Sentence semantic coding method based on reinforcement learning

Abstract

The present invention relates to the technical field of artificial intelligence, natural language processing, and more particularly, to a sentence semantic coding method based on reinforcement learning. The sentence semantic encoding method based on reinforcement learning comprises the following steps: the invention realizes the reading mode similar to human behavior by reinforcement learning; andthe reinforcement learning function of the invention is to locate the text to be read in the next step. The invention is innovative in that a reinforcement learning network is introduced to learn a reading strategy similar to human reading behavior. As that human intensively reads an article, reading and processing of the text is not in disorder or fixed order, but skipping, rereading and other reading behaviors are added, so these behaviors are given to LSTM through the reinforcement learning network, so that LSTM can encode text in a more similar way to human reading behavior, which will enable the model to have better text understanding ability.

Description

technical field [0001] The present invention relates to the technical fields of artificial intelligence and natural language processing, and more specifically, relates to a sentence semantic coding method based on reinforcement learning. Background technique [0002] Semantic Representation is the basic task in NLP, so Sentence Embedding is also a very important research direction in NLP. The current modeling of Sentence Embedding can be roughly divided into two types, disordered model and sequential model. The unordered model does not use the sequence information of the symbols in the input text. For example, Joulin et al. use the bag-of-words model to classify texts. Kim.Y proposed a classic CNN-based text classification architecture. The pooling layer at the highest layer of CNN The extracted features are out of order, so it is considered that the modeling method of CNN is also out of order. In the ordered model, the classic approach is to use the RNN network such as th...

AI Technical Summary

Problems solved by technology

In general, unordered models are faster to encode but less effective; sequential models are slower to encode but generally perform better than unordered

Images