Knowledge Graph Embedding: A Survey of Approaches and Applications - Leitura de Artigo 2

Q. Wang, Z. Mao, B. Wang and L. Guo, "Knowledge Graph Embedding: A Survey of Approaches and Applications," in IEEE Transactions on Knowledge and Data Engineering, vol. 29, no. 12, pp. 2724-2743, 1 Dec. 2017, doi: 10.1109/TKDE.2017.2754499.

1. Techniques that conduct embedding using only facts observed in the KG
2. Techniques that further incorporate additional information besides facts. 
3. How embeddings can be applied to and benefit a wide variety of tasks (in-KG applications and out-of-KG applications)

(2)

There is a wide variety of additional information that can be incorporated to further improve the task, e.g., entity types, relation paths, textual descriptions, as well as logical rules. 

The first kind of additional information we consider is entity types, i.e., semantic categories to which entities belong. For example, AlfredHitchcock has the type of Person, and Psycho the type of CreativeWork. This kind of information is available in most KGs, usually encoded by a specific relation and stored also in the form of triples, e.g., (Psycho , IsA, CreativeWork).  

Semantically smooth embedding (SSE), which requires entities of the same type to stay close to each other in the embedding space,

Entity types can also be used as constraints of head and tail positions for different relations, e.g., head entities of relation DirectorOf should be those with the type of Person, and tail entities those with the type of CreativeWork. Such constraints in the training process, particularly during the generation of negative training examples. Negative examples that violate entity type constraints are excluded from training.

The second kind of additional information we consider is relation paths, i.e., multi-hop relationships between entities.A relation path is typically defined as a sequence of relations r1→⋯→rℓ through which two entities can be connected on the graph. For example, BornIn→LocatedIn is a path linking AlfredHitchcock to England, via an intermediate node Leytonstone . Relation paths contain rich semantic cues and are extremely useful for KG completion, e.g., BornIn→LocatedIn is indicative of the relation Nationality between AlfredHitchcock and England.

A key challenge then is how to represent such paths in the same vector space along with entities and relations. A straightforward solution is to represent a path as a composition of the representations of its constituent relations, since the semantic meaning of the path depends on all these relations.

Extension of TransE to model relation paths, referred to as path-based TransE (PTransE).  

Most KGs there are concise descriptions for entities which contain rich semantic information about them. Besides entity descriptions stored in KGs, it can be extended to incorporate more general textual information (de fontes externas)

Jointly embedding utilizes information from both structured KGs and unstructured text. KG embedding and word embedding can thus be enhanced by each other. Moreover, by aligning these two types of information, jointly embedding enables the prediction of out-of-KG entities, i.e., phrases appearing in web text but not included in the KG yet.

Finally we consider the incorporation of logical rules, particularly those represented in terms of first-order Horn clauses, e.g., ∀x,y:HasWife(x,y)⇒HasSpouse(x,y) stating that any two entities linked by the relation HasWife should also be linked by the relation HasSpouse. Such logical rules contain rich background information and have been widely studied in knowledge acquisition and inference ... Recently, there has been growing interest in combining logical rules with KG embedding models.

Most KG embedding techniques do not explicitly distinguish between relations and attributes. Take the tensor factorization model RESCAL as an example. In this model, each KG relation is encoded as a slice of the tensor, no matter it indicates a true relation or just an attribute.