A dynamically updated intelligent customer service knowledge base system

Abstract

The application discloses a kind of dynamically updated wisdom customer knowledge base systems, including wisdom customer knowledge base, knowledge point creation module, text preprocessing module, keyword extraction module, unstructured data translation module, knowledge confusion module and knowledge confusion detection module;Knowledge point creation module forms entry knowledge point according to the user question extracted;Keyword extraction module extracts the special keyword in text information;Unstructured data translation module identifies the text area in user picture information, and outputs text information;Knowledge confusion module finds confusion knowledge point and high-quality confusion knowledge point;Knowledge confusion detection module carries out confusion detection, completes the dynamic update of knowledge base;The system can identify user voice, text, picture and other unstructured questions, can carry out confusion detection to knowledge entry in knowledge base, form dynamic knowledge base, improve user experience, improve the statistical accuracy of artificial user question.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a dynamic updating intelligent customer service knowledge base system. BACKGROUND

[0002] The knowledge base is a system (or expert system) based on knowledge and intelligence, not all intelligent programs have a knowledge base, only knowledge-based systems have a knowledge base. Many application programs use knowledge, some of which have reached a very high level, but these application programs may not be knowledge-based systems, and they do not have a knowledge base. The general application program is to implicitly encode the problem solving knowledge in the program, while the knowledge-based system explicitly expresses the problem solving knowledge of the application field and forms a relatively independent program entity. At present, the knowledge base creation cannot identify the creation of knowledge points in real time according to user problems, the creation efficiency of the knowledge base is low, the manual comparison time is long, and it cannot adapt to the demand of dynamic updating of the knowledge base. In addition, the current knowledge base cannot identify the non-structured data such as voice, picture and the like sent by the user, resulting in poor user experience. SUMMARY

[0003] In order to overcome the defects of the prior art, the present application provides a dynamic updating intelligent customer service knowledge base system, which can automatically identify user voice, text, picture and other non-structured questions, use intelligent word segmentation technology to identify user intent in all directions, automatically form question and answer knowledge points for the problems and artificial solutions proposed by the user, and expand the problems to form knowledge entries. The system can perform confusion detection on the knowledge entries in the library to form a dynamic knowledge base, improve user experience, and improve the statistical accuracy of artificial user problems.

[0004] The purpose of the present application can be achieved by the following technical solutions:

[0005] A dynamic updating intelligent customer service knowledge base system, comprising an intelligent customer service knowledge base, a knowledge point creation module, a text preprocessing module, a keyword extraction module, a non-structured data translation module, a knowledge confusion module and a knowledge confusion detection module;

[0006] The knowledge point creation module uses intelligent word segmentation technology to identify the extracted user voice information, user picture information and user text information according to the extracted user problems, and then combines the user problems and the agent reply information into question and answer knowledge points according to the agent reply information, and expands the question and answer knowledge points to form entry knowledge points;

[0007] The text preprocessing module performs text preprocessing on the entry knowledge points, removes the stop words in the entry knowledge points to enhance the learning efficiency of text classification, and adds them to the intelligent customer service knowledge base.

[0008] The keyword extraction module screens the user text information, uses a TF-IDF algorithm to identify and extract special keywords in the text information, and checks the keywords of each knowledge point in the system using the extracted special keywords to determine whether a new knowledge point needs to be added to the intelligent customer service knowledge base;

[0009] The unstructured data translation module uses a CNN+RNN+CTC model to identify the text information and text range in the user picture information, convert the text area into character information, and output the text information.

[0010] The knowledge confusion module determines a confusion knowledge point candidate set according to the keyword overlap between multiple knowledge points in the intelligent customer service knowledge base, finds confusion knowledge points and high-quality confusion knowledge points.

[0011] The knowledge confusion detection module is used to perform confusion detection on the user text information, the output text information, the confusion knowledge points and the high-quality confusion knowledge points, and the existing intelligent customer service knowledge base, compares whether there are similar problems in the existing knowledge base, compares whether there are similar problems in the existing knowledge base, determines whether the knowledge point needs to be expanded, if there are no similar problems, the knowledge point is expanded, if there are similar problems, the problem is incorporated into the expansion problem of the existing knowledge point, and the dynamic update of the knowledge base is completed.

