Training method and system for multi-intent recognition model
A technology for identifying models and intents, which is applied in the field of training methods and systems for multi-intent identification models, which can solve the problem of not directly optimizing the objective function, and achieve the effect of improving performance and simplifying the training process.
Active Publication Date: 2022-07-12
AISPEECH CO LTD
View PDF0 Cites 0 Cited by
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
[0009] In order to at least solve the problem that the training method in the prior art does not directly optimize the objective function related to the F1 value macro-average, and cannot be directly used as the loss function of the classification model training
Method used
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View moreImage
Smart Image Click on the blue labels to locate them in the text.
Smart ImageViewing Examples
Examples
Experimental program
Comparison scheme
Effect test
Embodiment approach
[0061] As an embodiment, the method further includes:
[0062] Identify the text corresponding to the user input sentence, and convert the text into a sentence vector through an encoder in the multi-intent recognition model;
[0063] The probability value of each intent in the sentence vector is determined by the classifier in the multi-intent recognition model, and at least one predicted intent whose probability value is higher than a preset threshold is output.
[0064] In this embodiment, as image 3 As shown in the flowchart of the inference and prediction stage, the text to be predicted is input into the sentence vector encoder in the trained multi-intent recognition model, so as to obtain the sentence vector of the text to be predicted.
[0065] The classifier in the multi-intent-based recognition model determines the probability value of each intention in the sentence vector of the text to be predicted. For example, the probability value of "Express field" in "Yes, I ...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More PUM
Login to View More Abstract
Embodiments of the present invention provide a training method for a multi-intent recognition model. The method includes: encoding the original labeled training data through an encoder to obtain a sentence vector; determining the probability of the true example, true negative example, false negative example and false positive example of each intention in the sentence vector by a classifier; based on the true example, True negative examples, false negative examples, and false positive examples determine the differentiable soft-f loss function; use the soft-f loss function to back-propagate the multi-intent recognition model to optimize the parameters of the classifier and encoder until the multi-intent recognition model is trained. The intent recognition model training is complete. The embodiment of the present invention also provides a training system for a multi-intent recognition model. The embodiment of the present invention modifies the calculation method of the F1 value and constructs a differentiable loss function, which means that these back-propagation algorithms can be used for optimization, which greatly simplifies the training process and improves the performance of intention field recognition.
Description
technical field [0001] The invention relates to the field of intelligent speech, in particular to a training method and system for a multi-intent recognition model. Background technique [0002] In the dialogue system, all reasonable intentions in the user's sentence need to be identified, for example, "Yes, I want to send express", two intentions of "confirm" and "send express" need to be identified. Usually the problem of multi-intent recognition can be modeled as a multi-label classification problem, in the form of one-vs-all, training multiple binary classification models to recognize the intent. [0003] The Binary cross entropy loss function is usually used, which is the most commonly used loss function for training two-class models; the Hinge loss function is also used, which is a two-class model for training the largest grid classification. loss functions, such as Support Vector Machines (SVMs). Compared with the cross-entropy loss, the hinge loss usually brings be...
Claims
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More Application Information
Patent Timeline
Login to View More Patent Type & Authority Patents(China)
IPC IPC(8): G06F16/35G06F16/332G06N3/08
CPCG06F16/353G06F16/355G06F16/3329G06N3/084
Inventor 刘枭
Owner AISPEECH CO LTD