Weak supervision relation extraction method based on multi-source semantic representation fusion

Abstract

The invention provides a weak supervision relation extraction method based on multi-source semantic representation fusion. Firstly, distributed word vectors are adopted to initialize context semanticfeatures of text statements, a natural language processing tool is adopted to analyze mass discretized symbol features describing the text features, and a universal first-order logic rule between statement instances and features in a relation extraction task is designed; then, a logic rule is combined with the factor graph to establish a relationship between the text characteristics and the statement instances, modeling is performed from the perspective of human perception through joint statistical reasoning, and a low-dimensional relationship semantic vector for describing the text characteristics is learned; and the semantic information of the statement content encoded by the bidirectional gating loop unit is used as a context content semantic vector. And finally, text characteristic semantic vectors are finely adjusted in the neural network, vector representations of two different characteristic sources are fused to obtain text semantic characteristic representations with higher robustness, and weak supervision relationship extraction work is guided together with entity pair embedded representations.

Description

technical field [0001] The present invention belongs to the technical field of information extraction, and is suitable for relation extraction in the general field, and in particular relates to the extraction of the "entity-relationship-entity" triplet relation in a sentence under the condition that the training samples corresponding to weak supervision are inaccurate and unbalanced, specifically A Weakly Supervised Relation Extraction Method Based on Multi-source Semantic Representation Fusion. Background technique [0002] In real life, unstructured text information is like dark matter, buried in massive network data, and difficult to process due to lack of structure. In-depth study of entity relationship extraction technology is to use the computer's ability to efficiently process text to extract a unified format of relational fact representations from massive, unstructured network texts. By mining the semantic information of the target entity in the text sentence, it pr...

AI Technical Summary

Problems solved by technology

However, in order to save labor costs, weakly supervised signals are often insufficiently supervised, which makes the relationship extraction method face problems such as wrong labels and uneven sample distribution during the training process.

Most of the existing relationship extraction algorithms based on weakly supervised learning focus on the embedded information of the original corpus content to alleviate the impact of insufficient supervision resources. They lack the integration and full utilization of semantic information at different levels, and the expression of semantic features is relatively simple, which can easily lead to serious extraction results. Prefer relationship types with more training samples

Images