An intelligent pushing method and system based on intention recognition
By reducing data dimensionality through the LightGBM algorithm and processing user intent using the BiGRU-Attention model, the problem of inaccurate intent recognition in traditional push systems is solved, enabling more efficient and accurate push of demand solutions.
Patent Information
- Application Number
- CN202211524964.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-01
- Publication Date
- 2025-11-04
- Estimated Expiration
- 2042-12-01
AI Technical Summary
In traditional push systems, user questions often contain complex intents, have weak contextual relevance, and are often spoken in a colloquial manner, leading to irrelevant responses from intelligent customer service and low accuracy in calculating question similarity.
The LightGBM algorithm is used to reduce data dimensionality, and the BiGRU-Attention model is used to process user intent. Key information is output through the BiGRU layer and attention mechanism layer, and the most similar solution is found in the database and pushed to the user.
It improves push efficiency and accuracy, enabling it to more precisely meet user needs.
Smart Images
Figure CN115809326B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, in particular to an intelligent pushing method and system based on intention recognition. BACKGROUND
[0002] With the continuous development and popularization of mobile Internet, the needs of customers are more and more diversified. In the traditional pushing system, the user's question has the problems of complex consulting intention, weak context relevance, diverse questions, missing reference, and serious colloquialism, etc., which often causes the intelligent customer service to have low accuracy in calculating the similarity of the question, and easily causes the situation of answering the wrong question.
[0003] In order to meet the current business needs and accurately push the demand scheme that meets the needs of customers to customers, an intelligent pushing method and system based on intention recognition are designed to construct key information according to the current language information and historical consulting information of the user, recognize the user's intention, and then push the demand scheme matched with the user to the user. SUMMARY
[0004] The present application provides an intelligent pushing method and system based on intention recognition to meet the needs and deficiencies of current technology development.
[0005] Firstly, the intelligent pushing method based on intention recognition of the present application adopts the following technical solutions to solve the above technical problems:
[0006] An intelligent pushing method based on intention recognition comprises the following steps:
[0007] S10, collect language information, text information and historical information of the user, and map the collected information to a high-dimensional space;
[0008] S20, learn the high-dimensional features of the information using the LightGBM algorithm, and then remove the unimportant features according to the feature importance and reduce the data dimension;
[0009] S30, based on the BiGRU-Attention model, process the output data of step S20 to output the key information reflecting the user's intention;
[0010] S40, find the demand scheme similar to the key information from the database, and intelligently push the demand scheme with the highest similarity to the user.
[0011] Optionally, the BiGRU-Attention model comprises an input layer, a hidden layer and an output layer, the hidden layer comprises a BiGRU layer and an attention mechanism layer, wherein,
[0012] The BiGRU layer is composed of a forward GRU and a backward GRU stacked together, and there is no information flow between the forward GRU and the backward GRU,
[0013] The attention mechanism layer includes a full connection layer and a Tanh layer, the forward GRU and the backward GRU are connected to the full connection layer at the same time, the Tanh layer introduces the weight parameters of the forward GRU and the backward GRU, and the output results of the full connection layer are calculated through the activation function.
[0014] Further, the calculation process of the involved hidden layer is as follows:
[0015] Step 1), assuming that the input of time step t is x t , the activation function of the Tanh layer is θ, in the BiGRU layer, the forward GRU state at t time is The backward GRU state is As shown in formula (1) and formula (2),
[0016]
[0017]
[0018] Where, W fx , W fh , W bx , W bh are weight parameters, b fh , b bh are corresponding biases;
[0019] Step 2), the weight parameters α and β of the forward GRU and the backward GRU are introduced in the Tanh layer, and the output result GRU_O t of the hidden layer is calculated by formula (3)
[0020]
[0021] Where, W q is the output layer weight parameter, b q is the output layer bias, α and β represent the weights of the forward output and the backward output respectively, and the final GRU_O t is obtained by combination.
