TW202133027A - Dialogue system and method for human-machine cooperation - Google Patents

Abstract

A dialogue system and a dialogue method for human-machine cooperation are provided. The system and the method receive dialog text input by a user and predict the possibility of crisis and the intention and the emotion of the user based on the dialog text. The system and the method automatically determine whether to transfer the dialog to a customer service representative to reply to the user or automatically reply to the user according to the possibility of crisis and the intention and the emotion of the user.

Description

Human-machine cooperation dialogue system and method

The present invention relates to natural language analysis, and in particular relates to a human-machine collaborative dialogue system and method.

At present, many product and service providers need a customer service system to answer users' questions or assist in solving users' problems. The use of dialogue robots can greatly reduce the labor and cost expenditure of text customer service, and the dialogue robot has the characteristics of responding to users instantly, serving many users at the same time, and serving 24 hours a day, all year round. However, when completely relying on the dialogue robot to answer, there will often be questions that are not answered. This will increase the user's confusion and reduce the user's willingness to use. Therefore, how to improve the service quality of dialogue robots is one of the issues that the industry pays attention to.

According to past research and most products on the market, most traditional technologies use conversational text customer service interfaces to provide the following two ways to switch robots and transfer real person services: 1. Set up a special button on the interface for users to click on their own , Trigger the transfer to the live service; 2. Provide the transfer to the live service link during the conversation, and ask the user whether to transfer to the live service. Among them, method 1 is likely to cause the user to directly click the transfer to the live service as soon as the user enters the system, and the dialogue robot is useless, and the method 2 is mostly to set the keyword detection method during the dialogue (for example, "I want to transfer" appears during the dialogue. service personnel" Wait for the set sentence) or set the confidence level to be lower than a certain preset threshold, then provide a link to transfer real customer service personnel. However, the above-mentioned methods will cause the users to have a poor experience with the text customer service robot, and they believe that the text customer service robot cannot handle related problems, thereby reducing their willingness to use.

Therefore, it is the goal of the current intelligent text customer service robot system or dialogue robot to achieve a good response by matching real people and robots, and at a certain level of user satisfaction, seeking to minimize the investment cost of real human customer service.

In order to solve the above problems, the present invention provides a human-machine collaborative dialogue system, including a text pre-processing module, a crisis recognition module, an intention classification module, an emotion detection module, a transfer recognition module, a smart auxiliary reply module, and Robot reply module. The word pre-processing module is used to receive the dialogue text input by the user, extract the dialogue vocabulary in the dialogue text, and convert the dialogue vocabulary into a dialogue vector. The crisis recognition module is used to output the probability of crisis confidence value according to the dialogue vocabulary. The intention classification module is used to output the probability distribution of intention classification according to the dialogue vector. The emotion detection module is used to output the probability distribution of the emotion contour according to the dialogue vector. The transfer recognition module is used to output the transfer customer service probability according to the probability of crisis confidence value, the probability distribution of intention classification and the probability distribution of emotional profile. The smart auxiliary reply module is used to reply to the user with the first reply sentence entered by the customer service staff when the probability of transferring the customer service is greater than or equal to the threshold value. The robot reply module is used to generate a second reply sentence according to the distribution of the intention classification probability when the probability of transferring the customer service is less than the threshold value, and reply the user with the second reply sentence.

The present invention also provides a human-machine collaborative dialogue method, including: receiving dialogue text input by a user, extracting dialogue vocabulary in the dialogue text to convert the dialogue vocabulary into a dialogue vector; outputting a crisis confidence value probability according to the dialogue vocabulary; and according to the dialogue vector Output intent classification probability distribution; according to dialogue vector Output the probability distribution of emotional profile; output the transfer customer service probability according to the crisis confidence value probability, the intention classification probability distribution and the emotional profile probability distribution; when the transfer customer service probability is greater than or equal to the threshold value, the user will be replied with the first reply sentence input by the customer service staff ; And, when the probability of transferring the customer service is less than the threshold value, a second reply sentence is generated according to the distribution of the intention classification probability, and the second reply sentence is used to reply to the user.

The human-machine collaborative dialogue system and method of the present invention have the following characteristics and effects: 1. Automatically identify the user's problem intention, instantly transfer real-person customer service or robot customer service, and hand over easy-to-reply questions to robots to reply, reducing the workload of customer service staff. 2. Combining crisis and emotion detection functions, it can provide smarter customer service with a warmer temperature compared to completely responding with a robot system and transferring the response from a real person according to the user's instructions. 3. Provide customer service personnel with a smart auxiliary interface to quickly and effectively grasp user problems, which can speed up the response speed and quality of customer service personnel.