[0012] Further, the intelligent customer service knowledge base includes a knowledge base platform and a knowledge base. The knowledge base platform provides hardware devices and software environments for the knowledge base. The hardware devices include computing devices, storage devices, security facilities, and network devices. The software environment includes Python and Matlab that can be used for machine learning.

[0013] The knowledge base includes knowledge base data and keyword data. The knowledge base data is a collection of common problems and hot issues of each system and each business, including various types of business data stored in categories. The business data includes text data, picture data, voice data, semi-structured and unstructured data.

[0014] The keyword data is a keyword collection of common problems and hot issues of each system and each business, including special professional terms of each business data.

[0015] Further, the TFIDF method is used to determine the keywords of each knowledge point in the system.

[0016] Further, the steps of finding confusion knowledge points include:

[0017] S41: Determine the keywords corresponding to each knowledge point, match using the keywords, and obtain a candidate set of confused knowledge points composed of at least two knowledge points with a keyword coincidence degree higher than a preset coincidence degree;

[0018] S42: The candidate set of confused knowledge points includes at least a first knowledge point and a second knowledge point, and the semantic similarity between the first knowledge point and the second knowledge point is calculated, including calculating the pairwise semantic similarity between the multiple questions of the first knowledge point and the multiple questions of the second knowledge point, and determining the semantic similarity between the first knowledge point and the second knowledge point according to the pairwise semantic similarity;

[0019] If the semantic similarity is higher than a preset similarity, the first knowledge point and the second knowledge point are determined to be confused knowledge points.

[0020] Further, the steps of finding high-quality confused knowledge points include:

[0021] S51: Determine the keywords corresponding to each knowledge point, match using the keywords, and obtain a candidate set of confused knowledge points composed of at least two knowledge points with a keyword coincidence degree higher than a preset coincidence degree;

[0022] S52: The candidate set of confused knowledge points includes at least a first knowledge point and a second knowledge point, and the semantic similarity between the first knowledge point and the second knowledge point is calculated;

[0023] S53: According to the order of semantic similarity from large to small, in the multiple question pairs composed of the questions corresponding to the first knowledge point and the questions corresponding to the second knowledge point, the keyword coincidence degree of the first M question pairs is determined, and the extended coincidence degree between the first knowledge point and the second knowledge point is determined according to the keyword coincidence degree of the first M question pairs, where M is an integer greater than 0;

[0024] If the extended coincidence degree is greater than a preset extended coincidence degree, the confused knowledge points composed of the first knowledge point and the second knowledge point are determined to be high-quality confused knowledge points.

[0025] Further, the steps of the knowledge confusion detection module for confusion detection include:

[0026] S61: The user text information, output text information is used as input text information, and whether there is a domain keyword in the input text information is matched by checking the set keyword library. If there is a domain keyword, the keyword is screened out alone, and the input text information is cut into several words by a word segmentation tool to form an initial word.

[0027] S62: After segmenting the input text information into several words, the text information is represented in the form of vectors to form word vectors for numerical calculation in the model. The position of the word vector in the matrix is ​​determined to obtain the vector representation of the word. The vector representation of the keyword is obtained to form a vector matrix. The set of vector representations forms a vector matrix. Each row of the word vector matrix is ​​the vector corresponding to a certain word.

[0028] S63: A convolutional neural network model is used for knowledge confusion training. The convolutional neural network model includes a TextCNN convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer uses a three-channel input and uses word vector tools such as word2vec and fastText to perform unsupervised learning on the existing intelligent customer service knowledge base to obtain specific word vectors of the words, which are used as the input of the convolutional layer of the convolutional neural network model. The convolutional output after feature extraction is used as the input of the pooling layer. The average pooling method is used to obtain the output of the pooling layer. A fully connected layer is constructed based on the output of the pooling layer, and then softmax is applied to obtain the knowledge confusion detection result.

