A Distant Supervised Relation Extraction Method Based on Multi-instance Cooperative Adversarial Training

Abstract

The invention discloses a remote supervisory relationship extraction method based on multi-instance cooperative confrontation training to solve the problem of low data usage efficiency in the traditional multi-instance learning framework in the remote supervisory relationship extraction task. The problem of low data usage is caused by the tendency of multi-instance learning frameworks to only focus on high-quality sentences within the bag, while ignoring a large number of potentially noisy sentences. The method of the present invention cooperates with virtual confrontation training and confrontation training to respectively constrain the noise samples in the package and the accurate characteristics of the package level, and further strengthen the model performance while solving the problem of data utilization. This method is better than some mainstream correlation algorithms in recent years.

Description

technical field [0001] The invention relates to natural language processing, in particular to a remote supervision relation extraction method based on multi-instance cooperative confrontation training. Background technique [0002] Natural Language Processing (NLP) is an interdisciplinary subject integrating linguistics and computer science. Relation Extraction (RE) is a key subtask in information extraction, and it plays a vital role in the process of building an automated knowledge base. Its main goal is to judge the relationship category between entity pairs based on the given context (Context) sentences and specified entity (Entity) pairs. For non-relational entity pairs, a special relationship category (NA) is generally used to refer to them. [0003] Traditional relational extraction models rely on a large amount of manually labeled data, and the process of obtaining these data is often extremely time-consuming and laborious. Therefore, the distance supervision meth...

AI Technical Summary

Problems solved by technology

[0005] Although multi-instance learning alleviates the data noise problem to a certain extent, it actually sacrifices a certain amount of data utilization.

Specifically, in order to obtain a more reliable bag-level representation, multi-instance learning only focuses on those sentences with high attention scores, and does not give full play to the role of a large number of sentences with low attention scores.

In fact, the attention score distribution of the sentences in the bag is a long-tailed distribution, and the attention scores of most sentences in the bag are relatively low, which means that there is a large amount of potential information that has not been utilized by the multi-instance learning framework.

Images