Abstract
Neural tensor networks (NTNs) were adapted from standard neural network
models for use in the field of knowledge graph embeddings, and serve to infer
relations between entities. This dissertation introduces two novel adaptations
of the NTN model, both of which are centered around the concept of a
relation. Each of the two major pieces of research presented is accompanied
by a respective international, peer-reviewed publication, and both publications
were later included in Springer archive journals. This dissertation is
also accompanied by an extensive literature review, which covers both NTNs
themselves as well as surrounding, relevant knowledge from the field of artificial
intelligence.
In our first contribution, we adapt the manner in which NTNs may be
trained. As with other knowledge graph embedding models, NTNs make use
of corrupted triplets during their training. In this research, we demonstrate
that there are both technical and semantic distinctions between corrupting a
tail entity - as is done by the original NTN implementation - and corrupting
a relation. We assess the two means of producing corrupted triplets in an
experimental context, finding notable distinctions between the resultant performance
of the NTN model under the different corruption mechanisms. We
also demonstrate experimentally the importance of balancing the respective
impacts of the valid and corrupted triplets during the training of the NTN.
In our second contribution, we augment the NTN model by providing it
with the means to leverage inter-relation patterns. We illustrate that the
intelligence of the NTN for any given relation is entirely isolated from the
other relations being considered. In response, we introduce cross-relational
reasoning: a novel means of coordinating relation-specific NTN outputs such
that any patterns existing between the relations in problem domain (such as
correlation) might be utilized during inference. Our novel approach is evaluated
against the original in an experimental context, and we demonstrate
that cross-relational reasoning can significantly augment the inferential capabilities
of the NTN. Furthermore, our second contribution also explores
the potential of adapting the NTN’s original activation function to improve
its efficacy.