10: Human-machine collaboration dialogue system

21~29: Process steps

100: text pre-processing module

110: Dialogue text

120: Sentence Regularizer

130: sentence breaker

140: Vocabulary Vectorizer

150: database

200: Crisis Recognition Module

220: Sensitive Word Database

230: Crisis Recognizer

240: Crisis confidence check-in rate

300: Intent Classification Module

320: Intent Classifier

330: Intent Classification Probability Distribution

400: Emotion detection module

420: emotional vocabulary

430: Mood Detector

440: Probability Distribution of Emotional Contours

500: Transfer identification module

540: Transfer identifier

550: Probability of transfer to customer service

600: Context Summary Module

610: database

620: Context Summarizer

630: summary sentence

700: Smart Assistant Reply Module

705: Smart Assistant Interface

710: Context dialogue summary information screen

720: Robot auxiliary sentence screen

730: Reply sentence input interface

740: Sensitive word display screen

750: User mood indicator light

800: Robot Reply Module

830: Q&A Knowledge Base

840: Reply Generator

850: reply sentence

Figure 1 is a block diagram of a human-machine collaborative dialogue system according to an embodiment of the present invention.

Figure 2 is a flowchart of a method for human-machine collaborative dialogue according to an embodiment of the present invention.

Figure 3 is a schematic diagram of the word pre-processing module in Figure 1.

Figure 4 is a schematic diagram of the crisis identification module in Figure 1.

Figure 5 is a schematic diagram of the intention classification module in Figure 1.

Figure 6 is a schematic diagram of the emotion detection module in Figure 1.

Figure 7 is a schematic diagram of the transfer identification module in Figure 1.

Figure 8 is a schematic diagram of the context summary module in Figure 1.

Figure 9 is a schematic diagram of the smart assistant interface of the smart assistant reply module in figure 1.

Figure 10 is a schematic diagram of the robot reply module in Figure 1.

The following specific examples illustrate the implementation of the present invention. Those familiar with the art can easily understand the other advantages and effects of the present invention from the contents disclosed in this specification.

FIG. 1 is a block diagram of a human-machine collaborative dialogue system 10 according to an embodiment of the present invention. The human-machine collaborative dialogue system 10 includes a text pre-processing module 100, a crisis recognition module 200, an intention classification module 300, an emotion detection module 400, a transfer recognition module 500, a context summary module 600, and a smart auxiliary reply module. Group 700, and robot reply module 800. Among them, the text pre-processing module 100 is coupled to the crisis recognition module 200, the intention classification module 300, and the emotion detection module 400. The transfer recognition module 500 is coupled to the crisis recognition module 200, the intention classification module 300, the emotion detection module 400, the context summary module 600, the smart assistant reply module 700, and the robot reply module 800. The context summary module 600 is coupled to the transfer recognition module 500, the smart assistant reply module 700, and the robot reply module 800.

In one embodiment, each of the above-mentioned modules is hardware, such as a server or other electronic devices with data processing capabilities or program execution capabilities. In another embodiment, each of the aforementioned modules is software or firmware, such as a computer program, which can be executed by a single electronic device or distributed by multiple electronic devices. In another embodiment, some of the above-mentioned modules are hardware, and the rest are software or firmware.

In one embodiment, the sensitive vocabulary 220, the emotional vocabulary 420, and the question-and-answer knowledge base 830 are all databases, and these databases may be part of the human-machine collaborative dialogue system 10, or may be independent databases.

Please refer to Figure 1 and Figure 2 together. Figure 2 is a flowchart of the human-machine collaborative dialogue method executed by the human-machine collaborative dialogue system 10, wherein the flow of the human-machine collaborative dialogue method starts from step 21.

In step 21, the text pre-processing module 100 receives the dialogue text 110 input by the user, and extracts the dialogue vocabulary in the dialogue text 110 to convert the dialogue vocabulary into a dialogue vector. Please also refer to FIG. 3, the word preprocessing module 100 includes a sentence normalizer 120, a sentence breaker 130, and a vocabulary vectorizer 140. Like the modules shown in FIG. 1, the sentence normalizer 120, sentence breaker 130, and vocabulary vectorizer 140 can be hardware, firmware, or software.

