Multidimensional relational knowledge embedding for coreference resolution

Currently, the co-reference resolution model using a knowledge base mainly faces two problems: first, the knowledge is complex and diverse, and it is difficult to add appropriate knowledge to complement the conceptual relationships between entities; second, it is difficult to integrate the obtained external knowledge into the model. In this paper, we propose a multidimensional relational knowledge model (MDR) for co-reference resolution, which extends in both high-dimensional and low-dimensional directions according to the antecedent words to be parsed, abstracts upwards to high-dimensional concepts to represent the essential relations of things, and diffuses downwards to find intra-sentence words to make the knowledge closer to the sentence meaning, providing the model with more generalised multidimensional relational knowledge and higher sentence relevance. At the same time, in order for the model to make full use of the knowledge, the attention mechanism is adjusted to use external knowledge to guide the intra-sentence relationship changes and adjust the degree of knowledge dominance according to the back-propagation of the neural network. The knowledge noise reduction module is designed based on a multiplexed hybrid approach, using a hybrid approach to dilute the proportion of knowledge in the total information and reduce noise generation. The multidimensional relational knowledge model is evaluated on the Definite Pronoun Resolution Dataset and Winograd Schema Challenge datasets, showing its cross-sectional comparison with existing models in experiments and ablation experiments, and the role of knowledge is analyzed for experimental cases to demonstrate that the multidimensional relational knowledge is helpful for model co-reference resolution ability.