A service automatic push system and method based on natural language processing technology

By utilizing a service automatic push system based on natural language processing technology and employing recurrent neural networks and classifier models, the shortcomings of existing natural language text push technologies have been addressed. This system enables efficient and accurate automatic push of internet services, thereby improving the efficiency of matching user needs.

CN109614486BActive Publication Date: 2024-11-26YUJIE ORIENTAL (BEIJING) TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201811436561.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2018-11-28
Publication Date
2024-11-26
Estimated Expiration
2038-11-28

AI Technical Summary

Technical Problem

Existing service push technologies are mostly based on internet user behavior records and user identity characteristics, lacking precise push solutions based on natural language text, making it difficult to achieve automated and precise push of internet services.

Method used

The system employs an automatic service push system based on natural language processing technology, which includes a vocabulary maintenance module, text acquisition, preprocessing, a model module, and an interaction module. It utilizes recurrent neural networks and classifier models to automatically identify the matching degree between text and keywords by training the neural network model, and then decides whether to push the service.

Benefits of technology

It enables efficient and accurate service delivery based on natural language text, improving the matching efficiency between consumers and services and avoiding the drawbacks of consumers manually searching for services and service providers passively waiting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN109614486B_ABST
    Figure CN109614486B_ABST
Patent Text Reader

Abstract

The present invention discloses a service automatic push system and method based on natural language processing technology. The system includes: M1, vocabulary maintenance module; M2, text acquisition module; M3, text preprocessing module; M4, model module; M5, interaction module; M6, storage module. Attached Figure 1 is a diagram of the system architecture. The method includes learning and use modes. The learning mode includes the following steps: S1, keyword table maintenance; S2, text data acquisition; S3, data annotation; S4, data preprocessing; S5, data set division; S6, model construction, verification, and testing. The use mode includes the following steps: U1, real-time text acquisition; U2, text discrimination; U3, text classification; U4, service push. The left and right boxes in Attached Figure 2 are flow charts of the learning and use modes, respectively. The present invention avoids the drawbacks of consumers manually searching for services and service providers passively waiting for consumers, and greatly improves the matching efficiency of consumers and services.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a system and method for automatically pushing services based on natural language processing technology. Background Art

[0002] With the rapid development of the Internet, the amount of text information on the Internet has exploded. Text information is a kind of unstructured data, which cannot be processed according to the traditional structured data processing method, making it difficult to achieve the automatic solution of various tasks based on text information. Natural language processing is a technology that was created to address this problem, and it is dedicated to enabling computers to automatically understand and process natural language.

[0003] In recent years, a large number of Internet-based consumer service models have emerged, and disruptive innovations have been achieved in many industries. For example, Internet-based shopping services, life services, and travel services are typical representatives of this service model. This service model contains a large number of requirements for automatic recognition, classification, prediction, retrieval, etc. based on natural language. The problem to be solved by this invention is the problem of automatic and accurate push of services based on natural language processing technology.

[0004] On the one hand, existing service push technologies are mostly based on information such as Internet user behavior records and user identity characteristics, and there are basically no service push methods based on natural language text; on the other hand, natural language text on the Internet is often a more direct information that reflects user needs. Therefore, the automatic service push method based on natural language processing technology is an urgently needed technical solution. Summary of the invention

[0005] In view of the above problems, the technical problem to be solved by the present invention is to provide an efficient and accurate method for automatically pushing Internet services.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a service automatic push system and method based on natural language processing technology, which has the following characteristics.

[0007] A service automatic push system based on natural language processing technology includes the following modules: M1, a vocabulary maintenance module, which is used for adding, deleting, deduplicating, merging, replacing, etc. entries in the vocabulary; M2, a text acquisition module, which is used to obtain natural language text related to the entries in the vocabulary; M3, a text preprocessing module, which performs text cropping, uncommon word replacement, word segmentation, keyword processing, etc.; M4, a model module, which is the core algorithm module of the system, and is used to automatically identify the matching degree between text and keywords, and then decide whether to push the service; M5, an interaction module, which is used to interact with operators, mainly including parameter setting, task selection, requesting help, etc.; M6, a storage module, which is used to store data required for system operation.

