A model training method and text representation method for academic heterogeneous network embedding

Abstract

The invention relates to a model training method for embedding an academic heterogeneous network. Firstly, papers are used to generate an academic heterogeneous network. The academic heterogeneous network includes paper nodes, multiple paper feature nodes, edges and text content; multiple paper nodes are selected as queries Nodes, walk the academic heterogeneous network with meta-paths composed of different paper features, and generate close relationship node sets for each query node under various paper feature query conditions; for each query node, close relationship node set and academic The heterogeneous network is sampled to obtain multiple triplet data representing the relationship between the query node and other nodes; the language representation model is trained according to multiple triplet data, so that the model embeds the structural relationship information of the academic heterogeneous network into the text representation vector. The model trained by the invention can simultaneously embed the text semantics and the structural semantics of the academic heterogeneous network into the text representation vector in the academic field to improve the representation effect.

Description

technical field [0001] The invention belongs to the technical field of text representation, and in particular relates to a model training method embedded in an academic heterogeneous network and a text representation method. Background technique [0002] Text representation in the academic field is an important basis for accurate and efficient scientific literature search, academic expert search, academic community discovery and other services. [0003] Academic papers form a variety of rich associations through intermediate entities such as authors, topics, fields, conferences, and journals, such as cooperative relationships, citation relationships, and so on. These academic texts, intermediate entities and their relationships constitute an academic heterogeneous network of papers, such as the internationally renowned DBLP academic network. [0004] The existing research work on academic text representation can be divided into two categories: the first is to use random wal...

AI Technical Summary

Problems solved by technology

However, this method does not consider the content semantic features of the text when representing the text. It is only a representation of the academic network structure, that is, the relationship information between academic papers, which affects the accuracy of academic text representation.

The second is to represent academic papers based on the text content representation method, such as using the model to directly vectorize the text content of the paper. This method does not consider the relationship between a large number of useful academic papers in the academic heterogeneous network , although some academic papers use different vocabulary, they are actually highly related to each other

In this case, the representation method based on text content loses the relationship information among academic papers

Images