RNN Language Processing Model-Driven Spoken Dialogue System Modeling Method.

Xia Zhu,Xin Ning

doi:10.1155/2022/6993515

Abstract

Speech recognition and semantic understanding of spoken language are critical components in determining the dialogue system's performance in SDS. In the study of SDS, the improvement of SLU performance is critical. By influencing the factors before and after the input text sequence information, RNN predicts the next text information. The RNN language model's probability score is introduced, and the recognition's intermediate result is rescored. A method of combining cache RNN models to optimize the decoding process and improve the accuracy of word sequence probability calculation of language model on test data is proposed to address the problem of mismatch between test data and training data in recognition. The results of the experiments show that the method proposed in this paper can effectively improve the recognition system's performance on the test set. It has the potential to achieve a higher SLU score. It is useful for future research on spoken dialogue and SLU issues.

Highlights

Spoken Dialogue System (SDS) is composed of automatic speech recognition, spoken language understanding (SLU), dialogue management, language generation, speech synthesis, and so on
In order to prove the practicability of the Recurrent Neural Networks (RNN) model proposed in this paper, this paper uses the data set widely used in experiments such as SLU, that is, the experiment of oral semantic understanding on the data set of Air Travel Information System (ATIS)
We propose the Deep Neural Network (DNN) structure of feature fusion through the Neural Networks (NN) structure described earlier and combine the features of DNN to concentrate multiple features in a time node for calculation. is paper introduces the experiments of RNN, Convolution Neural Networks (CNN), and Condition Random Fields (CRF) network structures on the database. e results show that the proposed RNN with external memory is effective for SLU tasks, has good accuracy and robustness, and has certain advantages in training error and convergence speed

Summary

Introduction

Spoken Dialogue System (SDS) is composed of automatic speech recognition, spoken language understanding (SLU), dialogue management, language generation, speech synthesis, and so on. E SLU system is a key component of SDS, which aims to help the computer “understand” the text recognized by the speech recognition module [1]. An important task in SLU is to automatically extract and classify semantic intentions or to fill in a set of parameters or slots embedded in the semantic framework so as to achieve the goal in manmachine dialogue [2]. Natural language understanding is an important factor that determines the usability and naturalness of human-computer SDS. Due to the randomness of user sentences in spoken dialogue and the imperfection of speech recognition module, the traditional whole sentence analysis method cannot achieve a correct understanding of user sentences, but it is necessary to introduce a robust understanding mechanism to extract the key information in sentences.

Methods

Results

Conclusion