[0008] As an improvement of the system described in the present invention, its model module represents text information through a recurrent neural network (RNN) or its improved structure, processes the output vector of the RNN with the word vector of the keyword in the vocabulary to form a high-dimensional real number vector that can reflect the text information, keyword information and their relationship, and inputs the vector into a classifier model for classification.

[0009] As an improvement to the system described in the present invention, the processing method of the output vector of the RNN in its model module and the keyword vector in the vocabulary can be element-by-element averaging, element-by-element summing, or concatenation.

[0010] As an improvement of the system described in the present invention, the classifier model in its model module is a two-category classification model.

[0011] A method for automatically pushing services based on natural language processing technology, the method comprising two modes: learning and using.

[0012] The learning mode of the method includes the following steps: S1, keyword table maintenance, maintaining a keyword table according to needs; S2, text data acquisition, acquiring news text related to the above keywords in the network, and forming a data pair of two elements of keyword-text; S3, data annotation, judging whether the semantics of the text obtained in the above steps conforms to the keyword semantics, if so, annotating the text as 1, otherwise annotating it as 0, and finally forming a data pair of three elements of label-keyword-text; S4, data preprocessing, preprocessing the text data in the above data pairs, including text clipping, word segmentation, replacement of uncommon words, keyword processing, etc.; S5, data set division, dividing the data obtained in S4 into a training set, a validation set and a test set according to a certain ratio; S6, building a model, and using the training set and validation set obtained in S5 for model training and model adjustment, until the performance of the model on the test set meets certain requirements and is ready for deployment and use.

[0013] The usage mode of the method includes the following steps: U1, real-time acquisition of news text related to keywords, forming a keyword-text two-element data pair; U2, using the qualified model obtained by the learning mode, automatically calculating the probability of the data obtained by U1 being classified into each label category. If the probability is greater than the model automatic classification threshold, the data is determined to belong to the label category. If the probability of each type of label is less than the model automatic classification threshold and greater than the manual judgment threshold, then the manual expert is asked to judge. If it is less than the manual expert judgment threshold, the data is discarded; U3, according to the result obtained by U2, the data is put into the database as labeled data for model training; U4, according to the result obtained by U2, decide whether to push related services.

[0014] As an improvement of the method described in the present invention, the probability threshold for determining the category label of the news text is between (0.5, 1). The larger the threshold, the higher the accuracy. Therefore, the threshold can be determined according to the requirement for the accuracy of the result.

[0015] The beneficial effects of the present invention are: using the manually annotated news data pairs to train the constructed neural network model; by deploying the trained and qualified neural network model, the labels of the news data pairs generated in real time can be automatically determined. Compared with the existing methods, this method can use natural language text information to model and automatically learn user needs, avoiding the drawbacks of consumers manually searching for services and service providers passively waiting for consumers, and greatly improving the matching efficiency of consumers and services. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The structure framework diagram of the system of the present invention.

[0017] Figure 2 It is a flow chart of the method of the present invention in two modes, wherein the left dotted box is a learning mode flow chart, and the right dotted box is a use mode flow chart. DETAILED DESCRIPTION

[0018] In order to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0019] like Figure 1As shown, in an embodiment of a service automatic push system based on natural language processing technology of the present invention, the system includes the following modules: M1, a vocabulary maintenance module, which is used for adding, deleting, deduplicating, merging, replacing, etc. of entries in the vocabulary; M2, a text acquisition module, which is used to obtain natural language text related to the entries in the vocabulary; M3, a text preprocessing module, which performs text clipping, uncommon word replacement, word segmentation, keyword processing, etc.; M4, a model module, which is the core algorithm module of the system, is used to automatically identify the matching degree between text and keywords, and then decide whether to push the service; M5, an interaction module, which is used to interact with the operator, mainly including parameter setting, task selection, requesting help, etc.; M6, a storage module, which is used to store the data required for the operation of the system. Specifically, in this embodiment, the vocabulary maintenance module can manually add, delete, and edit keywords; the interaction module can specify the size of the text acquisition amount, switch the use mode of the model, set the training parameters of the model, and request manual judgment, etc.