[0029] Compared with the prior art, the present invention has the following technical effects:

[0030] (1) The knowledge base established by this invention can be automatically expanded by user questions, automatically complete the automatic confusion detection of dynamic knowledge base, and form dynamic knowledge base;

[0031] (2) The knowledge base established by this invention can identify various unstructured data questions from users, improve user experience, and increase the accuracy of manual statistics on user questions. Attached Figure Description

[0032] Figure 1 This is a structural block diagram of the present invention; Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the invention, any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art to all other embodiments obtained without creative effort should be included within the protection scope of the present invention.

[0034] like Figure 1 As shown, a dynamically updated intelligent customer service knowledge base system includes an intelligent customer service knowledge base, a knowledge point creation module, a text preprocessing module, a keyword extraction module, an unstructured data translation module, a knowledge confusion module, and a knowledge confusion detection module.

[0035] The knowledge point creation module identifies the user voice information, user picture information and user text information according to the extracted user question by using intelligent word segmentation technology, and merges the user question and the agent reply information into a question and answer knowledge point according to the agent reply information, and forms an entry knowledge point by expanding the question of the question and answer knowledge point;

[0036] The text preprocessing module performs text preprocessing on the entry knowledge point, removes stop words in the entry knowledge point to enhance the learning efficiency of text classification, and adds the entry knowledge point to the intelligent customer service knowledge base;

[0037] The keyword extraction module filters the user text information, identifies and extracts special keywords in the text information by using a TF-IDF algorithm, and checks the keywords of each knowledge point in the system by using the extracted special keywords to determine whether a new knowledge point needs to be added to the intelligent customer service knowledge base;

[0038] The unstructured data translation module identifies the text information and text range in the user picture information by using a CNN+RNN+CTC model, converts the text area into character information, and outputs the text information;

[0039] The knowledge confusion module determines a confusion knowledge point candidate set according to the keyword overlap degree between multiple knowledge points in the intelligent customer service knowledge base, finds confusion knowledge points and high-quality confusion knowledge points;

[0040] The knowledge confusion detection module is used for confusion detection of the user text information, the output text information, the confusion knowledge points and the high-quality confusion knowledge points with the existing intelligent customer service knowledge base, comparison of whether there is a similar question in the existing knowledge base, comparison of whether there is a similar question in the existing knowledge base, judgment of whether knowledge point expansion is needed, if there is no similar question, the knowledge point expansion is performed, if there is a similar question, the question is incorporated into the expansion question of the existing knowledge point, and the dynamic update of the knowledge base is completed.

[0041] The intelligent customer service knowledge base includes a knowledge base platform and a knowledge base. The knowledge base platform provides hardware devices and software environments for the knowledge base. The hardware devices include computing devices, storage devices, security facilities and network devices. The software environments include Python and Matlab which can be used for machine learning.

[0042] The knowledge base includes knowledge base data and keyword data. The knowledge base data is a collection of common questions and hot questions of each system and each business, including classified storage of various business data. The business data includes text data, picture data, voice data, semi-structured and unstructured data.

[0043] The keyword data is a keyword set of common questions and hot issues of each system and each business, including specific professional terms of each business data.

[0044] The keyword of each knowledge point in the system is determined by using the term frequency inverse document frequency (TFIDF) method.

[0045] The step of finding the confused knowledge point includes:

[0046] S41: Determine the keywords corresponding to each knowledge point, use the keywords for matching, and obtain a confused knowledge point candidate set composed of at least two knowledge points with a keyword coincidence degree higher than a preset coincidence degree;

[0047] S42: The confused knowledge point candidate set at least includes a first knowledge point and a second knowledge point, and the semantic similarity between the first knowledge point and the second knowledge point is calculated, including calculating the pairwise semantic similarity between the multiple questions of the first knowledge point and the multiple questions of the second knowledge point, and determining the semantic similarity between the first knowledge point and the second knowledge point according to the pairwise semantic similarity;

[0048] If the semantic similarity is higher than a preset similarity, the first knowledge point and the second knowledge point are determined as confused knowledge points.

