TWI733012B - Dialogue system and method of integrating intentions and hot keys - Google Patents

Abstract

A dialogue system and a method of integrating intentions and hot keys are provided. In embodiments, an intention identification module generates features of intention, and an intention parameter identification module generates features of intention parameters. Then, according to the features of intention and the features of intention parameters, a dialog management module generates the status of conversation and the task of intention, and a systematic creation module for content of reply obtains, from the database storing content of reply, the systematic content of reply as the reply to the user, based on the status of conversation and the task of intention. Thereby, the effects of quick and effective reply could be achieved.

Description

Dialogue system and method combining intention and shortcut keys

The present invention relates to a dialogue technology, in particular to a dialogue system and method combining intention and shortcut keys.

In traditional dialogue systems, customer service personnel must be stationed online regularly to answer questions raised by system users at any time. However, such a service method will increase with the number of system users and require a large number of customer service personnel to effectively serve users. In addition, with such a manual response method, customer service personnel often need to spend time searching for the corresponding response content, so the effect of immediate response cannot be achieved in terms of time.

In view of the foregoing problems, it is really necessary to provide an effective dialogue system and method to achieve a quick and effective response effect, thereby solving the problems faced by users.

Based on the problems in the prior art, the present invention discloses a dialogue system and method combining intention and shortcut keys. Compared with the prior art, an embodiment of the present invention discloses the use of an intention recognition module, an intention parameter recognition module, a dialog management module, and a system response content creation module to generate a dialog reply. Content, in order to have real-time and effective dialogue with users.

An embodiment of the present invention provides a dialogue system that combines intentions and shortcut keys, including: an intention recognition module that converts a text pre-processing result into an intention feature; an intention parameter recognition module based on the text pre-processing The results and the intention features generate intention parameter features; a dialogue management module that generates dialog states and intention tasks based on the intention features and the intention parameter features and based on shortcut key information; and a system response content creation The module obtains the system reply content from a reply content database according to the dialogue state and the intention task.

In another embodiment, the intention features include at least one of the following: an intention attribute, an intention domain category, an intention domain, and an intention name.

In another embodiment, the intent parameter features include at least one of the following: an intent parameter category, an intent parameter name, and an intent parameter representative name.

In another embodiment, the shortcut key information includes a user's current dialog information and/or a user-specific shortcut key list.

In another embodiment, the intention task is a task based on the combination of the intention features and the intention parameter features.

In another embodiment, a text pre-processing module is further included, which processes a text input content to obtain the text pre-processing result.

In another embodiment, the text pre-processing module includes: a text sentence normalization module that filters and encodes specific symbols and language in the text input content to generate a normalized text content; The sentence segmentation module, which separates the normalized text content in word units, and then generates a word-separated vocabulary result; and a vocabulary vectorization module, which separates the normalized text content The result of the vocabulary after the word is converted into a vector representation as the result of the pre-processing of the word.

In another embodiment, a text input module is further included for a user to input the text input content.

Another embodiment of the present invention provides a dialogue method combining intentions and shortcut keys, including the following steps: converting a text pre-processing result into intention features; generating intention parameter features based on the text pre-processing results and the intention features; According to the intention features and the intention parameter features, and based on the shortcut key information, a dialog state and an intention task are generated; and based on the dialog status and the intention task, the system reply content is obtained from a reply content database.

In another embodiment, the intention features include at least one of the following: an intention attribute, an intention domain category, an intention domain, and an intention name.

In another embodiment, the intent parameter features include at least one of the following: an intent parameter category, an intent parameter name, and an intent parameter representative name.

In another embodiment, the shortcut key information includes a user's current dialog information and/or a user-specific shortcut key list.

In another embodiment, the intention task is a task based on the combination of the intention features and the intention parameter features.

In another embodiment, before converting a text pre-processing result into an intent feature, processing a text input content to obtain the text pre-processing result.

In another embodiment, the step of processing the text input content to obtain the text pre-processing result includes: inputting the text into the content Filter and code conversion of specific symbols and language to generate a normalized text content; separate the normalized text content in word units to generate a word-separated vocabulary result; and the word-separated vocabulary result Convert to vector notation as the result of pre-processing of the text.

It should be understood that the subject matter described above can be implemented as a computer-controlled device, computer program, computing system, or as a product, such as a computer-readable storage medium.