The sentence normalizer 120 is used to filter out the words in the dialogue text 110 other than the preset language, and filter out the less important redundant symbols in the dialogue text 110. For example, in one embodiment, if the default languages of the human-machine collaborative dialogue system 10 are Chinese and English, the sentence normalizer 120 will filter out the dialogue text 110 except for Chinese and English, and only retain Chinese and English. The sentence normalizer 120 can also convert the encoding of the dialogue text 110, for example, between traditional Chinese and simplified Chinese. For example, when the dialogue text 110 entered by the user is in simplified Chinese "My bill is deducted wrong, it’s really bad @#$@", after filtering and encoding conversion by the sentence normalizer 120, the dialogue text 110 is converted to Traditional Chinese "My bill was deducted wrongly, it sucks."

The sentence breaker 130 is configured to receive the dialogue text 110 output by the sentence normalizer 120, and then divide the dialogue text 110 in units of words, and remove the words listed in the stop vocabulary in the dialogue text 110. The words in the stop vocabulary are all redundant words, so the sentence breaker 130 can remove the dialogue text 110 Superfluous words in. For example, the dialogue text received by the sentence word breaker 130 is "My bill is deducted wrongly, really bad", and the separated dialogue text 110 is "My. , And suppose that the words "I", "的" and "了" appear in the stop vocabulary, so the result after removing the redundant words is "Bill. Deduction wrong. Payment. Really bad". The aforementioned dialogue vocabulary is the dialogue text 110 processed by the sentence normalizer 120 and the sentence breaker 130. If the previous example is continued, the dialogue vocabulary is "Bill. Wrong deduction. Payment. Really bad".

The vocabulary vectorizer 140 is used to receive the dialogue vocabulary output by the sentence breaker 130, and then convert the dialogue vocabulary into a dialogue vector according to the dictionary. For example, if the One-Hot vector representation is adopted, the length of the vector of each vocabulary is the number of vocabulary contained in the dictionary, and each dimension represents a vocabulary in the dictionary. The One-Hot vector of each vocabulary is only if its only representative dimension is 1, and the other dimensions are 0. For example, according to a certain preset dictionary, the One-Hot vector of the "bill" is represented as [1,0,0, 0,0,0,0], the One-Hot vector of "deduction" is represented as [0,1,0,0,0,0,0], and the One-Hot vector of "section" is represented as [0,0 ,1,0,0,0,0], the "really bad" One-Hot vector is represented as [0,0,0,1,0,0,0]. The above-mentioned dialogue vector is a set of vectors for the conversion of each word in the dialogue vocabulary. The final output of the word preprocessing module 100 is the above-mentioned dialogue vocabulary and dialogue vector. In one embodiment, the text pre-processing module 100 stores the dialogue vocabulary and dialogue vectors in a database 150 for use by other modules. The database 150 may be a part of the human-machine collaborative dialogue system 10.

Next, in step 22, the crisis identification module 200 outputs the crisis confidence value probability according to the dialogue vocabulary in the database 150. Please also refer to FIG. 4, the crisis recognition module 200 includes a crisis recognizer 230. The sensitive vocabulary 220 stores the pre-collected sensitive vocabulary. The crisis recognizer 230 is an artificial intelligence model pre-trained using sensitive words in the sensitive vocabulary 220 for crisis recognition. Using the crisis recognizer 230, the crisis recognition module 200 can output a crisis letter according to whether the dialogue vocabulary contains the sensitive vocabulary in the sensitive vocabulary 220 and whether the prefix vocabulary of the sensitive vocabulary contained in the dialogue vocabulary contains a negative vocabulary. The heart rate is 240. For example, when the dialogue vocabulary includes any vocabulary in the sensitive vocabulary 220 (such as "complaint", "bad attitude" or "really bad", etc.), and its prefix vocabulary does not include negative vocabulary, then the crisis recognizer 230 Identify as a crisis. If the dialogue vocabulary does not contain sensitive words, or contains sensitive words but its prefix vocabulary contains negative words, the crisis recognizer 230 recognizes that there is no crisis. For example, when the dialogue text input by the user is “customer service attitude will not be bad”, it will not be recognized as a crisis due to the negative word “no”, and the crisis recognizer 230 will output a lower crisis confidence value probability 240. When the dialogue vocabulary is "Bill. Wrong deduction. Payment. Really bad", it will be identified as a crisis, and the crisis recognizer 230 will output a higher probability of crisis confidence value 240, such as [0.88].

