Entity and relation joint extraction system based on translation mechanism and use method

Abstract

According to the entity and relation joint extraction system based on the translation mechanism and the using method thereof, the thought of the translation mechanism is utilized by the system, and the problem of processing a symmetric relation by a traditional translation mechanism is avoided; a correct result and an error result of relation prediction are bounded by using an NA (non-relation) relation as a dynamic threshold value. In the entity labeling part, four 0 / 1 labeling devices are used for independently labeling a head entity and a tail entity, so that the problem of single entity superposition can be solved. For each group of head and tail entities, the thought of a translation mechanism is used, h + r is approximately equal to t, the head entity is subtracted from the tail entity to obtain vector representation of the relationship, then similarity calculation is performed on vectors of all the relationships in the relationship set, the similarity with the NA relationship is used as a dynamic threshold value, all the relationships with the similarity larger than that of the relationship are used as prediction results, so the problem of entity pair overlapping is solved. Experiments on a public data set show that the method of the system can reach a good level.

Description

technical field [0001] The invention relates to the technical fields of natural language processing and deep learning, in particular to a translation mechanism-based entity and relationship joint extraction system and usage method. Background technique [0002] The joint extraction of entities and relations is defined as a task of extracting structured knowledge in the form of triples (head, relation, tail) from unstructured text, which is very important in natural language processing and automatic construction of knowledge graphs. Such as the sentence "The Summer Palace is in Beijing, the capital of China", the triples that can be formed are (Summer Palace, located in, Beijing), (Beijing, located in, China), (Beijing, is the capital of..., China), (Summer Palace, located in ,China). [0003] Traditional pipeline methods regard entity recognition and relation extraction as two independent tasks, while ignoring their associations, so there is an error propagation problem. I...

AI Technical Summary

Problems solved by technology

In recent years, many researchers have focused on building joint models, and although great success has been achieved, there is currently no work that addresses three key issues well within a unified system:

This logically separated paradigm makes the task easy to handle and makes each component more flexible, but the relationship classifier in it may be misled by some entity pairs that have no relationship, because N entities will generate nearly N 2 entity pairs, but most of them do not have any relationship

[0005] The second problem is the overlap problem

Existing research approaches to this problem usually convert the task to sequence labeling or text clustering, although they have achieved good results, considering each relationship individually will inevitably ignore the connections between them, making it difficult to discover implicit triplet

[0006] The third problem is relational association

Although such a situation is common for knowledge graph tasks, it is difficult to capture the correlation between relationships for joint extraction tasks that do not contain entities in the input.

Images