[0022] Preferably, the values of α and β can be the same or different.
[0023] Further, the output data of the BiGRU-Attention model involved in step S20 is processed, and the specific processing process is as follows:
[0024] S31, the input layer of the BiGRU-Attention model pre-processes the output data of step S20 to process the output data into a sequence vector form that can be directly received and processed by the BiGRU layer;
[0025] S32, the hidden layer of the BiGRU-Attention model processes the sequence vector output by step S31 and outputs the weighted result of the sequence vector;
[0026] S33, the output layer of the BiGRU-Attention model uses a softmax function to calculate the output result of step S32 and outputs the key information reflecting the user's intention.
[0027] Secondly, the intelligent push system based on intention recognition of the application solves the above technical problems by adopting the following technical solutions:
[0028] An intelligent push system based on intention recognition comprises:
[0029] An information collection module is configured to collect language information, text information and historical information of a user, and map the collected information to a high-dimensional space;
[0030] A LightGBM module is configured to learn high-dimensional features of the information using a LightGBM algorithm, and then remove unimportant features and reduce the data dimension according to the feature importance;
[0031] A BiGRU-Attention model is configured to process the output data of the LightGBM module and output key information reflecting the user's intention;
[0032] A search and push module is configured to search for a demand scheme similar to the key information from a database and intelligently push the demand scheme with the highest similarity to the user.
[0033] Optionally, the BiGRU-Attention model comprises an input layer, a hidden layer and an output layer, the hidden layer comprises a BiGRU layer and an attention mechanism layer, wherein,
[0034] The BiGRU layer is composed of a forward GRU and a backward GRU stacked together, and there is no information flow between the forward GRU and the backward GRU,
[0035] The attention mechanism layer comprises a full connection layer and a Tanh layer, the forward GRU and the backward GRU are connected to the full connection layer at the same time, the Tanh layer introduces the weight parameters of the forward GRU and the backward GRU, and calculates the output result of the full connection layer through an activation function.
[0036] Further optionally, the calculation process of the involved hidden layer is as follows:
[0037] Step 1), assuming that the input at time step t is x t , the activation function of the Tanh layer is θ, in the BiGRU layer, the forward GRU state at time t is The backward GRU state is As shown in formula (1) and formula (2),
[0038]
[0039]
[0040] Wherein, W fx , W fh , W bx , W bh are weight parameters, b fh , b bh are corresponding biases;
[0041] Step 2), introducing the weight parameters α and β of the forward GRU and the backward GRU in the Tanh layer, and calculating the output result GRU_O of the hidden layer by formula (3) t ,
[0042]
[0043] Wherein, W q is the output layer weight parameter, b q is the output layer bias, α and β represent the weights of the forward output and the backward output respectively, and the final GRU_O is obtained by combination t .
[0044] Preferably, the values of α and β can be the same or different.
[0045] Further optionally, the BiGRU-Attention model processes the output data of the LightGBM module, and the specific processing process is as follows:
[0046] The input layer of the BiGRU-Attention model preprocesses the output data of the LightGBM module to process the output data into a sequence vector form that can be directly received and processed by the BiGRU layer;
[0047] The hidden layer of the BiGRU-Attention model processes the sequence vector and outputs the weighted result of the sequence vector;
[0048] The output layer of the BiGRU-Attention model utilizes a softmax function to calculate the output result of the hidden layer, and outputs key information.
