Few-time text classification method under meta-learning framework based on measurement

Abstract

The invention discloses a few-time text classification method under a meta-learning framework based on measurement, which comprises the following steps: in an input layer, inputting a support instanceand a query instance; in a word embedding layer, mapping discrete words into a vector space by searching a pre-trained word embedding table; optimizing, at a context encoder layer, a local representation of each word in sentences supporting instances and query instances by considering contexts; in the bidirectional attention layer, firstly coupling a query instance with each support instance, andthen generating matching information between the query instance and each support instance; in the model layer, forming feature vectors for the query instances and the support instances, and calculating weights of the support instances by a given query instance-level attention module to dynamically generate a prototype; at an output layer, providing prediction for query instances by measuring similarity scores between queries and prototypes. According to the method, a few-time text classification framework using a bidirectional attention mechanism and cross-class knowledge is provided, so thatthe few-time text classification method is more effective.

Description

technical field [0001] The invention belongs to the technical field of natural language processing in artificial intelligence, and relates to a few-shot text classification method under a measure-based meta-learning framework. Background technique [0002] Text classification is a key task in natural language processing, which serves a range of downstream applications such as information retrieval and opinion mining. The task is defined as selecting an appropriate label for a given unlabeled text from a set of candidate classes. Recent developments in deep learning have aroused interest in text classification models supervised by neural networks. In fact, these methods require a large amount of labeled training data. However, acquiring such high-quality data is labor-intensive, and the manual labeling process is time-consuming. [0003] To alleviate this problem, "few-shot learning" (FSL) is proposed to train classifiers for new categories that require only a few labeled ...

AI Technical Summary

Problems solved by technology

And if the supporting instances of different classes are close to each other in the embedding space, the generated prototypes are indistinguishable

Nonetheless, existing research on few-shot text classification largely ignores this cross-category knowledge

Images