Causal analysis method and system for proving evidence based on prompts

Abstract

The invention relates to a causal analysis method and system for proving evidence based on prompts, and belongs to the field of natural language processing. The method mainly comprises the following steps of: (1) formalizing a structural causal model of an interaction relationship of each variable in a proving process based on a prompt; the structural causal model totally comprises 11 key variables, and describes causal relationships in four key processes of model pre-training, prompt selection, natural linguistic test set generation and performance evaluation. And (2) based on a structural causal model, identifying a real causal relationship expected to be evaluated and back door paths of three confused evaluation results, and analyzing three deviations caused by the three back door paths: prompt preference deviation, instance natural linguistic deviation and sampling difference deviation. And (3) a causal intervention method for eliminating probing process deviation based on a backdoor criterion. The method can effectively identify, understand and eliminate the deviation in the probing process, and obtains a stable, accurate and reliable evaluation result.

Description

technical field [0001] The invention relates to a causal analysis method and system for probing based on prompts, belonging to the field of natural language processing. Background technique [0002] With the great success of large-scale pre-training models in the field of natural language processing, many studies have focused on exploring what knowledge is contained in existing pre-training models. Among them, prompt-based probing is one of the most widely used probing methods: by constructing task-specific prompts, the pre-trained model is asked questions to evaluate the knowledge of the model on this task. For example, to evaluate whether the pre-trained model knows the birthday of "Michael Jordan", we can input "Michael Jordan was born in[MASK]" to the pre-trained model, where "Michael Jordan" is the natural languageized query input and "was "born in" is a prompt that is translated into natural language, and "[MASK]" is a placeholder for the pre-trained model to output p...

AI Technical Summary

Problems solved by technology

[0003] However, many studies have found that the current paradigm of probing using prompts is inaccurate, unstable, and unreliable.

The deviations in these probing processes will make the true ability of the pre-training model unable to be accurately evaluated, mislead our understanding of the pre-training model, and even produce wrong decisions

Therefore, in order to accurately evaluate the task-specific capabilities of pre-trained models, three core questions urgently need to be answered: 1) What are the biases in the existing prompt-based exploratory paradigm? 2) Where do these deviations come from? 3) How to eliminate these biases?

Images