[0049] The intelligent pushing method and system based on intention recognition have the beneficial effects compared with the prior art:
[0050] The LightGBM algorithm is used to reduce the dimension of data features, the BiGRU-Attention model is used to process low micro data, and key information reflecting user intention is output, similar demand schemes are found from the database, and the demand scheme with the highest similarity is intelligently pushed to the user, so that the pushing efficiency and pushing accuracy can be effectively improved. BRIEF DESCRIPTION OF DRAWINGS
[0051] FIG. 1 is a method flowchart of the embodiment one of the present application; Figure 1
[0052] FIG. 2 is a module connection diagram of the embodiment two of the present application; Figure 2
[0053] FIG. 3 is a connection block diagram of the hidden layer and the output layer in the present application. Figure 3
[0054] The information of each number in the drawing represents:
[0055] 1, information collection module, 2, LightGBM module,
[0056] 3, BiGRU-Attention model, 4, finding and pushing module. DETAILED DESCRIPTION
[0057] In order to make the technical solutions, the technical problems solved and the technical effects of the present application clearer, the technical solutions of the present application are described clearly and completely in combination with specific embodiments.
[0058] Embodiment one:
[0059] In combination with the Figure 1 Taking an intelligent question and answer system as an example, the embodiment proposes an intelligent pushing method based on intention recognition, which includes the following steps:
[0060] S10, collect language information, text information and historical information of the user, and map the collected information to a high-dimensional space.
[0061] S20, use the LightGBM algorithm to learn the high-dimensional features of the information, and then remove unimportant features according to the feature importance, and reduce the data dimension.
[0062] S30, based on the BiGRU-Attention model, processing the output data of step S20 to output key information reflecting the user's intention.
[0063] In this embodiment, the BiGRU-Attention model includes an input layer, a hidden layer and an output layer, combined with the attached Figure 3 , the hidden layer includes a BiGRU layer and an attention mechanism layer, wherein: the BiGRU layer is composed of a forward GRU and a backward GRU stacked together, and there is no information flow between the forward GRU and the backward GRU; the attention mechanism layer includes a fully connected layer and a Tanh layer, the forward GRU and the backward GRU are connected to the fully connected layer at the same time, the Tanh layer introduces the weight parameters of the forward GRU and the backward GRU, and calculates the output results of the fully connected layer through the activation function.
[0064] In detail, the calculation process of the hidden layer is as follows:
[0065] Step 1), assuming that the input of time step t is x t , the activation function of the Tanh layer is θ, in the BiGRU layer, the forward GRU state at time t is , and the backward GRU state is As shown in formula (1) and formula (2),
[0066]
[0067]
[0068] , wherein W fx , W fh , W bx , W bh are weight parameters, and b fh , b bh are corresponding biases.
[0069] Step 2), introducing the weight parameters α and β of the forward GRU and the backward GRU in the Tanh layer, and calculating the output result GRU_O t of the hidden layer by formula (3)
[0070]
[0071] , wherein W q is the output layer weight parameter, b q is the output layer bias, and α and β represent the weights of the forward output and the backward output respectively, and the value is 0.5, and the final GRU_O t is obtained by combination.
[0072] It should be noted that the values of a and β can be the same or different.
[0073] In more detail, the BiGRU-Attention model processes the output data of step S20, and the specific processing process is as follows:
[0074] S31, the input layer of the BiGRU-Attention model pre-processes the output data of step S20 to process the output data into a sequence vector form that can be directly received and processed by the BiGRU layer;
[0075] S32, the hidden layer of the BiGRU-Attention model processes the sequence vector output by step S31 and outputs the weighted result of the sequence vector;
[0076] S33, the output layer of the BiGRU-Attention model uses a softmax function to calculate the output result of step S32, and outputs the key information reflecting the user's intention, which is the y t . Figure 3
[0077] S40, find a demand scheme similar to the key information from the intelligent question and answer system database, and intelligently push the demand scheme with the highest similarity to the user.
[0078] In this step, the key information contains multiple effective words, and the similarity between the demand scheme and the key information is determined according to the number of effective words in the key information contained in the demand scheme.
