A method for natural language inference based on conjunction tagging and reinforcement learning

Abstract

The present invention provides a natural language inference method based on conjunction marking and reinforcement learning, comprising the following steps: (1) training a conjunction prediction model on the conjunction prediction task; (2) inputting premise and conjecture in the natural language inference model Encode the text and use the encoder to obtain the expression matrix; (3) Embed the encoder of the conjunction prediction model into the natural language reasoning model, and send the premise and guessed word vector obtained in the encoding process of step (2) to the pre-set The trained encoder and output the expression vector; (4) interact the encoder of the natural language inference model with the encoder of the embedded conjunction prediction model in the attention mechanism to obtain an attention vector; (5) combine the attention Convert the vector to a probability distribution and output the result. By using the present invention, the accuracy rate of natural language reasoning tasks on large-scale data sets is greatly improved by transferring the knowledge learned from other supervised learning tasks.

Description

technical field [0001] The invention relates to the field of natural language processing, in particular to a natural language reasoning method based on conjunction marking and reinforcement learning. Background technique [0002] Natural language inference has always been one of the hottest and toughest problems in natural language processing. In the most popular tasks at present, given two sentences (premise sentence, conjecture sentence), we need to judge whether the logical relationship between the two sentences is entailment, neutrality or contradiction. An efficient natural language inference model can be widely used in many fields based on semantic understanding, such as dialogue robots, question answering systems and text generation systems. [0003] Early natural language reasoning tasks are based on small datasets, and the methods used are also traditional methods such as natural logic method and shadow method. Such datasets cannot support complex models with larg...

AI Technical Summary

Problems solved by technology

This approach is too simple to achieve good results on large-scale datasets

Images