Next, in step 23, the intention classification module 300 outputs an intention classification probability distribution according to the dialogue vector in the database 150. Please also refer to FIG. 5, the intent classification module 300 includes an intent classifier 320. The intent classifier 320 is an artificial intelligence model pre-trained using pre-collected intent annotation data, and is used to classify the user's intent. The intention classifier 320 outputs an intention classification probability distribution 330 through a softmax function (or called a normalized exponential function) according to the above-mentioned dialogue vector. The intention classification probability distribution 330 includes probability distributions in a plurality of preset intention categories. For example, assuming there are three preset intention categories, namely "bill problem", "mobile phone problem" and "communication problem", the probability distribution of intent classification 330 can be a three-dimensional vector, such as [0.8,0.15,0.05], Among them, "bill problems" have the highest probability. In the intention classification probability distribution 330, the sum of the probability values of each dimension is 1.

Next, in step 24, the emotion detection module 400 outputs an emotion profile probability distribution according to the dialogue vector in the database 150. Please also refer to FIG. 6, the emotion detection module 400 includes an emotion detector 430. The emotion detector 430 is an artificial intelligence model pre-trained using pre-collected emotion annotation data to detect the user's emotions. In addition, the emotion vocabulary pre-collected and stored in the emotion vocabulary 420 can be used to enhance the detection capability of the emotion detector 430. The emotion detector 430 based on the above dialogue Vector, and through the softmax function, the emotional contour probability distribution 440 is output. The emotion profile probability distribution 440 includes probability distributions in a plurality of preset emotion indicators. For example, suppose there are six preset emotion indicators: [happy, scared, surprised, angry, sad, nauseous], the probability distribution 440 of the emotional contour is a 6-dimensional vector, for example [0.01,0.1,0.03,0.8,0.02,0.04 ], the probability of being "angry" is the highest. The probability distribution of emotion profile 440 in which the sum of the probability values of each dimension is 1.

Next, in step 25, the transfer recognition module 500 outputs the transfer customer service probability according to the crisis confidence value probability 240, the intention classification probability distribution 330, and the emotional profile probability distribution 440. Please also refer to FIG. 7, the transfer identification module 500 includes a transfer identifier 540. The transfer recognizer 540 is an artificial intelligence model pre-trained using pre-collected data. The transfer recognition module 500 concatenates the crisis confidence value probability 240 generated by the crisis recognition module 200, the intention classification probability distribution 330 generated by the intention classification module 300, and the emotion profile probability distribution 440 generated by the emotion detection module 400 into A long vector is input to the switch recognizer 540. If the previous example is continued, the probability of crisis confidence value 240 is a 1-dimensional vector [0.88], the probability distribution of intention classification 330 is a 3-dimensional vector [0.8, 0.15, 0.05], and the probability distribution of emotion profile 440 is a 6-dimensional vector [0.01, 0.1,0.03,0.8,0.02,0.04], a 10-dimensional vector composed of the three in series [0.88,0.8,0.15,0.05,0.01,0.1,0.03,0.8,0.02,0.04] is the input of the transfer identifier 540 . Then the transfer identifier 540 outputs the transfer customer service probability 550, for example, its value is [0.68].

Next, in step 26, the context summary module 600 extracts summary sentences in all conversation content with the user. The summary sentence only contains important content, for example, greetings will be eliminated. All the above-mentioned dialogue content also includes dialogue text 110. The dialogue text 110 is usually the most recent dialogue among all dialogue contents. Please also refer to FIG. 8, the context summary module 600 records all the content of the conversation during the conversation with the user, for example, stores all the content of the conversation in the database 610. The database 610 may be a part of the human-machine collaborative dialogue system 10. The context summary module 600 includes context Summarizer 620. The context summarizer 620 is a pre-trained artificial intelligence model. For example, a sequence to sequence training method of deep learning may be used. During training, a large number of long sentences and corresponding short summary sentences need to be collected, and the weights are revised repeatedly through the gradient descent method to obtain the context summarizer 620, which can automatically convert all the lengthy dialogue content in the database 610 into short summary sentences 630 . For example, if the user has input "Hello, I have a billing question to ask" before inputting "The wrong deduction is bad", then the summary sentence 630 output by the context digester 620 may be "Bill deduction is wrong". The summary sentence 630 may support the intelligent auxiliary reply module 700 to provide customer service personnel for reference, and may also be provided to the robot reply module 800 to automatically generate a reply sentence.

