Intention and slot joint recognition method based on multi-task learning

Abstract

The invention relates to an intention and slot joint recognition method based on multi-task learning. Input texts of a user, such as utterance / query are processed to output intention labels and slot labels. The method comprises the following steps of sequentially processing the text sequence input by the user through a long-term and short-term memory network and a convolutional neural network toform an LSTM-CNN shared representation feature; according to the difference between the intention label information and the slot position label information, respectively establishing a Bi-LSTM intention recognition model / slot position recognition model with an attention mechanism based on the shared representation features; and constructing a total loss function of the intention recognition modeland the slot position recognition model by using a weighted calculation method based on a gradient descent method, and performing joint optimization solution on the total loss function. The multi-tasklearning thought is applied to the construction process of the vertical dialogue system, joint recognition of the input text intention and the slot position can be achieved, and the recognition accuracy and F value of the input text intention and the slot position of the vertical dialogue system are effectively improved.

Description

technical field [0001] The invention belongs to the field of human-computer interaction, relates to natural language processing, a vertical dialogue system, etc., and particularly relates to a multi-task learning-based joint recognition method of intent and slot. Background technique [0002] Intent recognition and slot recognition can convert the user input text in the vertical dialogue system into semantic representation, and provide support for the system to take the next action, which is a key step in the natural language understanding module of the vertical dialogue system. The intent recognition task focuses on predicting the intent of the input text, and the slot recognition aims to extract semantic concepts as constraints of natural language, that is, to assign appropriate semantic labels to each word in a given input text. In the current research, intent recognition and slot recognition are usually carried out independently in a "pipeline" manner. Some recent studie...

AI Technical Summary

Problems solved by technology

In the current research, intent recognition and slot recognition are usually carried out independently in a "pipeline" manner. Some recent studies have adopted the method of joint recognition of intent and slot, but these models do not fully consider the strong correlation between intent and slot This will lead to the fact that in the process of man-machine dialogue, as the dialogue progresses, the user's intention may continue to shift, and the slot value may also change continuously, so it is necessary to carry out user intention and slot value Continuous Match Verification

Images