[0049] The step of finding the high-quality confused knowledge point includes:

[0050] S51: Determine the keywords corresponding to each knowledge point, use the keywords for matching, and obtain a confused knowledge point candidate set composed of at least two knowledge points with a keyword coincidence degree higher than a preset coincidence degree;

[0051] S52: The confused knowledge point candidate set at least includes a first knowledge point and a second knowledge point, and the semantic similarity between the first knowledge point and the second knowledge point is calculated;

[0052] S53: According to the order from large to small of the semantic similarity, in a plurality of question pairs composed of a plurality of questions corresponding to the first knowledge point and a plurality of questions corresponding to the second knowledge point, the keyword coincidence degree of the first M question pairs is determined, and the extended coincidence degree between the first knowledge point and the second knowledge point is determined according to the keyword coincidence degree of the first M question pairs, wherein M is an integer greater than 0;

[0053] If the extended coincidence degree is greater than a preset extended coincidence degree, the confused knowledge point composed of the first knowledge point and the second knowledge point is determined as a high-quality confused knowledge point.

[0054] The step of the knowledge confusion detection module for confusion detection includes:

[0055] S61: output the user text information as input text information, check the set keyword library, match whether there is a domain keyword in the input text information, if there is a domain keyword, separate the keyword, cut the input text information into several words through a word segmentation tool, and form an initial word;

[0056] S62: after cutting the input text information into several words, the text information is represented in the form of a vector to constitute a word vector for numerical calculation in the model, the position of the word vector in the matrix is determined to obtain the vector representation of the word, the vector representation of the keyword is obtained, and a vector matrix is constituted; the set of vector representations forms a vector matrix, and each row of the word vector matrix is a vector corresponding to a word;

[0057] S63: knowledge confusion training is performed by using a convolutional neural network model, the convolutional neural network model includes a TextCNN convolutional layer, a pooling layer and a full connection layer, the convolutional layer uses three channel inputs, word2vec, fastText and other word vector tools are used to learn the existing wisdom customer service knowledge base unsupervisedly, obtain specific word vectors of words, use the specific word vectors as inputs of the convolutional layer of the convolutional neural network model, use convolution outputs after feature extraction as inputs of the pooling layer, use a mean pooling method to obtain a pooling layer output result, construct a full connection layer according to the output of the pooling layer, and then obtain a knowledge confusion detection result through softmax.

[0058] By using the method, the knowledge base constructed can be dynamically updated, and knowledge points with many questions raised by users can form knowledge entries, the wisdom customer service system can improve the question and answer accuracy through the dynamically updated knowledge base, and in addition, the user experience can be enhanced by identifying questions of non-structured data such as voice and pictures.

Claims

1. A dynamically updated intelligent customer service knowledge base system, characterized in that, The system includes a smart customer service knowledge base, a knowledge point creation module, a text preprocessing module, a keyword extraction module, an unstructured data translation module, a knowledge obfuscation module, and a knowledge obfuscation detection module. The knowledge point creation module uses intelligent word segmentation technology to identify and extract user voice information, user image information, and user text information based on the extracted user questions. Then, based on the agent's reply information, it merges the user questions and agent's reply information into question-and-answer knowledge points, and expands the question-and-answer knowledge points to form item knowledge points. The text preprocessing module performs text preprocessing on the entry knowledge points, removes stop words from the entry knowledge points, and adds them to the intelligent customer service knowledge base. The keyword extraction module filters the user text information, uses the TF-IDF algorithm to identify and extract unique keywords from the text information, and uses the extracted unique keywords to check the keywords of each knowledge point in the system. The unstructured data translation module uses a CNN+RNN+CTC model to identify text regions in the user image information, convert the text regions into character information, and output text information. The knowledge obfuscation module determines a candidate set of obfuscated knowledge points based on the keyword overlap between multiple knowledge points in the intelligent customer service knowledge base, and searches for obfuscated knowledge points and high-quality obfuscated knowledge points. The knowledge obfuscation detection module is used to perform obfuscation detection on user text information, output text information, obfuscated knowledge points and high-quality obfuscated knowledge points with the existing intelligent customer service knowledge base, compare whether similar problems exist in the existing knowledge base, and determine whether knowledge points need to be expanded based on the comparison results, so as to complete the dynamic update of the knowledge base.

