Entity disambiguation method based on neural network combined with knowledge description

Abstract

The invention discloses an entity disambiguation method based on a neural network combined with knowledge description, and relates to the technical field of natural language processing. The method comprises the following steps: 1, modeling by using a referring context text and a candidate entity, and calculating the similarity between the referring context text and the candidate entity; 2, modeling by using text information described by candidate entity knowledge and referred context text; 3, extracting candidate entity knowledge description keywords; 4, establishing a local model of entity disambiguation; 5, establishing a global model of entity disambiguation; and 6, introducing a loss function, and training to find the target formula in the step 4. According to the short text entity disambiguation method, entity operation is carried out simultaneously from the candidate entities and the vectors of the context, the global model is used for entity disambiguation of the short text, and the optimized local model is combined, so that the problems that a global model corpus lacks and the local model lacks other entity information of the text are solved.

Description

technical field [0001] The invention relates to the technical field of natural language processing, in particular to a method for entity disambiguation based on a neural network combined with knowledge description. Background technique [0002] Entity disambiguation is a sub-task of entity linking in NLP. For example, the word "House of Flying Daggers" can be a song sung by singer Eason Chan, or a movie directed by director Zhang Yimou, or a Chinese idiom or a pipa song . [0003] At present, the main entity disambiguation methods are mainly divided into machine learning and deep learning methods: traditional machine learning methods mainly include retrieval-based methods, sorting model-based methods, space vector-based methods, and topic model-based methods according to model classification. etc.; Entity disambiguation based on deep learning is the current mainstream research method, and it can be divided into local model and global model according to the information used ...

AI Technical Summary

Problems solved by technology

However, these methods represent entities by training entity vectors offline, which makes the disambiguation model have a natural information representation loss problem.

[0004] The local model uses the local textual context information around the entity reference to independently solve the ambiguity problem of each entity reference, ignoring the semantic connection between different entities in the same document; the global model usually constructs the entity reference and its candidate entities in the document It is a graph structure, in which the nodes are entities, and the edges represent their relationships. Using the relationship between entity references, candidate entities, and entity references and candidate entities for collaborative reasoning, there is also the problem of lack of data. At the same time, due to the joint reasoning mechanism, the calculation amount Huge, when the document is long, the entity graph contains too many nodes, resulting in a large amount of calculation

Images