Dependency parsing field self-adaption method based on web search

Abstract

A dependency parsing field self-adaption method based on web search comprises the following steps: establishing a reorder model according to K-Best candidate dependency parsing trees, and generating a group of candidate dependency parsing trees by a reference model, wherein each candidate dependency parsing tree corresponds to a probability to define initial sequences of dependency parsing results, and on the basis of the initial sequences, the reorder model tries to use the syntax characteristics newly increased in a target field to improve the initial syntax tree sequence; evaluating the dependent relationship between words of the candidate dependency parsing trees based on the web search: at first, splitting a dependency parsing tree into the collection of a group of word relations, wherein the intensity of each word semantic relation in the collection is calculated on the basis of the web search respectively, and then the evaluation level is determined according to the weight of the word relations. Experiment results show that a maltparser trained on a penn Chinese treebank shows favorable performance in biomedicine data.

Description

technical field [0001] The invention relates to the technical field of natural language processing, in particular to a domain adaptive method using the dependency syntax of network search. Background technique [0002] Syntactic analysis is a basic research problem in natural language processing, and it plays an important supporting role in applications such as machine translation, question answering systems, and information extraction. Dependency syntax has attracted much attention for its concise form, easy labeling, and easy application. In recent years, the research on dependency parsing has made great progress. At present, the best dependency parser (sometimes, also called "dependency parsing model") has an accuracy rate of more than 90% for English dependency parsing, and the accuracy rate for Chinese dependency parsing has also reached More than 80%. However, this is only achievable performance in restricted domains such as news. However, in upper-level applicatio...

AI Technical Summary

Problems solved by technology

These new features can be learned through collaborative training (co-training) and self-training (self-training). To learn these new features, this bootstrap method is not capable of self-learning for new features

Images