Knowledge map reasoning -- hybrid neural network and distributed representation reasoning

original text ：Schlichtkrull M, Kipf TN, Bloem P, Berg RVD, Titov I, Welling M. Modeling relational data with graph convolutional networks. arXiv Preprint arXiv: 170306103, 2017.https://link.springer.com/chapter/10.1007/978-3-319-93417-4_38

Introduce

Knowledge maps can be applied in many ways , Such as question and answer and information retrieval . Although many knowledge maps are large, they are still incomplete , such as Yago, DBPedia perhaps Wikidata. This article applies R-GCNs Complete two tasks ： Link to predict （ Recover the lost facts , Triple ） And entity classification （ Recover the missing entity attribute values ）. The missing information segment is predicted by the graph encoded by neighborhood structure .
The knowledge spectrum is represented by triples （ Entity , Relationship , Entity ） Such as ：(Mikhail Baryshnikov, educated at, Vaganova Academy), Every entity （entity） There are types （type）

chart ：

In order to accomplish the above tasks , This paper realizes ：
Entity classification model ： Each node in the figure uses softmax classifier , The classifier accepts RGCN The provided node represents , And make prediction labels .
Link prediction model ： Encoder ,RGCN Generate entity latent feature representation ; decoder , A tensor factorization model uses these representations to predict the edges of markers , Factorization method :distmult.

Main contributions

1. Is the first proof GCN The framework can be applied to people who model relational data , In particular, link prediction and entity classification tasks .
2. The technology of parameter sharing and strengthening sparse constraints is introduced , And use them to R-GCNs Apply to multiple graphs with a large number of relationships .
3. With DistMult For example , The author proves that the performance of the factorization model can be enriched by the encoder model that performs multiple information propagation steps in the relational graph

Relational convolution network

The symbolic representation of the graph ：

To be continued …