Method for realizing distributed intelligent interaction by adopting natural language, and system

Abstract

The invention relates to the technical field of intelligent interaction, and discloses a method for realizing distributed intelligent interaction by adopting a natural language. The method comprises the steps of: receiving a user voice signal, and performing pre-emphasis, windowing and framing processing on user voice; detecting noise and a mute frame in the user voice signal by using a VAD detection method based on step-by-step segmentation; performing feature extraction on the preprocessed user voice signal by using a WF-MFCC algorithm to obtain WF-MFCC features of the user voice signal; extracting the WF-MFCC features by using an LSTM model combining the weight and a self-attention mechanism to obtain user semantic features; encoding and decoding the user semantic features by using an encoding-decoding process based on the information weight; and realizing interactive generation of natural language by using an attention adjustment process based on user information. The invention further provides a system for distributed intelligent interaction by adopting the natural language. According to the method and the system, intelligent interaction based on the natural language is realized.

Description

technical field [0001] The invention relates to the technical field of intelligent interaction, in particular to a method and system for realizing distributed intelligent interaction using natural language. Background technique [0002] With the continuous progress of human-computer interaction technology and the development of information technology represented by Internet technology, dialogue-based interaction technology has been increasingly valued and used more widely. People obtain a large amount of information closely related to life and work on the Internet, and language is the most direct type of information. How to feed back appropriate and important information from a large number of language information is particularly important. As a basic technology that has a great impact on human production and life in the information age, human-computer interaction has been widely valued. [0003] There are two traditional intelligent interaction strategies using natural lan...

AI Technical Summary

Problems solved by technology

However, with the continuous complexity of application scenarios and the continuous improvement of user requirements for interactive experience, the method based on rule matching cannot be performed in real time.

Although the search-based chatbot can guarantee the grammatical rationality and fluency of the reply sentence, it is limited by the richness of the content in the training data. If the reply the user needs is not in the dialog database, the dialog system cannot give more satisfactory answer

[0004] Existing deep learning methods mainly use an end-to-end automatic speech recognition model to recognize the user's voice, and then use the traditional intelligent interaction strategy using natural language to interact. However, the existing automatic speech recognition model has the following problems. On the one hand, The CTC speech recognition model assumes that the output units are independent of each other, but in fact this is unreasonable for speech recognition that is closely related to the context. On the other hand, the speech recognition model based on the attention mechanism will irregular input and output Flexible alignment, but usually speech recognition has a strictly monotonic input and output, so it is possible for the recognition result to contain deletion and insertion errors

Moreover, the existing language generation model mainly uses the encoder-decoder structure to train the input corpus. Due to the lack of external knowledge, the model can only learn its own information from the social corpus produced by different people, while the existing decoding model Always choose the words with higher probability as the output, even if these words may have semantic errors or inconsistencies

Therefore, many current end-to-end language generation models cannot provide coherent and informative responses that contain personal characteristics.

Images