Controllable text generation method based on natural language reasoning type under condition of few samples

Abstract

The invention relates to a controllable text generation method based on a natural language reasoning type under the condition of few samples, and the method comprises the steps: converting a premise p into a hypothesis h corresponding to a specific logic relation c through a controllable text generation model, and generating a controllable text, the controllable text generation model comprises a retriever and a generator, the retriever retrieves and obtains an example most similar to the premise p from a training set, the premise p, the logic relation c and the example are all subjected to normative processing through a pre-constructed template, and the generator generates and obtains a final hypothesis h by adopting a dynamic demonstration algorithm based on the normative premise p, the logic relation c and the example. Compared with the prior art, the method has the advantages of being high in generation accuracy, capable of better adapting to current tasks and fields and the like.

Description

technical field [0001] The invention relates to the field of natural language processing, in particular to a controllable text generation method based on a natural language inference type under few samples. Background technique [0002] The natural language inference type of text generation is given a premise p, a logical relationship c, to generate a hypothesis h that conforms to the logical relationship c. There are three types of logical relationship c: [0003] Entailment: A hypothesis can be deduced from a premise to be true. [0004] Neutral: The assumption cannot be deduced from the premises to be true or false. [0005] Contradiction: A false assumption can be deduced from a premise. [0006] A simple example is the premise p: Xiao Ming is playing basketball. If it is implication, the hypothesis can be: Xiao Ming is moving. If it is neutral, the hypothesis can be: Xiaoming likes Xiaohong. If it is contradictory, the hypothesis can be: Xiao Ming is playing footb...

AI Technical Summary

Problems solved by technology

[0009] There is a flaw in the above related research techniques: a large amount of labeled data is required to train the deep learning model

[0015] (1) The related research on controllable text generation of natural language reasoning type does not consider the solution under the condition of few samples

[0016] (2) The best-performing method for automatically generating templates in paradigm learning, the results on the verification set are unreliable when used for the problem of (1)

[0017] (3) The existing demonstration learning method does not make specific vector adjustments according to the current domain and problem when selecting examples

Images