Multi-stage unsupervised domain adaptive causal relationship identification method

Abstract

The invention discloses a multi-stage unsupervised domain adaptive causal relationship identification method. The method comprises the following steps of (1) dividing a data set, (2) carrying out pre-training by utilizing self-adaptive contrast learning, (3) performing adversarial learning in combination with knowledge distillation, (4) filtering multi-stage data to obtain a seed set, (5) filtering single-stage data to obtain a pseudo label set, (6) obtaining a subclass prototype by using a k-means clustering method, (7) enhancing introduction consistency by adopting feature-level data, and (8) carrying out self-training by using the filtered pseudo labels. According to the method, rich causal relationship knowledge in a source domain is obtained through self-adaptive comparative learning, then the source domain knowledge is migrated to a target domain in a knowledge distillation and adversarial learning mode, and multi-stage subclass prototypes and false labels of unlabeled samples are obtained through data filtering. Feature-level data enhancement is conducted by using a prototype so as to introduce consistency loss, and self-training is conducted in a target domain by using a pseudo label.

Description

technical field [0001] The invention belongs to the field of natural language processing, and relates to self-adaptive contrastive learning in natural language, adversarial transfer learning combined with knowledge distillation, and an unsupervised domain-adaptive causality recognition method based on feature-level data enhancement and self-training strategies of consistency loss. Background technique [0002] Causal recognition is fundamental to reasoning and decision making. The feature projection of causality in natural language enables machines to better understand the surrounding environment, providing key clues for downstream tasks such as logical reasoning and question answering systems. Due to the lack of background knowledge, difficulty in understanding the context and the lack of logical reasoning ability in traditional causal relationship identification methods, as human society enters an era of information explosion, traditional methods are increasingly unable to...

AI Technical Summary

Problems solved by technology

At present, there are few related studies on the unsupervised domain adaptation problem of causality, and the existing methods have the following disadvantages: (a) general unsupervised domain adaptation methods by pulling in the distance between source domain features and target domain features will produce Forgetting disaster and fuzzy classification boundary, that is, the model will lose the classification ability of the task during the migration process; (b) the existing methods do not have the ability to learn the knowledge of the target domain, which greatly limits the improvement of the model ability, when the source domain When the knowledge and data scale cannot meet the demand, the recognition effect will drop significantly; (c) the existing pseudo-labeling method is easy to introduce noise, which greatly affects the recognition ability of the model

Images