In order to make the above-mentioned features and advantages of the present invention more comprehensible, embodiments are specifically described below in conjunction with the accompanying drawings. In the following description, the additional features and advantages of the present invention will be partially explained, and these features and advantages will be partly obvious from the description, or can be learned by practicing the present invention. The features and advantages of the present invention are realized and achieved by means of the elements and combinations specifically pointed out in the scope of the patent application. It should be understood that both the foregoing general description and the following detailed description are only illustrative and explanatory, and are not intended to limit the claimed scope of the present invention.

100‧‧‧Starting dialogue module

110‧‧‧System start dialogue

120‧‧‧System default initial sentence content

130‧‧‧User-specific shortcut key list

200‧‧‧Text Input Module

300‧‧‧Text preprocessing module

400‧‧‧Intent Recognition Module

410‧‧‧Intent attribute recognition module

420‧‧‧Intent domain category recognition module

430‧‧‧Intent Domain Recognition Module

440‧‧‧Intent name recognition module

450‧‧‧Intention tree

451‧‧‧Intent attribute

452‧‧‧Intent domain category

453‧‧‧Field of Intent

454‧‧‧Intent name

460‧‧‧Training data

500‧‧‧Intent parameter recognition module

510‧‧‧Intent parameter recognition module

513‧‧‧Intent name

520‧‧‧Intent parameter represents name recognition module

530‧‧‧Intent parameter tree

531‧‧‧Required parameter category

532‧‧‧Optional parameter category

533‧‧‧Representative name of parameter 1

534‧‧‧The first parameter alias

535‧‧‧Representative name of parameter 2

536‧‧‧Second parameter alias

540‧‧‧Training data

600‧‧‧Dialog Management Module

610‧‧‧Dialog Management Module

620‧‧‧Dialog State Machine Module

630‧‧‧Intent attribute table

640‧‧‧Dialog State Table

650‧‧‧Current conversation information

660‧‧‧User-specific shortcut key list

700‧‧‧System response content creation module

710‧‧‧Continue dialogue module

720‧‧‧Response content database search module

730‧‧‧Transfer personnel processing module

800‧‧‧Reply and Improvement Module

810‧‧‧ Reply to User Module

811‧‧‧Auxiliary Staff Module

812‧‧‧Automatic Reply Module

820‧‧‧Response data correctness judgment module

830‧‧‧Dialogue system performance improvement module

831‧‧‧Response result marking module

900‧‧‧Shortcut key creation and update module

910‧‧‧User-specific shortcut keys to create and update modules

920‧‧‧User conversation record

930‧‧‧User-specific shortcut keys table

1000‧‧‧Dialog System

Figure 1 shows a schematic architecture diagram of a dialogue system according to an embodiment of the present invention; Figure 2 shows a system initial dialogue module according to an embodiment of the present invention; Figure 3 shows a system Is an intention recognition module according to an embodiment of the present invention; Figure 4 shows an intention parameter recognition module according to an embodiment of the present invention; Figure 5 shows a dialog management module according to an embodiment of the present invention; Figure 6 shows a reply and refinement module according to an embodiment of the present invention; Figure 7 shows a dialogue management module according to an embodiment of the present invention. A shortcut key creation and update module according to an embodiment of the invention; Figure 8 shows an intention recognition module according to an embodiment of the invention; Figure 9 shows an intention according to an embodiment of the invention Tree; Figure 10 shows an intention tree according to an embodiment of the present invention; Figure 11 shows an intention parameter identification module according to an embodiment of the present invention; Figure 12 shows an intention parameter recognition module according to an embodiment of the present invention; For example, an intention parameter tree based on a dual model of an embodiment of the invention; Figure 13 shows an intention parameter tree according to an embodiment of the present invention; Figures 14(a) and 14(b), respectively The figure shows the state and finite state machine of a dialogue state machine module according to an embodiment of the present invention; Figure 15 shows a system response content creation module according to an embodiment of the present invention; and Figure 16 shows It is a reply and refinement module according to an embodiment of the present invention.

The following describes the implementation of the present invention with specific specific embodiments In this way, a person familiar with this technology can easily understand the other advantages and effects of the present invention from the content disclosed in this specification, and can also be implemented or applied by other different specific embodiments.

The dialog system and method for combining intentions and shortcut keys disclosed in the present invention are described in detail as follows.

