ANAS: Sentence Similarity Calculation Based on Automatic Neural Architecture Search

Sentence similarity calculation is one of several research topics in natural language processing. It has been widely used in information retrieval, intelligent question answering and other fields. Traditional machine learning methods generally extract sentence features through manually defined feature templates and then perform similarity calculations. This type of method usually requires more human intervention and constantly has to deal with domain migration problems. Automatically extracting sentence features using certain deep learning algorithms helps to solve domain migration problems. The training data is of the domain is required to complete the extraction process. However, the neural network structure design in the deep learning model usually requires experienced experts to carry out multiple rounds of the tuning design phase. Using an automatic neural architecture search (ANAS) technology deals with all the network structure design problems. This paper proposes a sentence similarity calculation method based on neural architecture search. This method uses a combination of grid search and random search. The method in this paper is tested on the Quora “Question-Pairs” data set and has an accuracy of 81.8%. The experimental results show that the method proposed in this paper can efficiently and automatically learn the network structure of the deep learning model to achieve high accuracy.