[0079] Embodiment two:
[0080] In combination with the accompanying Figure 2 , the embodiment proposes an intelligent push system based on intention recognition, which includes:
[0081] An information collection module 1 is used to collect language information, text information and historical information of the user, and map the collected information to a high-dimensional space;
[0082] A LightGBM module 2 is used to learn the high-dimensional features of the information using the LightGBM algorithm, and then remove unimportant features according to the feature importance and reduce the data dimension;
[0083] A BiGRU-Attention model 3 is used to process the output data of the LightGBM module 2 and output the key information reflecting the user's intention;
[0084] A search and push module 4 is used to find a demand scheme similar to the key information from the database, and intelligently push the demand scheme with the highest similarity to the user.
[0085] In this embodiment, the BiGRU-Attention model 3 includes an input layer, a hidden layer and an output layer.
[0086] In combination with the accompanying drawings Figure 3 , the hidden layer includes a BiGRU layer and an attention mechanism layer, wherein,
[0087] The BiGRU layer is composed of a forward GRU and a backward GRU stacked together, and there is no information flow between the forward GRU and the backward GRU,
[0088] The attention mechanism layer includes a fully connected layer FC layer and a Tanh layer, the forward GRU and the backward GRU are connected to the fully connected layer FC layer at the same time, the Tanh layer introduces the weight parameters of the forward GRU and the backward GRU, and calculates the output results of the fully connected layer FC layer through an activation function.
[0089] Based on the structure of the hidden layer, the calculation process is as follows:
[0090] Step 1), assuming that the input at time step t is x t , the activation function of the Tanh layer is θ, in the BiGRU layer, the state of the forward GRU at time t is The state of the backward GRU is As shown in formula (1) and formula (2),
[0091]
[0092]
[0093] Wherein, W fx , W fh , W bx , W bh are weight parameters, b fh , b bh are corresponding biases;
[0094] Step 2), introducing the weight parameters α and β of the forward GRU and the backward GRU in the Tanh layer, and calculating the output result GRU_O t of the hidden layer by formula (3)
[0095]
[0096] Wherein, W q is the weight parameter of the output layer, b q is the bias of the output layer, α and β represent the weights of the forward output and the backward output respectively, and the value is set to 0.5 this time, and the final GRU_O t is obtained by combination.
[0097] It should be noted that the values of a and β can be the same or different.
[0098] In this embodiment, the BiGRU-Attention model 3 processes the output data of the LightGBM module 2, and the specific processing process is as follows:
[0099] The input layer of the BiGRU-Attention model 3 preprocesses the output data of the LightGBM module 2 to process the output data into a sequence vector form that can be directly received and processed by the BiGRU layer;
[0100] The hidden layer of the BiGRU-Attention model 3 processes the sequence vector and outputs the weighted result of the sequence vector;
[0101] The output layer of the BiGRU-Attention model 3 calculates the output result of the hidden layer by using a softmax function, and outputs key information, that is, the y t . Figure 3
[0102] As can be seen from the above, the intelligent pushing method and system based on intent recognition can effectively improve the pushing efficiency and pushing accuracy.
[0103] The principles and implementation modes of the present application are described in detail above by using specific examples, and these examples are only used to help understand the core technical content of the present application. Based on the above specific embodiments of the present application, any improvement and modification of the present application made by those skilled in the art without departing from the principles of the present application shall fall within the scope of the patent protection of the present application.
Claims
1. An intelligent push method based on intent recognition, characterized in that... It includes the following steps: S10. Collect users' language information, text information, and historical information, and map the collected information to a high-dimensional space; S20. Use the LightGBM algorithm to learn the high-dimensional features of the information, and then remove unimportant features according to their importance to reduce the data dimensionality. The S30 BiGRU-Attention model includes an input layer, a hidden layer, and an output layer. The hidden layer includes a BiGRU layer and an attention mechanism layer. The BiGRU layer is composed of a forward GRU and a backward GRU stacked on top of each other, with no information flow between the forward GRU and the backward GRU. The attention mechanism layer includes a fully connected layer and a Tanh layer. The forward GRU and the backward GRU are connected to the fully connected layer. The Tanh layer introduces the weight parameters of the forward GRU and the backward GRU and calculates the output of the fully connected layer through an activation function. Based on the BiGRU-Attention model, the output data of step S20 is processed to output key information reflecting the user's intent. The specific processing steps are as follows: S31, the input layer of the BiGRU-Attention model preprocesses the output data of step S20 to convert it into a sequence vector form that the BiGRU layer can directly receive and process; S32, the hidden layer of the BiGRU-Attention model processes the sequence vector output in step S31 and outputs the weighted result of the sequence vector; S33, the output layer of the BiGRU-Attention model uses the softmax function to calculate the output result of step S32 and outputs key information reflecting the user's intent. S40. Search the database for requirement solutions similar to the key information, and intelligently push the requirement solution with the highest similarity to the user.