Next, in step 27, it is checked whether the transfer customer service probability 550 output by the transfer identification module 500 is greater than or equal to a preset threshold. Step 27 can be performed by the transfer recognition module 500, or can also be performed by the smart assistant reply module 700 and the robot reply module 800. If the transfer customer service probability 550 is greater than or equal to the threshold, the smart assistant reply module 700 responds to the user with the reply sentence input by the customer service staff in step 28; otherwise, the robot reply module 800 automatically generates a reply sentence in step 29 to reply to the user. For example, suppose that the transfer customer service probability 550 is [0.68] and the threshold value is 0.5, then the process proceeds to step 28, and the customer service personnel and the user dialogue are transferred. For example, assuming that the probability of transferring customer service 550 is [0.68] and the threshold value is 0.7, the process proceeds to step 29, and the robot reply module 800 automatically talks to the user.

Please refer to Fig. 9, in order to assist the customer service personnel, the smart assistance reply module 700 can display the smart assistance interface 705 on a display screen of the human-machine collaborative dialogue system 10 for the customer service personnel to use. The smart assistant interface 705 includes a contextual dialogue summary information screen 710, a robot assistant sentence screen 720, a reply sentence input interface 730, a sensitive word display screen 740, and a user emotion display light 750 for displaying auxiliary information and receiving reply sentences input by customer service personnel . Context dialogue summary information screen 710 can display the summary sentence 630 provided by the context summary module 600, the robot auxiliary sentence screen 720 can display the reply sentence 850 automatically generated by the robot reply module 800 (details will be described later), and the reply sentence input interface 730 can receive the reply sentence input by the customer service staff , The sensitive word display screen 740 can display the sensitive words in the dialogue text 110, and the user emotion display light 750 can display the user emotion according to the emotion profile probability distribution 440. In another embodiment, the user emotion display light 750 may be changed to directly display the emotion profile probability distribution 440. In another embodiment, in addition to displaying the summary sentence 630, the context dialogue summary information screen 710 can also display all the dialogue contents in the database 610.

Continuing the previous example, if the transfer identification module 500 generates a transfer customer service probability 550 of [0.68], and the default threshold of the transfer identification module 500 is 0.5, then the transfer customer service probability 550 is greater than the threshold. In this embodiment, the smart auxiliary reply module 700 is switched. In order to assist the customer service staff, the summary sentence 630 displayed on the context dialogue summary information screen 710 is "Bill deduction", and the reply sentence 850 automatically generated by the robot reply module 800 displayed on the robot assistant sentence screen 720 is "Please log in to your personal account to check Relevant withholding information and contact the customer service staff", the sensitive word display screen 740 displays the sensitive word "really bad", and the user emotion display light 750 shows that the current user's emotional profile is "angry" with the highest indicator. The customer service staff uses the above auxiliary information to make a judgment, and enters the reply sentence "I'm sorry for the trouble, can you confirm your personal information?" in the reply sentence input interface 730. Then, after the customer service personnel directly conduct identity verification and other confirmation procedures with the user, they can query the user's billing information in the human-machine collaboration dialogue system 10 and reply to the user "Is the wrong deduction for X month X day OO?". In this way, you can quickly respond to related issues and directly confirm the process of related issues with users.

All the dialogue content, summary sentences, automatic reply sentences, sensitive vocabulary, and user emotions displayed on the smart assistant interface 705 can help the customer service staff grasp the current key points and respond to users quickly and correctly to improve service efficiency. Smart assistant reply module 700 will use customer service in step 28 The person responds to the user with the reply sentence entered in the reply sentence input interface 730. In addition to the reply sentences input by the customer service personnel, according to the operations or instructions issued by the customer service personnel on the smart assistant interface 705, the smart assistant reply module 700 can directly reply the user with the automatic reply sentence 850 in step 28, or use the customer service personnel The modified automatic reply sentence 850 replies to the user.

In another embodiment, the smart assistant interface 705 can be simplified. For example, part of the above-mentioned assistant information can be omitted, or the context dialogue summary information screen 710, the robot assistant sentence screen 720, the reply sentence input interface 730, and the sensitive word display screen can be omitted. At least one of 740 and the user emotion display light 750.

