| Summary: | Knowledge graph embedding aims to construct a low-dimensional and continuous space, which is able to describe the semantics of high-dimensional and sparse knowledge graphs. Among existing solutions, translation models have drawn much attention lately, which use a relation vector to translate the head entity vector, the result of which is close to the tail entity vector. Compared with classical embedding methods, translation models achieve the state-of-the-art performance; nonetheless, the rationale and mechanism behind them still aspire after understanding and investigation. In this connection, we quest into the essence of translation models, and present a generic model, namely, GTrans, to entail all the existing translation models. In GTrans, each entity is interpreted by a combination of two states-eigenstate and mimesis. Eigenstate represents the features that an entity intrinsically owns, and mimesis expresses the features that are affected by associated relations. The weighting of the two states can be tuned, and hence, dynamic and static weighting strategies are put forward to best describe entities in the problem domain. Besides, GTrans incorporates a dynamic relation space for each relation, which not only enables the flexibility of our model but also reduces the noise from other relation spaces. In experiments, we evaluate our proposed model with two benchmark tasks-triplets classification and link prediction. Experiment results witness significant and consistent performance gain that is offered by GTrans over existing alternatives.
|
|---|
