (KBs) like Wikidata, DBpedia, and Freebase are foundational for semantic web applications, yet they frequently suffer from incompleteness and inconsistency. Traditional schema constraints typically focus on basic properties like domain, range, or subproperty relationships.
However, KBs often store information about sets of entities in two distinct ways:
- Counting Predicates: Expressing the size of a set as an integer (e.g.,
numberOfChildren). - Enumerating Predicates: Listing the actual member entities of that set (e.g.,
child).
Standard KBs lack explicit alignments between these two variants. Without these links, it is difficult to cross-check data quality (e.g., ensuring the listed number of children matches the actual number of linked child entities) or provide alternative answers when one form of the data is missing.
Methods
CounQER addresses this gap using a two-step offline pipeline followed by an online query interface:
- Predicate Classification: The system uses supervised classification to automatically categorize predicates into either counting predicates (which take integer values) or enumerating predicates (which take entity values).
- Predicate Alignment: It pairs counting and enumerating predicates across the two variants by applying ranking methods based on statistical co-occurrence and lexical similarity metrics.
The online demonstrator allows users to run Subject-Predicate-Object (SPO) queries. When a user queries a specific predicate, the system fires the main query to the KB's SPARQL endpoint and simultaneously surfaces the top-five highest-scoring aligned predicate facts along with KB-specific predicate statistics.
Datasets
The authors evaluated and deployed CounQER across multiple variations of popular knowledge bases:
- Wikidata-truthy
- DBpedia-mapped (using structured, mapped extractions)
- DBpedia-raw (using raw, unmapped extractions)
- Freebase (accessible via the system's alignment browser)
Key Insights and Ideas
- Leveraging Schema Redundancy: Instead of treating duplicate representations of set data as a flaw, CounQER turns redundancy into an asset for data validation and query enrichment.
- Automated Curation Signals: Comparing the integer value of a counting predicate against the count of unique entities in an aligned enumerating predicate provides an immediate, automated signal for KB curation. If the numbers don't match, the data is either incomplete or inconsistent.
- Graceful Query Fallbacks: If a counting predicate is unpopulated for a specific entity, an aligned enumerating predicate can serve as a fallback, allowing QA systems to compute the count on the fly from the listed entities.
Key Results
The paper demonstrates the utility of CounQER through three practical scenarios:
- QA Enhancement: Querying the number of employees at Microsoft in DBpedia-mapped returns the explicit count ($114,000$) but also surfaces relevant entity lists from inverse predicates like
employer_invandoccupation_inv.
- Inconsistency Detection: A query for Charlie Chaplin's children on Wikidata reveals a data conflict: the
number of childrenpredicate states $6$, but the KB contains $9$ distinct statements for thechildpredicate.
- Zero-Result Recovery: When querying gold medals won by Roger Federer in DBpedia-raw, the primary predicate returns no results. However, CounQER successfully surfaces closely aligned counting predicates like
doublestitles($8$) andsinglestitles($88$) to provide partial answers.
Limitations and Future Work
While CounQER successfully identifies and visualizes these alignments, the authors note several areas for future development:
- Automated Resolution: The current system highlights inconsistencies but does not automatically resolve or flag them within the underlying KB.
- Multi-Hop Alignments: The alignment model is currently restricted to direct, single-hop predicate pairs and does not capture complex relationships that span multiple edges.
- OpenIE Expansion: Extending this alignment methodology to Open Information Extraction (OpenIE) domains remains an open challenge.
Summary updated