[0020] Preferably, the model module of the system represents the text information through a recurrent neural network (RNN) or its improved structure, processes the output vector of the RNN with the word vector of the keyword in the vocabulary to form a high-dimensional real number vector that can reflect the text information, keyword information and their relationship, and inputs the vector into a classifier model for classification. Specifically, in this embodiment, 512 unidirectional LSTMs (long short-term memory) are selected to form a layer of network, which is spliced ​​with the keyword word vector and connected to a three-layer fully connected neural network to form the architecture of the entire model.

[0021] More preferably, the processing method of the output vector of the RNN in the model module of the system and the keyword vector in the vocabulary can be element-by-element averaging, element-by-element summing or concatenation. Specifically, the processing method in this embodiment is concatenation.

[0022] More preferably, the classifier model in the model module of the system is a two-class classification model. Specifically, the two-class classification model in this embodiment is a three-layer fully connected neural network, the number of neurons in the middle layer is 10, and the number of neurons in the output layer is 2.

[0023] like Figure 2 As shown, in an embodiment of the method for automatically pushing services based on natural language processing technology of the present invention, two modes are included: learning and using.

[0024] The learning mode of the method includes the following steps: S1, keyword table maintenance, maintaining a keyword table according to needs; S2, text data acquisition, acquiring news text related to the above keywords in the network, and forming a data pair of two elements of keyword-text; S3, data annotation, judging whether the semantics of the text obtained in the above steps conforms to the keyword semantics, if so, annotating the text as 1, otherwise annotating it as 0, and finally forming a data pair of three elements of label-keyword-text; S4, data preprocessing, preprocessing the text data in the above data pairs, including text clipping, word segmentation, uncommon word replacement, keyword processing, etc.; S5, data set division, dividing the data obtained in S4 into a training set, a validation set and a test set according to a certain ratio; S6, building a model, and using the training set and validation set obtained in S5 to train and adjust the model until the performance of the model on the test set meets certain requirements. Specifically, according to the task requirements of this embodiment, the keyword table in this example maintains 128 keywords; 320,000 microblog data are obtained based on the keywords, totaling 81.4MB; the keywords with the same meaning in the text are replaced with the corresponding unified keywords in the vocabulary; after the data annotation is completed, the ratio of positive and negative categories is counted, and in order to prevent the impact of data imbalance on the model, the ratio between classes is controlled within 5 times. If it exceeds, the large category data is deleted until the ratio is reduced to within 5; the data set is divided based on keywords, 70% of the data under the keywords are classified into the training set, and 20% are classified into the training set. The data under the keywords are divided into the validation set, and the data under the remaining 10% of the keywords are divided into the test set; the word vector dimension is 100, each training step contains 16 data pairs, the model training round is 5, and the accuracy of the model on the training set and the validation set is calculated after every 40 iterations, and the accuracy change curve of the model on the training set and the validation set during the entire training process is made. After the training, the accuracy of the model on the training set is 96.7%, the accuracy on the validation set is 85.3%, and when the classification threshold is 85%, the accuracy on the test set is 88.9%.

