Research on the similarity between nodes with hypernymy-hyponymy relations based on IC and taxonomical structure

Abstract EN

The similarity method has an important effect on some tasks of natural language processing, such as information retrieval, automatic translation and named entity recognition.

Hypernymy/hyponymy relations are widespread in semantic webs and knowledge graphs, so computing the similarity of hypernymy/hyponymy is a key issue in the text processing field.

All measures of both feature-based and IC-based methods have obvious deficiencies.

The feature-based method estimated the similarity by the depth of the node, and the IC-based method computed the similarity by the position of the deepest common parent.

The deficiency of the feature-based method and IC-based method is that they include one parameter, so the performance is slightly inaccurate and unstable.

To address this deficiency, our paper proposed a hybrid method that computes the similarity of hypernymy/hyponymy by a hybrid parameter (dhype(lch)) that implies two parameters: depth of the node and position of the deepest common parent.

Compared with several similarity methods, the proposed method achieved better performance in terms of accuracy rate, Pearson correlation coefficient and artificial fitting effect.