Explaining and Querying Knowledge Graphs by Relatedness - Leitura de Artigo (DEMO)

Valeria Fionda and Giuseppe Pirrò. 2017. Explaining and querying knowledge graphs by relatedness. Proc. VLDB Endow. 10, 12 (August 2017), 1913–1916. DOI:https://doi.org/10.14778/3137765.3137807

ABSTRACT

We demonstrate RECAP, a tool that explains relatedness between entities in Knowledge Graphs (KGs) and implements a query by relatedness paradigm that allows to retrieve entities related to those in input. 

 

One of the peculiarities of RECAP is that it does not require any data preprocessing and can combine knowledge from multiple KGs. The underlying algorithmic techniques are reduced to the execution of SPARQL queries plus some local refinement. This makes the tool readily available on a large variety of KGs accessible via SPARQL endpoints. 

 

To show the general applicability of the tool, we will cover a set of use cases drawn from a variety of knowledge domains (e.g., biology, movies, co-authorship networks) and report on the concrete usage of RECAP in the SENSE4US FP7 project. 

 

We will underline the technical aspects of the system and give details on its implementation. The target audience of the demo includes both researchers and practitioners and aims at reporting on the benefits of RECAP in practical knowledge discovery applications.

 

1. INTRODUCTION

 

LOD datasets can be queried via SPARQL, an expressive query language, while Google’s KG and the like are typically accessible only via keywords. In the first case, users require knowledge of the SPARQL query language while in the second case, the lack of a query language makes it difficult to precisely express what one is looking for. 

 

One peculiarity of KGs underlying Google and other search engines is that they also suggest entities related to the entity in input.

 

* Não é só busca, contém recomendação * 

 

In this demo we demonstrate a novel approach that combine relatedness explanations and querying capabilities called RECAP. The user inputs a pair of entities and the system produces different types of explanations that capture the essence of their relatedness according to different aspects (e.g., path diversity, informativeness).

 

* Explicabilidade da relação *

 

One peculiarity of RECAP is the possibility to work with any existing KG without data preprocessing. Because of this design choice, the algorithmic techniques are reduced to the execution of a set of queries plus some local refinement. We will focus on an implementation of RECAP using the SPARQL query language for RDF data, which makes the tool readily available on a variety of KGs accessible via SPARQL endpoints.

 

* Uso imediato pq não requer pré processamento e trata de relação. Na ferramenta de similaridade do kgtk tem o cálculo de embeddings e é similaridade *

 

Here, relatedness explanations provided by RECAP are useful to investigate and show to the user topic connectivity, thus enabling to find out previously unknown relevant information, understand how it is of relevance, and navigate it.

 

2. SYSTEM DESCRIPTION

 

In this demo we focus on KGs encoded in the RDF data format and accessed via a SPARQL endpoint. Nevertheless, our approach is generic enough to capture data in other graph data formats.

     

[Path Builder] A user information request is expressed as a pair of entities plus a maximal distance. ... The first step of the evaluation of an information need consists in retrieving paths interlinking the specified entities by issuing a set of (automatically constructed) SPARQL queries to one or more endpoints. 

 

* Acredito que seja um conjunto de queries BGP simulando uma BFS saindo da origem, percorrendo a distância especificada, até encontrar o nó de destino *  

* RDF é grafo direcionado, deve considerar somente essa direção para achar o caminho *

 

[Explanation Builder] Paths interlinking entities form the basis to generate different kinds of query explanations. One immediate type of explanation is the merge of all retrieved paths.  

 

* Agrupo todos os caminhos encontrados e forma um subgrafo *

 

However, this kind of explanation becomes quickly unreadable as the number of intermediate entities involved can be very large. To cope with this issue, RECAP allows to flexibly chose the set of paths that will form an explanation.

 

* Critérios para ordernar e podar os caminhos que vão ser incluído na explicação * 

• Informativeness: it is estimated by investigating RDF predicates in a path via the notion of Predicate Frequency Inverse Triple Frequency.
• Pattern informativeness: a path pattern generalizes a path by replacing nodes with variables. Pattern informativeness is computed by counting the number of paths sharing a certain path pattern.
• Diversity: it takes into account the variety of predicates in a set of paths; diversity guarantees to rank high paths that contain rare predicates.
  

* Predicate Frequency Inverse Triple Frequency: número de vezes que o predicado aparece no caminho / número de predicados distintos dos caminhos - CONFIRMAR * 

 

The user can use one or more path ranking strategies to build an explanation....

 

[KG Query Engine] Explanations capture the essence of relatedness between a pair of entities. RECAP leverages explanations to learn SPARQL queries that can be used to find other pairs of entities similar to those in input. The tool considers three different approaches for learning SPARQL queries: 

(i) edge-isomorphic, where nodes in an explanation are replaced by variables while keeping the same join structure; 

(ii) onto-relaxed queries obtained from edge-isomorphic queries by substituting a subset of predicates with more general ones (taken from the ontology); 

(iii) edge-relaxed queries obtained from edge-isomorphic queries by substituting a subset of predicates with variables. Predicates to be relaxed can be either indicated by the user or automatically selected on the basis of predicates’ informativeness. 

 

* Três opções para construir as queries. Para a segunda opção é necessario ter uma ontologia e entendi que precisa usar reasoning para usar um predicado pai e ainda assim recuperar caminhos que passam por predicados filho * 

 

The last component of the query engine involves the ranking of results. RECAP at the moment implements two different strategies; one based on Pagerank and the other based on the Katz index introduced to estimate the relatedness of actors in a social network.

 

* Ordenação dos resultados gerados com as consultas SPARQL extraídas da explicação entre os dois nós de entrada *  

3. RUNNING EXAMPLE

 

We have developed a main running example that will serve as a basis to explain the ideas underlying our framework.

 
The example is on the domain of movies and involves the pair of entities <A> and <B>. The users inputs this pair to RECAP ... he interface also enables to chose the maximum path distance to be considered ... An example of explanation  ... combines information from Freebase and DBpedia. The explanation includes the top-20 most informative paths (out of 240) at max. distance 2 for the pair ... 

 

* Encontra caminhos e cada caminho possui um cálculo de relevância para ordenação dos resultados. Os top-K, onde K é definido na aplicação, são apresentados *  

 

.... allows to filter an explanation according to certain types of predicates. ...

 

* Poderia filtrar por contexto se as consultas recuperassem o contexto dos caminhos *  

* O filtro é aplicado na interface depois que os caminhos são recuperados, ordenados e apresentados ao usuário *

The system suggests different types of SPARQL queries ... query can be (manually) refined and then executed to find other pairs of entities. The top-k pairs of entities found, and ranked by their popularity, are show  ...

 

* As consultas BGP/CGP geradas para representar os caminhos encontrados podem ser aplicadas a outros pares de nós <origem> e <destino>, usando os nós intermediários como variáveis *

 

4. DEMONSTRATION SCENARIO

 

We will pick specific examples having small and large explanations and underline the following aspects: (i) filtering explanations is a necessary step when the size of the explanation starts to grow; (ii) each type of explanation has pros and cons. 

 

The idea is to underline the pros and cons of our approach with particular emphasis on the following aspects: (i) balance between path length and running time; (ii) motivate the design choice to work on top of KGs and access them only via endpoints.  

* Não faz referência ao deo online *

* A abordagem de encontrar caminhos entre dois nós (usados como seed / semente) e depois extrair padrões de queries BGP desses caminhos para uma forma de aprendizado mas é feito por regras *