[0025] The usage mode of the method includes the following steps: U1, real-time acquisition of news text related to keywords, forming a keyword-text two-element data pair; U2, using the qualified model obtained by the learning mode, automatically calculate the probability of the data obtained by U1 being classified into each label category. If the probability is greater than the model automatic classification threshold, the data is determined to belong to the label category. If the probability of each type of label is less than the model automatic classification threshold and greater than the threshold for manual discrimination, then the manual expert is asked to judge. If it is less than the threshold for manual expert discrimination, the data is discarded; U3, according to the result obtained by U2, the data is put into the database as labeled data for model training; U4, according to the result obtained by U2, decide whether to push related services. Specifically, in this embodiment, obtain new text data of any keyword in the keyword table; take the automatic classification threshold as 85%, if the model output value is greater than 85%, it is automatically classified into this category, if it is between 70% and 85%, it is asked for manual discrimination, and if it is less than 70%, it is discarded.

[0026] Preferably, the probability threshold for distinguishing the category label of the news text is between (0.5, 1). The larger the threshold, the higher the accuracy. Therefore, the threshold can be determined according to the accuracy requirement of the result. Specifically, in this embodiment, according to the task requirements, 85% is selected as the threshold for automatic classification of the model, and 70% is the threshold for manual discrimination.

[0027] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Therefore, equivalent changes made according to the scope of the patent application of the present invention are still within the scope covered by the present invention.

Claims

1. An automatic service push system based on natural language processing technology, It is characterized in that The system includes the following modules: M1, vocabulary maintenance module, used for adding, deleting, removing duplicates, merging and replacing entries in the vocabulary; M2, a text acquisition module, used to acquire natural language text related to an entry in the vocabulary; M3, text preprocessing module, performs text clipping, rare word replacement, word segmentation, and keyword processing; M4, model module, is the core algorithm module of the system, which is used to automatically identify the matching degree between text and keywords, and then decide whether to push the service; the model module represents the text information through a recurrent neural network (RNN), and processes the output vector of the RNN with the word vector of the keyword in the vocabulary to form a high-dimensional real number vector that can reflect the text information, keyword information and their relationship. The relationship refers to whether the text semantics conforms to the keyword semantics, and the vector is input into a classifier model for classification; the processing method of the output vector of the RNN and the keyword word vector in the vocabulary is element-by-element averaging, element-by-element summing or splicing; M5, interactive module, used to interact with operators, including parameter setting, task selection, and help request functions; M6, storage module, used to store data required for system operation.

2. The system according to claim 1, It is characterized in that The classifier model is a two-class classification model.

3. A method for automatically pushing services based on natural language processing technology, It is characterized in that The method includes two modes: learning and using; The learning mode of the method comprises the following steps: S1. Keyword table maintenance: maintain a keyword table according to the needs; S2, text data acquisition, obtaining news texts related to the above keywords on the Internet to form a data pair of keyword-text elements; S3, data labeling, judging whether the semantics of the text obtained in the above steps conform to the semantics of the keyword, if so, labeling the text as 1, otherwise labeling it as 0, and finally forming a data pair of the three elements of label-keyword-text; S4, data preprocessing, preprocessing the text data in the above data pairs, including text clipping, word segmentation, rare word replacement, and keyword processing; S5, data set division, divide the data obtained in S4 into training set, validation set and test set according to a certain ratio; S6: Build a model and use the training set and validation set obtained in S5 to train and adjust the model until the performance of the model on the test set meets certain requirements and is ready for deployment; The usage mode of the method comprises the following steps: U1, real-time acquisition of news text related to keywords, forming a keyword-text data pair; U2, using the qualified model obtained by the learning mode, automatically calculates the probability of the data obtained by U1 being classified into each label category. If the probability is greater than the model automatic classification threshold, the data is judged to belong to the label category. If the probability of each label is less than the model automatic classification threshold and greater than the threshold for manual judgment, then the manual expert judgment is requested. If it is less than the threshold for manual expert judgment, the data is discarded. U3, based on the result of U2, puts the data into the database as labeled data for model training; U4, decides whether to push related services based on the results obtained in U2.

4. The method according to claim 3, It is characterized in that The threshold value is between (0.5, 1). The larger the threshold value, the higher the accuracy. Therefore, the threshold value can be determined according to the requirement for the accuracy of the result.