Fig. 1 is a schematic structural diagram illustrating a dialogue system according to an embodiment of the present invention. The dialog system 1000 may include an initial dialog module 100, a text input module 200, and a text preprocessing module 300, such as a four-module-based intention recognition module 400, such as a dual-module-based intention parameter recognition Module 500, dialog management module 600, system reply content creation module 700, reply and refinement module 800, shortcut key creation and update module 900. For details of each module in Figure 1, please refer to Figures 2 to 16 of the drawings respectively.

FIG. 2 is a diagram illustrating a system initial dialog module according to an embodiment of the present invention.

As shown in the figure, the initial dialogue module 100 can provide the initial dialogue content for the user to start a conversation. Among them, the initial dialog module 100 can combine the system default initial sentence content 120 and the user-specific shortcut key list 130 generated by the shortcut key creation and update module 900 to generate a user-specific and continuously updated The system initiates a dialogue 110.

Thereafter, after the user starts to input text content, the text input module 200 can receive the text content input by the user.

In one embodiment, if the user's input is voice content, the system can connect a voice recognizer to convert the voice input by the user into text After the content, the text input module 200 receives the converted text content.

The text pre-processing module 300 may perform pre-processing on the text content received by the text input module 200 to output a text pre-processing result.

In one embodiment, the word preprocessing module 300 includes, but is not limited to, a sentence normalization module, a sentence segmentation module, and a vocabulary vectorization module. The sentence normalization module can be used to normalize the text content (such as filtering out specific symbols or language in the text content and encoding conversion), and the sentence segmentation module can separate the normalized text content in word units. Then the segmented vocabulary is generated, and the vocabulary vectorization module can convert the segmented vocabulary into a vector representation.

In one embodiment, the vocabulary vectorization module can convert the segmented vocabulary into One-Hot vector representation. The length of the vector is the size of the dictionary, and each dimension of the vector represents a word 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: a kind of word "contract" The One-Hot vector is represented as [0,0,0,1,0,0,0]. Finally, the text normalization and segmentation results and the vocabulary vectorization results are integrated as the text pre-processing results, which are passed to the intent recognition module 400.

For example, when the user enters "I want #$%查票单#&", the text preprocessing module 300 will perform text preprocessing on the user input, and the result will include "I", "Yes", "Check", "Bill" and the result of word vectorization "I=[1,0,0,0,0,0,0]; Want=[0,1,0,0,0,0,0]; Check=[0,0,1,0,0,0,0]; bill [0,0,0,0,0,0,1]".

Thereafter, for example, the intention recognition module 400 based on four modules can perform intent on the text pre-processing result generated by the text pre-processing module 300 Recognize to generate intent features (such as: intent attributes, intent domain categories, intent domains, and intent names, etc.).

Fig. 3 illustrates an intention recognition module according to an embodiment of the present invention. As shown in the figure, for example, an intent recognition module 400 based on four modules may include an intent attribute recognition module 410, an intent domain category recognition module 420, an intent domain recognition module 430, and an intent name recognition module 440.

The intention attribute recognition module 410 can generate the intention attribute according to the result of text pre-processing. The intent domain category recognition module 420 can generate the intent domain category according to the text pre-processing result and the intent attribute. The intent domain recognition module 430 can generate the intent domain according to the text pre-processing result, the intent attribute, and the intent domain category. The intention name recognition module 440 can generate the intention name according to the pre-processing result of the text, the intention attribute, the intention domain category, and the intention domain.

Then, for example, the intent parameter recognition module 500 based on dual modules can perform intent parameter recognition on the text pre-processing results and the intent features generated by the intent recognition module 400 based on four modules to generate intent parameter features. (Such as: intent parameter category, intent parameter name, intent parameter representative name).

Fig. 4 illustrates an intention parameter identification module according to an embodiment of the present invention. As shown in the figure, for example, the intention parameter identification module 500 based on dual modules may include an intention parameter identification module 510 and an intention parameter representative name identification module 520.

The intention parameter recognition module 510 can be based on the text pre-processing result and the intention features generated by the intention recognition module 400 (such as: intention parameter category, intention Parameter name, intent parameter representative name), generate intent parameter category and intent parameter name. Then, the intent parameter representative name recognition model 520 can generate the intent parameter representative name according to the intent parameter category and the intent parameter name.

