Intelligent outbound calling methods based on prior knowledge

Abstract

This invention discloses an intelligent outbound calling method based on prior knowledge, comprising: acquiring training data; performing feature processing on the training data; manually labeling the processed training data; building an intent recognition model; incorporating prior knowledge into the intent recognition model; inputting the labeled training data into the intent recognition model for training; and inputting the corpus to be recognized into the trained intent recognition model for intent recognition to obtain the intent recognition result. The intelligent outbound calling method based on prior knowledge provided by this invention incorporates manually designed prior information into the intent recognition model, allowing the intent recognition model to learn some key features in advance, thereby improving the accuracy of intent recognition.

Description

Technical Field

[0001] This invention relates to an intelligent outbound calling method based on prior knowledge. Background Technology

[0002] Traditional outbound calling services suffer from high customer service staff turnover, high labor costs, lack of service standards, and low efficiency, leading to difficulties in recruitment and lengthy training periods for businesses, thus increasing related costs. Therefore, AI-powered intelligent outbound calling systems are introduced to provide standardized outbound calling services. While deep neural networks are typically used in the training of NLU models for intelligent outbound calling, end-to-end deep neural networks are essentially black boxes. Although they can automatically learn some highly discriminative features, they often fit to unimportant features, causing the model to locally collapse onto undesirable features. This severely impacts the accuracy of intent recognition and degrades the user experience.

[0003] Existing AI-powered outbound calling systems have the following drawbacks: some data-sensitive departments, such as banks and government agencies, have high data confidentiality requirements, making data collection difficult and training data limited; the data quality is low and features are not obvious in the early stages of model building; identification problems directly reduce the user experience and lead to a high complaint rate. Summary of the Invention

[0004] This invention provides an intelligent outbound calling method based on prior knowledge to solve the aforementioned technical problems, specifically adopting the following technical solution:

[0005] An intelligent outbound calling method based on prior knowledge includes the following steps:

[0006] Obtain training data;

[0007] Perform feature processing on the training data;

[0008] The processed training data is manually labeled;

[0009] Build an intent recognition model;

[0010] Incorporate prior knowledge into the consciousness recognition model;

[0011] The labeled training data is input into the intent recognition model to train it;

[0012] The corpus to be recognized is input into the trained intent recognition model to perform intent recognition and obtain the consciousness recognition result.

[0013] Furthermore, the specific method for incorporating prior knowledge into the consciousness recognition model is as follows:

[0014] Replace the network weights of the intent recognition model with the network weights of another trained model.

[0015] Furthermore, in replacing the network weights of the intent recognition model with the network weights of another pre-trained model, the other pre-trained model is the BERT model.

[0016] Furthermore, after manually labeling the processed training data, the intelligent outbound calling method based on prior knowledge also includes:

[0017] Extract feature words from training data;

[0018] The specific method for training the intent recognition model by inputting the labeled training data is as follows:

[0019] The labeled training data and its corresponding feature words are simultaneously input into the intent recognition model for training.

[0020] Furthermore, the specific method for extracting feature words from the training data is as follows:

[0021] Build a feature word library, which contains several keywords;

[0022] The training data is matched with a feature word library, and the keywords contained in the training data are extracted as its feature words.

[0023] Furthermore, after manually labeling the processed training data, the intelligent outbound calling method based on prior knowledge also includes:

[0024] Add location labels to the processed training data.

[0025] Furthermore, the specific method for adding location labels to the processed training data is as follows:

[0026] The training data is divided into three regions: front, middle, and back.

[0027] Based on the location of key information in the training data, assign at least one location label from the front, middle, and back of the training data.

[0028] Furthermore, after inputting the corpus to be recognized into the trained intent recognition model to obtain the consciousness recognition result, the intelligent outbound calling method based on prior knowledge also includes:

[0029] The auxiliary recognition model is trained by inputting the corpus to be recognized to obtain the auxiliary recognition result;

[0030] The final recognition result is obtained by combining the results of consciousness recognition and auxiliary recognition.

[0031] Furthermore, the auxiliary recognition model is an identity recognition model.

[0032] The advantage of this invention lies in the provision of an intelligent outbound calling method based on prior knowledge, which adds artificially designed prior information to the intent recognition model, allowing the intent recognition model to learn some key features in advance, thereby improving the accuracy of consciousness recognition. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of an intelligent outbound calling method based on prior knowledge according to the present invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0035] like Figure 1 The diagram illustrates an intelligent outbound calling method based on prior knowledge, comprising the following steps: S1: Acquire training data. S2: Perform feature processing on the training data. S3: Manually label the processed training data. S4: Build an intent recognition model. S5: Add prior knowledge to the intent recognition model. S6: Input the labeled training data into the intent recognition model for training. S7: Input the corpus to be recognized into the trained intent recognition model for intent recognition to obtain the intent recognition result. This intelligent outbound calling method based on prior knowledge adds manually designed prior information to the intent recognition model, allowing the model to learn some key features in advance, thereby improving the accuracy of intent recognition. The following details the above steps.

[0036] For step S1: Obtain training data.

[0037] Specifically, training data should be collected in advance, depending on the application scenario.

[0038] For step S2: Perform feature processing on the training data.

[0039] Feature processing refers to the process of transforming training data into feature vectors. Feature processing is a common technique in this field and will not be elaborated upon here.

[0040] For step S3: Manually label the processed training data.

[0041] In this application, the training data is manually labeled, and the intention label of each training data is manually marked.