Please refer to FIG. 10, the robot reply module 800 includes a reply generator 840. The robot reply module 800 selects the intent classification with the highest probability according to the intent classification probability distribution 330, uses this intent classification to query the Q&A knowledge base 830 to generate candidate reply sentences, and then inputs the candidate reply sentences and summary sentences 630 into the reply generator 840 together. To generate a sentence 850 for the robot to automatically reply, and then use the reply sentence 850 to reply to the user in step 29. Among them, the Q&A knowledge base 830 includes user's common questions and corresponding best answers. The response generator 840 is an artificial intelligence model pre-trained using pre-collected data. The input of the response generator 840, in addition to the user's current intention, also includes a summary sentence 630 of all conversations with the user in the database 610, so that a more profound background can be considered and a more intelligent response can be provided.

Continuing the previous example, if the transfer identification module 500 generates a transfer customer service probability 550 of [0.68] and the default threshold of the transfer identification module 500 is 0.7, then the transfer customer service probability 550 is less than the threshold. In this embodiment, the robot reply module 800 is transferred. The robot reply module 800 judges that the user’s intention is "bill problem" based on the intention classification probability distribution 330 generated by the intention classification module 300, and receives the summary sentence 630 "Bill deduction wrong" provided by the context summary module 600. The above two By asking Answer The response sentence 850 generated by the knowledge base 830 and the response generator 840 is "Please log in to your personal account to query relevant withholding information and contact customer service personnel." In addition, the robot reply module 800 can display a web page link for inquiring an account or a prompt button for transferring a customer service staff, so that the user can select a subsequent processing procedure. To handle such billing issues, you need to go through the user’s personal data identity verification process. If the user chooses the web link for inquiring about the account, it will connect to the relevant customer service website for subsequent authentication and query actions, and if the user chooses to transfer the customer service staff to prompt the button , As in the aforementioned process of transferring the smart assistant reply module 700, the real customer service performs user identity verification and the real customer service responds.

In addition, the order of execution of some steps shown in Figure 2 is adjustable. For example, the order of execution among the three steps of steps 22, 23, and 24 can be adjusted arbitrarily, and these three steps can also be executed at the same time. The execution order of step 26 can also be adjusted, as long as it is executed before step 28 and step 29.

In another embodiment, the context summary module 600, step 26, and summary sentence 630 may be omitted. In some embodiments, the summary sentence 630 can be replaced with all the conversation content in the database 610.

As described above, the crisis recognizer 230, the intent classifier 320, the emotion detector 430, the transfer recognizer 540, the context summarizer 620, and the response generator 840 are all artificial intelligence models. Each of these artificial intelligence models can use individual or combination of pattern matching, information retrieval, rule analysis, statistical methods, machine learning, and deep learning to construct a model. The above-mentioned statistical method is, for example, regression analysis. The above-mentioned machine learning can use one or more models of support vector machine (SVM), neural network, and decision tree, for example. The above-mentioned deep learning can use recurrent neural network (recurrent neural network; RNN for short), long short-term memory model (long short-term memory), for example; LSTM for short), deep neural network (DNN for short), and convolutional neural network (CNN for short) one or more models.

The human-machine collaborative dialogue system and method of the present invention have the following characteristics and effects: 1. Automatically identify the user's problem intention, instantly transfer real-person customer service or robot customer service, and hand over easy-to-reply questions to robots to reply, reducing the workload of customer service staff. 2. Combining crisis and emotion detection functions, it can provide smarter customer service with a warmer temperature compared to completely responding with a robot system and transferring the response from a real person according to the user's instructions. 3. Provide customer service personnel with a smart auxiliary interface to quickly and effectively grasp user problems, which can speed up the response speed and quality of customer service personnel. In addition, the human-machine collaborative dialogue system and method of the present invention can be applied to any industry and product that uses a dialogue interface to interact with users.

The above-mentioned embodiments only exemplarily illustrate the principles and effects of the present invention, and are not intended to limit the present invention. Anyone who is familiar with this technique can modify and change the above-mentioned embodiments without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the rights of the present invention should be listed in the scope of patent application described later.

