On multi-relational link prediction with bilinear models
Study the expressiveness and connectivity between different bilinear models. Many models such as RESCAL, TransE, DISTMULT, HolE, and ComplEX can be represented using bilinear models and additional constraints on the embeddings. These constraints give us an universal model.
Research which relation types can be represented by which models
In-depth analysis on model performance with different alternative datasets and training methods!
Knowledge graph embedding by flexible translation
Proposed flexible translation between entity and relation vectors to model the different complex relations between entities (symmetric, transitive, one-to-many, many-to-many etc). This involves coming up with a new scoring function that favors flexible translation without increasing model complexity. There are many methods explored for knowledge graph embeddings but they usually have limitations. General linear models can’t capture the correlations between entities and relations. Bilinear models can only model linear interactions and neural network is hard to scale. Lastly, the translation-based model works well but the underlying principle is too strict that it fails to model complex objects and relations. The Flexible Translation aims to fix this rigidness, whereby instead of strictly enforcing h + r = t, we only constrain the direction of h + r (or t – r) is the same as alpha*t (or alpha*h), where alpha is the flexible magnitude of the vectors. The new scoring function scores compatibility of a fact using the inner product between (the sum of head entity vector and relation vector) and tail vector. This is instead of the Euclidean distance!
Analogical inference for multi-relational embeddings
Proposed a novel framework that uses analogical properties of entities and relations to improve the latent representations. The framework also unified different multi-relational embeddings. The paper argue that analogical inference is desirable for knowledge completion since if system A is analogous to system B, then the unobserved triples in B could be inferred by mirroring the counterparts in A.
Explore analogical structures in other machine learning problems such as machine translation and image captioning