Thereafter, the intention feature, the intention parameter feature, the user's current dialog information, and the user-specific shortcut key list can be imported into the dialog management module 600 to generate the user's dialog status and intent tasks.

Figure 5 shows a dialog management module according to an embodiment of the present invention. As shown in the figure, the dialog management module 600 can generate user dialog status and intention tasks. The user's current dialogue information can record the user's latest dialogue state and intent task before this text input, and the intent task is a combination of intent features and intent parameter features.

In addition, importing the user-specific shortcut key table can achieve the following technical effects: when the system detects that the user enters a shortcut key, it can search for the corresponding intent task through the user-specific shortcut key table. The dialog management module 600 can generate the current dialog status and intent task through the user's text input, and determine the system processing mode and update the user's current dialog information and user dialog records according to the current dialog status. Among them, the user's current conversation information will be imported into the user's current conversation information after being updated, and the user conversation record will be imported into the user's conversation record after being updated.

In one embodiment, the current dialogue state and the intended task can be imported into the system reply content creation module 700. The system response content creation module 700 can query the system response content database (not shown) to determine the system response content. Among them, the system's reply includes but is not limited to replying to user questions, Continuous dialogue, confirmation with users, etc.

The response and improvement module 800 can receive the system response content generated by the system response content creation module 700, determine the way to respond to the user based on the system response content, and perform system performance improvement.

Fig. 6 shows a reply and refinement module according to an embodiment of the present invention. As shown in the figure, the reply and improvement module 800 may include a user reply module 810, a reply data correctness judgment module 820, and a dialogue system performance improvement module 830.

The user reply module 810 can determine the method of replying to the user and reply the message to the user. Then, the reply data correctness judgment module 820 can confirm whether the reply message is correct. Further, the reply data correctness judgment module 820 can correct the intent feature and the intent parameter feature of the user text input data whose reply is wrong to the correct intent feature and the correct intent parameter feature.

Then, the dialog system performance improvement module 830 can readjust the intention recognition module 400 and the intention parameter recognition module 500 according to the revised intention characteristics and intention parameter characteristics to achieve the dialog system performance improvement effect.

After that, the current dialog status and intent tasks can be imported into the shortcut key creation and update module 900 based on the user's dialog record to create and update the user-specific shortcut key list.

Figure 7 is a diagram illustrating a shortcut key creation and update module according to an embodiment of the present invention. As shown in the figure, the current dialog status, intent task, user dialog record 920, and user-specific shortcut key list 930 can be imported into the user-specific shortcut key creation and update module 910. Afterwards, the user’s exclusive express The shortcut key creation and update module 910 can determine whether to perform shortcut key creation and update based on the current dialog state.

If it needs to be updated, the user-specific shortcut key creation and update module 910 can calculate the latest shortcut key table applicable to the user based on the intention task and the user’s dialogue record, and then can update the user-specific shortcut key table 930 to the latest The shortcut key table is imported into the initial dialog module 100 to obtain the user-specific shortcut key table 130.

Referring again to Figure 2, in one embodiment, the system default initial sentence content 120 can be preset by manual editing. For example, the system default initial sentence content 120 can be set to "Dear user, you Okay, how can I help you?"

In addition, the shortcut key creation and update module 900 can also be used to calculate the intent names of the top few (eg, the top 3) in the system as the user-specific shortcut key table 130. Then, you can use the system default initial sentence content 120 as part of the system initial dialog 110, and then add "recommend your personal shortcut list:", "input 1:", {user-specific shortcut Key table 130 first place}, "input 2:", {user-specific shortcut key table 130 second place}, "input 3:", {user-specific shortcut key table 130 third}. And present the result of this concatenated text content to the user.

For example, in the user-specific shortcut key table 130, the first place is "check bill amount", the second place is "check my contract", and the third place is "check latest plan", then the system starts a dialogue The content of 110 is "Hello, dear users, what can I do for you? Recommend your personal shortcut list: input 1: check the bill amount; input 2: check my contract; input 3: check the most new plan".

In one embodiment, when the user uses the system for the first time or has not completed any intended tasks, the content of the user-specific shortcut key list 130 is blank, and the recommended shortcut key list will not be displayed at this time.

Then, the text input module 200 can receive the text content input by the user through the text input interface, and pass it to the text preprocessing module 300. After the text pre-processing module 300 performs pre-processing on the user's text input, the text pre-processing result can be transmitted to the intent recognition module 400.