10: Human-machine collaboration dialogue system

100: text pre-processing module

110: Dialogue text

200: Crisis Recognition Module

220: Sensitive Word Database

300: Intent Classification Module

400: Emotion detection module

420: emotional vocabulary

500: Transfer identification module

600: Context Summary Module

700: Smart Assistant Reply Module

800: Robot Reply Module

830: Q&A Knowledge Base

Claims (10)

A human-machine collaborative dialogue system includes: The word pre-processing module is used to receive the dialogue text input by the user, extract the dialogue vocabulary in the dialogue text, and convert the dialogue vocabulary into a dialogue vector; The crisis recognition module is used to output the probability of crisis confidence value according to the dialogue vocabulary; The intention classification module is used to output the probability distribution of intention classification according to the dialogue vector; The emotion detection module is used to output the probability distribution of the emotion contour according to the dialogue vector; The transfer recognition module is used to output the transfer probability of customer service according to the probability of the crisis confidence value, the probability distribution of the intention classification, and the probability distribution of the emotional profile; The smart auxiliary reply module is used to reply to the user with the first reply sentence entered by the customer service staff when the probability of the customer service transfer is greater than or equal to the threshold; and The robot reply module is used to generate a second reply sentence according to the intention classification probability distribution when the probability of transferring customer service is less than the threshold value, and reply the user with the second reply sentence. For example, the human-computer collaboration dialogue system described in item 1 of the scope of patent application, the text pre-processing module includes: The sentence regularizer is used to filter out the text in the dialog text except the default language, and filter out the redundant symbols in the dialog text; The sentence word breaker is used to separate the dialogue text by vocabulary unit, and remove the words listed in the stop vocabulary in the dialogue text, where the dialogue vocabulary passes through the sentence normalizer and the sentence word breaker The processed text of the dialogue; and The vocabulary vectorizer is used to convert the dialogue vocabulary into the dialogue vector according to the dictionary. For example, in the human-machine collaborative dialogue system described in item 1 of the scope of patent application, the crisis identification module is based on whether the dialogue vocabulary contains sensitive vocabulary in the sensitive vocabulary and whether the prefix vocabulary of the sensitive vocabulary contains negative vocabulary, Output the confidence value probability of the crisis. For the human-machine collaborative dialogue system described in the first item of the patent application, the probability distribution of the intention classification is included in the probability distribution of a plurality of preset intention categories, and the probability distribution of the emotion profile is included in a plurality of preset emotion indicators Probability distribution. For example, the human-machine collaboration dialogue system described in item 1 of the scope of patent application includes: The context summary module is used for extracting summary sentences in all dialogue contents of the user, wherein the all dialogue contents include the dialogue text. For example, in the human-machine collaboration dialogue system described in item 5 of the scope of patent application, the smart auxiliary reply module is further used to display a smart auxiliary interface, and display auxiliary information in the smart auxiliary interface and receive the first reply sentence, wherein, The auxiliary information includes at least one of the summary sentence, the second reply sentence, the sensitive words in the dialogue text, and the probability distribution of the emotional contour. For example, in the human-machine collaborative dialogue system described in item 5 of the scope of patent application, the robot reply module is further used to generate the second reply sentence according to the intent classification probability distribution, the question and answer knowledge base, and the summary sentence. A human-machine collaborative dialogue method includes: Receiving the dialogue text input by the user, extracting the dialogue vocabulary in the dialogue text, so as to convert the dialogue vocabulary into a dialogue vector; According to the probability of outputting the crisis confidence value of the dialogue vocabulary, the probability distribution of outputting the intention classification according to the dialogue vector, and the probability distribution of outputting the emotional contour according to the dialogue vector; According to the probability of the crisis confidence value, the probability distribution of the intention classification, and the probability distribution of the emotional profile, the probability of switching to the customer service is output; When the probability of transferring customer service is greater than or equal to the threshold, reply to the user with the first reply sentence entered by the customer service staff; and When the probability of transferring customer service is less than the threshold value, a second reply sentence is generated according to the intention classification probability distribution, and the second reply sentence is used to reply to the user. For example, the human-machine collaborative dialogue method described in item 8 of the scope of patent application includes: Excerpts and summary sentences in all dialogue contents of the user, wherein all dialogue contents include the dialogue text; and Display a smart auxiliary interface to display auxiliary information in the smart auxiliary interface and receive the first reply sentence, wherein the auxiliary information includes the summary sentence, the second reply sentence, the sensitive words in the dialogue text, and the emotional outline Probability distribution at least one of them. For example, the human-machine collaborative dialogue method described in item 8 of the scope of patent application includes: Excerpts and summary sentences in all dialogue contents of the user, wherein all dialogue contents include the dialogue text; and The second reply sentence is generated according to the probability distribution of the intention classification, the Q&A knowledge base, and the summary sentence.

Images