2. The intelligent push method based on intent recognition according to claim 1, characterized in that, The calculation process for the hidden layer is as follows: Step 1) Assume the input at time step t is x t The activation function of the Tanh layer is In the BiGRU layer, the forward GRU state at time t is: Backward GRU state is As shown in formulas (1) and (2), Among them, W fx W fh W bx W bh b is the weight parameter. fh b bh For the corresponding bias; Step 2) Introduce the weight parameters α and β of the forward GRU and backward GRU into the Tanh layer, and calculate the output result GRU_O of the hidden layer using formula (3). t , Among them, W q b represents the output layer weight parameters. q As the output layer bias, α and β represent the weights of the forward and backward outputs, respectively, and are combined to obtain the final GRU_O. t .
3. The intelligent push method based on intent recognition according to claim 2, characterized in that, The values of α and β can be the same or different.
4. An intelligent push system based on intent recognition, characterized in that, It includes: The information collection module is used to collect users' language information, text information, and historical information, and to map the collected information to a high-dimensional space. The LightGBM module is used to learn high-dimensional features of information using the LightGBM algorithm, and then removes unimportant features according to their importance to reduce the data dimensionality. The BiGRU-Attention model includes an input layer, hidden layers, and an output layer. The hidden layers consist of a BiGRU layer and an attention mechanism layer. The BiGRU layer is composed of a forward GRU and a backward GRU stacked vertically, with no information flow between them. The attention mechanism layer includes a fully connected layer and a Tanh layer. The forward and backward GRUs are simultaneously connected to the fully connected layer. The Tanh layer introduces the weights of the forward and backward GRUs and calculates the value using an activation function on the output of the fully connected layer. The model is used to process the output data of the LightGBM module, outputting key information reflecting the user's intent. The specific processing procedure is as follows: The input layer of the BiGRU-Attention model preprocesses the output data of the LightGBM module to convert it into a sequence vector form that the BiGRU layer can directly receive and process; the hidden layer of the BiGRU-Attention model processes the sequence vectors and outputs the weighted result of the sequence vectors; the output layer of the BiGRU-Attention model uses the softmax function to calculate the output result of the hidden layer and outputs the key information. The search and push module is used to find similar requirement solutions from the database and intelligently push the most similar requirement solution to the user.
5. The intelligent push system based on intent recognition according to claim 4, characterized in that, The calculation process for the hidden layer is as follows: Step 1) Assume the input at time step t is x t The activation function of the Tanh layer is In the BiGRU layer, the forward GRU state at time t is: Backward GRU state is As shown in formulas (1) and (2), Among them, W fx W fh W bx W bh b is the weight parameter. fh b bh For the corresponding bias; Step 2) Introduce the weight parameters α and β of the forward GRU and backward GRU into the Tanh layer, and calculate the output result GRU_O of the hidden layer using formula (3). t , Among them, W q b represents the output layer weight parameters. q As the output layer bias, α and β represent the weights of the forward and backward outputs, respectively, and are combined to obtain the final GRU_O. t .
6. The intelligent push system based on intent recognition according to claim 5, characterized in that, The values of α and β can be the same or different.