Fig. 8 shows an intention recognition module according to an embodiment of the present invention. The intent recognition module 400 can separately train the intent attribute recognition module 410, the intent domain category recognition module 420, the intent domain recognition module 430, and the intent name recognition module through the pre-built intent tree 450 and training data 460 in advance. 440.

In one embodiment, the training data 460 is the text pre-processing result of the text dialogue collected in advance. Then, the system can label each training data 460 according to the intent features defined by the intent tree 450, so that each training data 460 is labeled with the corresponding intent attribute, intent domain category, intent domain, and intent name.

Figure 9 illustrates an intent tree according to an embodiment of the present invention. As shown in the figure, several (eg, one or more) intent attributes 451, intent field categories 452, intent fields 453, and intent names 454 of sentence intent tasks can be defined in the intent tree 450 in advance.

For example, each intent attribute 451 may have multiple intent domain categories 452, and each intent domain category 452 may have multiple intent domains. 453. Each intent field 453 may have multiple intent names 454, and each intent name 454 may be connected to the intent parameter tree 530 corresponding to it.

Figure 10 illustrates an intent tree according to an embodiment of the present invention. As shown in the figure, the intent attribute 451 includes "question intent", "confirmation intent", and "chat intent". The intent field category 452 of "question intent" includes "mobile business", "fixed network business", and "mobile business". The intention field 453 includes "bill information", "contract information", and the intent name 454 of "bill information" includes "check billing" and "check billing cycle".

Through the definition of the intent tree 450, the training data 460 can be marked with the four intent features corresponding to the intent tree 450 one by one. Then, the training data 460 can be imported into each module according to the features required by the intent attribute recognition module 410, the intent field category recognition module 420, the intent field recognition module 430, and the intent name recognition module 440, respectively.

In one embodiment, a recursive neural network algorithm can be used to train the intent attribute recognition module 410, the intent domain category recognition module 420, the intent domain recognition module 430, and the intent name recognition module 440.

After the training is completed, these four modules can be used as recognition. The recognition process is as follows: import the text pre-processing result into the intent attribute recognition module 410 after the training is completed to generate the intent attribute. Then, import the text pre-processing results and intent attributes into the intent domain category recognition module 420 after the training is completed to generate the intent domain category.

Afterwards, import text pre-processing results, intent attributes, and intent domain categories into the intent domain recognition module 430 after the training is completed to generate intent domains. Finally, you can import text pre-processing results, intent attributes, and intent fields The category and intent domain are generated by the intent name recognition module 440 after the training is completed.

For example, when inputting "I want to check the bill" and its vocabulary vectorized result "I=[1,0,0,0,0,0,0]; Want=[0,1,0,0,0 ,0,0]; check=[0,0,1,0,0,0,0]; bill [0,0,0,0,0,0,1]" in the intention recognition module 400, The intent recognition module 400 can output 4 intent features: "1. Intent attribute: question intent; 2. Intent field category: integrated business; 3. Intent field: billing information; 4. Intent name: check billing."

Then, the intent feature can be imported into the intent parameter recognition module 500 to train the intent parameter recognition module 510 and the intent parameter representative name recognition module 520.

FIG. 11 is a diagram illustrating an intention parameter identification module according to an embodiment of the present invention. As shown in the figure, the intent parameter recognition module 500 can train the intent parameter recognition module 510 and the intent parameter representative name recognition module 520 through the pre-established intent parameter tree 530 and training data 540, respectively.

For details of the intent parameter tree 530, refer to FIG. 12, which illustrates an intent parameter tree based on a dual model, for example, according to an embodiment of the present invention. As shown in the figure, several types (such as one type, two types, etc.) related to the intent name 454 can be predefined in the intent parameter tree 530, the necessary parameter category 531, the optional parameter category 532, and the representative name 533 of parameter 1 The first parameter alias 534 and so on.

In one embodiment, each parameter category (e.g., required parameter category 531, optional parameter category 532) can have multiple parameter representative names (e.g., parameter 1’s representative name 533, parameter 2’s representative name 535); and Each parameter generation The table name (such as the representative name 533 of parameter 1 and the representative name 535 of parameter 2) can have multiple parameter aliases (such as the first parameter alias 534 and the second parameter alias 536).

