Vietnamese Event Entity Recognition Method Fusion Dictionary and Adversarial Transfer

Abstract

The present invention relates to a Vietnamese event entity recognition method integrating dictionary and adversarial transfer. The present invention takes Vietnamese as the target language, respectively takes English and Chinese as the source languages, and utilizes entity labeling information of the source language and a bilingual dictionary to improve the entity recognition effect of the target language. The invention first uses word-level adversarial migration to realize the semantic space sharing between the source language and the target language, and then integrates bilingual dictionaries for multi-granularity feature embedding to enrich the semantic representation of words in the target language, and then uses sentence-level adversarial migration to extract language-independent sequences. feature, and finally mark the entity recognition result through CRF. The experimental results on the Vietnamese news dataset show that when the source languages ​​are English and Chinese, the proposed model has improved entity recognition performance compared with the monolingual entity recognition model and the current mainstream transfer learning model. Compared with the monolingual entity recognition model, the F1 value increased by 19.61 and 18.73, respectively.

Description

technical field [0001] The invention relates to a Vietnamese event entity recognition method integrating dictionaries and confrontation migration, and belongs to the technical field of natural language processing. Background technique [0002] The goal of Vietnamese event entity recognition is to identify and assign specific types of labels in Vietnamese news texts, such as names of people, places, organizations, and specific political concepts. At present, most event entity recognition systems use a combination of BiLSTM-CRF based on bidirectional long short-term memory (BiLSTM) network and conditional random field (CRF) for entity recognition. The performance is very low on the task of low-resource linguistic entity recognition. The current method with better effect on the task of low-resource language event entity recognition is to use the idea of ​​transfer learning, that is, to use the annotation information of high-resource languages ​​to improve the labeling effect o...

AI Technical Summary

Problems solved by technology

[0003] Multi-task learning is that all tasks share a coding layer, and knowledge transfer can be performed through the shared coding layer. However, due to the different sequence structures of different languages, when encoding the language information of two different resources at the same time, the encoder cannot guarantee to extract language-independent information. The sequence information of the high-resource language can be better transferred to the label information of the high-resource language; the word-level confrontation realizes the bilingual word embedding representation and only performs confrontation training on the pre-trained word vectors of the two languages ​​to map the two languages ​​​​into the same semantic space. The sequence feature information of the two languages ​​is ignored, and the sequence features of the source language cannot be fully used to assist the target language for entity recognition; bilingual dictionaries realize bilingual word embedding representations, and use large-scale bilingual dictionaries to align the word vector spaces of the source language and the target language, thereby Migrate the source language annotation information to the target language space, but it is relatively difficult to artificially construct a large-scale bilingual dictionary and this method does not consider the polysemy problem of bilingual translation; the two-layer confrontation transfer is based on the BiLSTM-CRF network, using word-level confrontation transfer Integrate two languages ​​into the same semantic space, and use sentence-level adversarial transfer to extract language-independent sequence features, but the semantic representation of target language words is single and the effect of extracting language-independent sequence features is poor

Images