2. The dynamically updated intelligent customer service knowledge base system according to claim 1, characterized in that, The intelligent customer service knowledge base includes a knowledge base platform and a knowledge base. The knowledge base platform provides hardware devices and a software environment for the knowledge base. The hardware devices include computing devices, storage devices, security facilities, and network devices. The software environment includes Python and Matlab, which can be used for machine learning. The knowledge base includes knowledge base data and keyword data; the knowledge base data is a collection of common and hot issues of various systems and businesses, including various types of business data stored in categories, including text data, image data, voice data, semi-structured and unstructured data; The keyword data is a collection of keywords related to common and hot issues in various systems and business operations, including unique professional terms for each business data.

3. The dynamically updated intelligent customer service knowledge base system according to claim 2, characterized in that, The keyword of each knowledge point in the system is determined by using the inverse knowledge point frequency (TFIDF) method.

4. The dynamically updated intelligent customer service knowledge base system according to claim 3, characterized in that, The steps to identify confusing points of knowledge include: S41: Determine the keywords corresponding to each knowledge point, use the keywords to match, and obtain a candidate set of confusing knowledge points consisting of at least two knowledge points whose keyword overlap is higher than the preset overlap. S42: The candidate set of confusing knowledge points includes at least a first knowledge point and a second knowledge point. Calculating the semantic similarity between the first knowledge point and the second knowledge point includes calculating the pairwise semantic similarity between multiple questions of the first knowledge point and multiple questions of the second knowledge point, and determining the semantic similarity between the first knowledge point and the second knowledge point based on the pairwise semantic similarity. If the semantic similarity is higher than the preset similarity, then the first knowledge point and the second knowledge point are determined to be confused knowledge points.

5. The dynamically updated intelligent customer service knowledge base system according to claim 4, characterized in that, The steps to find high-quality, confusing knowledge points include: S51: Determine the keywords corresponding to each knowledge point, use the keywords to match, and obtain a candidate set of confusing knowledge points consisting of at least two knowledge points whose keyword overlap is higher than the preset overlap. S52: The candidate set of confused knowledge points includes at least a first knowledge point and a second knowledge point, and the semantic similarity between the first knowledge point and the second knowledge point is calculated; S53: In order of semantic similarity from largest to smallest, determine the keyword overlap of the first M question pairs formed by the questions corresponding to the first knowledge point and the questions corresponding to the second knowledge point, and determine the extended overlap between the first knowledge point and the second knowledge point based on the keyword overlap of the first M question pairs, where M is an integer greater than 0. If the expansion overlap is greater than the preset expansion overlap, then the confused knowledge point composed of the first knowledge point and the second knowledge point is determined to be a high-quality confused knowledge point.

6. The dynamically updated intelligent customer service knowledge base system according to claim 5, characterized in that, The steps for the knowledge obfuscation detection module to perform obfuscation detection include: S61: The user text information and output text information are used as input text information. By checking the set keyword library, the domain keywords are matched to see if there are any domain keywords in the input text information. If there are domain keywords, the keywords are filtered out separately and the input text information is segmented into several words by a word segmentation tool to form initial words. S62: After the input text information is segmented into several words, the text information is represented in the form of vectors to form word vectors for numerical calculation in the model. The position of the word vector in the matrix is ​​determined to obtain the vector representation of the word. The vector representation of the keyword is obtained to form a vector matrix. S63: A convolutional neural network model is used for knowledge confusion training. The convolutional neural network model includes a TextCNN convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer uses a three-channel input and uses word2vec and fastText word vector tools to perform unsupervised learning on the existing intelligent customer service knowledge base to obtain word vectors as input to the convolutional layer. After passing through the pooling layer and the fully connected layer, the knowledge confusion detection result is obtained.

Images