A zero-shot relation extraction method based on fine-grained matching

By employing fine-grained matching methods, manually designing and expanding relation descriptions, and combining the BERT model with a self-attention mechanism, the inference latency and accuracy issues in zero-sample relation extraction were resolved, achieving high-precision and low-latency relation extraction results.

CN117910475BActive Publication Date: 2026-06-19FUDAN UNIVERSITY +1
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
FUDAN UNIVERSITY
Filing Date
2022-10-10
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing deep learning methods suffer from high inference latency, low accuracy, and shortcut learning issues in zero-shot relation extraction, especially when dealing with emerging relation categories that lack labeled data.

Method used

We employ a fine-grained matching approach, which involves manually designing relationships to describe and expand the entity categories. We then combine the BERT model and self-attention mechanism for feature extraction and utilize projective distillation and a dual-tower structure for explicit matching, thereby improving matching accuracy and reducing inference latency.

Benefits of technology

It significantly improves the matching accuracy of zero-sample relation extraction, maintains the advantage of low inference latency, and achieves high-precision inference performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117910475B_ABST
    Figure CN117910475B_ABST
Patent Text Reader

Abstract

This invention provides a zero-shot relation extraction method based on fine-grained matching. First, relation descriptions are manually designed and their corresponding head and tail entity categories are labeled. For cases with multiple entity categories in the relation description, a sentence generation method based on related words to expand and fuse templates is proposed. This enriches the entity category information in the relation description and narrows the gap between the encoder and isolated entity categories, resulting in better entity category representation and improved matching accuracy. Second, a classifier with gradient inversion and a feature extraction method based on self-attention are used. Projective distillation is employed to remove redundant information irrelevant to the relation in the original sentence, obtaining a pure relation semantic representation and further improving matching accuracy. Finally, based on a dual-tower structure, cosine similarity is used as a metric to perform fine-grained structured matching of the relation description and the original sentence, enabling entity-to-entity and sentence-to-sentence matching, avoiding the accuracy loss caused by the pre-encoding characteristics of the dual-tower structure.
Need to check novelty before this filing date? Find Prior Art