If the user does not input the relevant parameters of the necessary parameter category 531, the system can continue to inquire and remind the user until the user inputs the relevant parameters of the necessary parameter category 531. Conversely, if the user does not input the relevant parameters of the optional parameter category 532, the system can still generate a reply to the user, so there is no need to continue to inquire the user.

Fig. 13 illustrates an intention parameter tree according to an embodiment of the present invention. As shown in the figure, the intent name 513 is "check bill", the required parameter category 531 of the intent parameter tree 530 is "time", the optional parameter category 532 is "status", and the representative name 533 of parameter 1 is "one". Month", the representative name 535 of parameter 2 is "February", the first parameter alias 534 is "January", and the second parameter alias 536 is "JAN".

Through the definition of the intent parameter tree 530, the training data attributable to the bill checking intent can be marked with the intent parameter characteristics corresponding to the intent parameter tree 530 one by one. Then, the training data 540 can be imported into each module according to the features required by the intent parameter recognition module 510 and the intent parameter representative name recognition module 520.

In one embodiment, a keyword matching method may be used to create the intention parameter identification module 510 and the intention parameter representative name identification module 520, and the intention parameter identification module 510 and the intention parameter representative name identification module 520 are used as the intention parameters. Identify use.

The recognition process is as follows: import the text pre-processing results and intent features Go to the intention parameter identification module 510 to generate the intention parameter category and the intention parameter name; then import the intention parameter category and the intention parameter name into the intention parameter representative name identification module 520 to generate the intention parameter representative name.

For example, when the user enters "I want to check the bills that have not been paid in January", the intent feature will be: "1. Intent attribute: question intent; 2. Intent field category: integrated business; 3. Intent Field: Billing Information; 4. Intention Name: Checking Bills" is imported into the intent identification module 400.

Next, return to FIG. 5, which illustrates a dialog management module according to an embodiment of the present invention. The dialog management module 600 can create a dialog state machine module 620 based on the pre-built intent attribute table 630 and the dialog state table 640, where the intent attribute table 630 is composed of the intent attribute categories that can be recognized by the intent recognition module 400 .

In one embodiment, the intention attribute 451 may include five intention attributes such as question intention, confusion intention, confirmation intention, chat intention, and other intentions, and the dialogue state table 640 is a state list presented by the dialogue state machine module 620.

Figures 14(a) and 14(b) respectively illustrate the state of the dialogue state machine module and the finite state machine according to an embodiment of the present invention.

As shown in FIG. 14(a), the state of the dialogue state machine module 620 may include, for example, the initial state, the confirmation state, the lack of intention parameter state, the confusion state, the task completion state, the chat state, and the transfer state. As shown in the figure, for example, a combination of the 7 dialogue states and question intention, confusion intention, confirmation intention, chat intention, and other intentions as shown in Figure 14(a) can be used to establish the 14th(b) The finite state machine shown in the figure.

In addition, the user's intention characteristics, intention parameter characteristics, and user information (such as the shortcut key information including the current dialog information 650 and/or the user-specific shortcut key table 660) can be imported into the dialog management module 610. The dialog management module 610 can use the dialog state machine module 620 to analyze and output the user's dialog status and intention tasks based on the aforementioned information, and update the user's current dialog information and user dialog records.

For example, if the user’s current dialogue state is the "initial state"; the user-specific shortcut key list is "no shortcut key"; the user's intention feature is "intent attribute: question intention; intent field category: integrated service: Intent field: billing information; intent name: check bill"; intent parameter feature is "intention parameter category 1: time; intent parameter category 2: status; intent parameter name 1: January; intent parameter name 2: not paid yet ;Intention parameter representative name 1: January; Intention parameter representative name 2: unpaid".

In this case, the dialog status can be output through the dialog management module 600: "Confirmation Status", Intent Task: "Intent Attribute: Question Intention; Intent Field Type: Integrated Business; Intent Field: Billing Information; Intent Name: Check Intention parameter recognition result: "Intention parameter category 1: time; intent parameter category 2: status; intent parameter name 1: January; intent parameter name 2: not yet paid; intent parameter representative name 1: January; Intent parameter represents name 2: Unpaid".

