Long Text Relationship Extraction Method for Complex Productse

Abstract

In the data management of complex products, relationship extraction of related texts can assist in constructing product data chains. This can realize the integration and fusion of data chains through nodes with relationships between different data chains. However, due to the complexity of complex products in terms of customer requirements, product technology, and manufacturing process, many related text data contain a large number of complex sentences, and a large amount of referential information is often lost in the relationship extraction of long text in these complex sentences, resulting in poor relationship extraction results. In this paper, we propose a long-text relationship extraction method for complex products, using a pre-trained language model to encode semantic information and obtain input text word vectors, then using a Gaussian graph generator (GGG) to construct potentially directed multi-views, learning graph features more deeply with the help of densely connected graph convolutional networks, and using dynamic time-regularized pooling operations to extract more relationship-dependent indicative words to assist the relationship. The extraction task is completed by combining the graph feature learning results with semantic information embedding representation for relationship extraction. Experiments are conducted on the DialogRE dataset, and the experimental results show that the F1 values reach 66.1% and 63.3% on the validation and test sets, respectively, and the F1 values still exceed 65% when the number of text words exceeds 400, which verifies the feasibility and effectiveness of the proposed method.