[0042] For step S4: Build an intent recognition model.

[0043] In this application, the intent recognition model is able to identify the intent of spoken text through an awareness recognition algorithm.

[0044] For step S5: Add prior knowledge to the consciousness recognition model.

[0045] Specifically, the method for incorporating prior knowledge into the consciousness recognition model is as follows:

[0046] Replace the network weights of the intent recognition model with the network weights of another trained model.

[0047] In this application, when the network weights of the intent recognition model are replaced with the network weights of another pre-trained model, the other pre-trained model is the Google BERT pre-trained model, which is trained on massive amounts of data and can better learn the common features in the corpus.

[0048] For step S6: Input the labeled training data into the intent recognition model to train it.

[0049] For step S7: Input the corpus to be recognized into the trained intent recognition model to perform intent recognition and obtain the consciousness recognition result.

[0050] As a preferred implementation, after manually labeling the processed training data, the intelligent outbound calling method based on prior knowledge further includes extracting feature words from the training data. Furthermore, the specific method for inputting the labeled training data into the intent recognition model for training is as follows: the labeled training data and its corresponding feature words are simultaneously input into the intent recognition model for training. In intent recognition, the feature words representing intent in the corpus and the entire corpus are used as two inputs to the neural network, and the information from the two inputs is fused at the back end of the network. This achieves the goal of focusing on both local and overall aspects.

[0051] Specifically, the method for extracting feature words from training data is as follows:

[0052] Build a feature word library containing several keywords. Match the training data with the feature word library, and extract the keywords contained in the training data as its feature words.

[0053] As a preferred implementation, after manually labeling the processed training data, the intelligent outbound calling method based on prior knowledge further includes adding location tags to the processed training data.

[0054] It is understandable that in a given intent corpus, key information is not distributed throughout the entire corpus, but rather concentrated in a specific part of the intent. If the intent recognition model can quickly identify the location of this key information and thus give more attention to that part, it will greatly improve the efficiency of semantic recognition. Therefore, in this application, after labeling the intent with intent tags, the corpus is also manually labeled with location tags. The intent recognition model is trained using training data that is simultaneously labeled with intent and location tags, enabling the model to quickly identify the location of key information within the intent.

[0055] As a preferred implementation method, the specific method for adding location labels to the processed training data is as follows:

[0056] The training data is divided into three regions: front, middle, and back. Based on the location of key information in the training data, at least one location label from the front, middle, and back regions is assigned to each region.

[0057] Understandably, the way training data is divided can be further refined as needed.

[0058] As a preferred implementation, after inputting the corpus to be recognized into a trained intent recognition model to obtain an intent recognition result, the intelligent outbound calling method based on prior knowledge further includes: inputting the corpus to be recognized into a trained auxiliary recognition model to obtain an auxiliary recognition result. The final recognition result is obtained by combining the intent recognition result and the auxiliary recognition result. Specifically, the auxiliary recognition model is an identity recognition model.

[0059] Understandably, outbound calling systems need to determine if the caller is the intended recipient. For non-representative intents, auxiliary tasks can be added to determine relationships such as kinship, colleague, and friend as supplementary classifications. Specifically, relationship data can be obtained from public platforms as prior knowledge, containing various labeled relationship data. This type of data is usually large in volume and well-labeled. This data can be trained using an auxiliary learning process. In this way, even subtle features like kinship and colleague can be recognized by the auxiliary model, thus strengthening the recognition of non-representative intents. Auxiliary learning can add weakly defined prior knowledge to the model, improving its recognition capabilities.

[0060] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. An intelligent outbound calling method based on prior knowledge, characterized in that, Includes the following steps: Obtain training data; The training data is then subjected to feature processing. The processed training data is then manually labeled. Build an intent recognition model; Prior knowledge is added to the intent recognition model; The labeled training data is input into the intent recognition model for training. The text to be recognized is input into the trained intent recognition model to perform intent recognition and obtain the intent recognition result. The specific method for incorporating prior knowledge into the intent recognition model is as follows: Replace the network weights of the intent recognition model with the network weights of another trained model; In the process of replacing the network weights of the intent recognition model with the network weights of another pre-trained model, the other pre-trained model is the BERT model. After manually labeling the processed training data, the intelligent outbound calling method based on prior knowledge further includes: Extract feature words from the training data; The specific method for inputting the labeled training data into the intent recognition model for training is as follows: The labeled training data and its corresponding feature words are simultaneously input into the intent recognition model for training. The specific method for extracting feature words from the training data is as follows: Construct a feature word library, which contains several keywords; The training data is matched with the feature word library, and the keywords contained in the training data are extracted as its feature words; After manually labeling the processed training data, the intelligent outbound calling method based on prior knowledge further includes: Add location labels to the processed training data; The specific method for adding location labels to the processed training data is as follows: The training data is divided into three regions: front, middle, and back. Based on the location of key information in the training data, assign at least one of the following location labels to the training data: front, middle, and back. After inputting the corpus to be recognized into the trained intent recognition model to obtain the intent recognition result, the intelligent outbound calling method based on prior knowledge further includes: The auxiliary recognition model is trained by inputting the corpus to be recognized to obtain the auxiliary recognition result; The final recognition result is obtained by combining the intent recognition result and the auxiliary recognition result.

2. The intelligent outbound calling method based on prior knowledge according to claim 1, characterized in that, The auxiliary recognition model is an identity recognition model.

Images