Then, the system can confirm the intent and intent parameters to the user. For example, the system can ask the question "Do you mean "January" or "Unpaid bill", is it correct or incorrect?", when the user responds When "correct", the current dialog status can be changed from the "confirmation status" through the dialog management module 600 Status" is changed to "Task completed status".

Figure 15 illustrates a system reply content creation module according to an embodiment of the present invention. The system response content creation module 700 can determine the content to be responded to according to the dialogue status and the intent task.

If the current dialogue state is one of the "missing intention parameter state", "confusion state", and "confirmation state", the dialogue continuing module 710 can be used to generate the dialogue continuing sentence, and the purpose of generating the dialogue continuing sentence is to The user asks for further information to obtain more dialogue information. For example: when the intent task is "check the bill" and it is in the "confirmed state", the system can ask the user "Do you mean to check the bill, is it right or wrong?".

If the current conversation status is one of the "task completed status" or the "chat status", the reply content database search module 720 can reply the corresponding answer to the user. For example: when the intent task is "check bill" and is "task completed status", the system can search the user's bill and send the search result back to the user.

If the current conversation status is "transfer status", the transfer processing module 730 can notify the relevant personnel to perform the subsequent reply procedure.

Figure 16 illustrates a reply and refinement module according to an embodiment of the present invention. As shown in the figure, it is possible to select whether to use the auxiliary staff module 811 or the fully automatic reply module 812 through the reply user module 810.

The auxiliary staff module 811 can send the system reply content to the customer service staff, and the customer service staff can refer to the system reply content before replying to the user. The fully automatic reply module 812 can reply directly to the user without being processed by the customer service staff.

Then, the reply data correctness judgment module 820 can confirm whether the reply message is correct. In one embodiment, the reply data correctness judgment module 820 can use the reply result labeling module 831 to judge whether the reply message is correct, which can focus on the content of the user's question and the intent feature and the intent parameter feature identified by the robot. During the comparison, if there are different items in the comparison process, the response result marking module 831 can mark the correct intention features and the correct intention parameter features for the different items one by one.

Thereafter, the dialog system performance improvement module 830 can be used to improve the dialog system performance. Among them, the dialogue system performance improvement module 830 can regard the content of the user’s question, the correct intention feature and the correct intention parameter feature as new training data, and re-import the new training data into the intention recognition module 400, such as double Retraining is performed in the model-based intention parameter recognition module 500 to improve the accuracy of the model.

Finally, the shortcut key creation and update module 900 based on user dialogue records creates and updates the user-specific shortcut key table. For details of the shortcut key creation and update module 900, please refer to Figure 7. As shown in the figure, the dialog status, intent task, user dialog record 920, and user-specific shortcut key list 930 can be imported into the user-specific shortcut key creation and update module 910 to create and update the user-specific shortcut key list. renew.

In one embodiment, when the current dialog status is "task completed status", the system will create and update shortcut keys. The shortcut key creation and update module 900 can use the calculation method such as formula (1):

Among them, T _i represents the number of intent i in the total number of historical queries within the maximum preset limit; W _i represents the weight value of intent i; Rank _ij represents the jth time of intent i from the most recent query In the query, the total number of user queries.

For example, when the number of historical queries N=10, and the user's intention ID in these 10 queries is in order: (estimated from the latest query) 1,4,1,2,3,3,4 ,3,2,1, and preset W ₁ =1, the score of intention ID=1 is:

According to the same calculation method, it can be calculated that the first intent is Intent 1, the second intent is Intent 4, the third intent is Intent 3, and the fourth intent is Intent 2. After calculating the ranking, you can then use the calculation method such as formula (2) to create or update the shortcut key:

Wherein M represents the maximum number of user-specific shortcut key table shortcuts; C representative of the total number of historical queries the user; argTop _N (TaskScore _i) representative of the N taken before the highest score in accordance with the intended fraction i sort of TaskScore _i.

In one embodiment, the shortcut key table can be updated according to the user's new TaskScore _{i value order.} If the total number of historical queries C of the user is less than M, there will be a total of M shortcut keys. At this time, the top C intentions i with the highest _{TaskScore i can be taken out as shortcut key options.}

On the other hand, if the total number of historical queries C of the user is greater than or equal to M, there are a total of M shortcut keys. At this time, the top M intentions i with the highest _{TaskScore i can be taken out as shortcut key options.} For example: set M=3, the system can display Intent 1, Intent 4, and Intent 3 as shortcut key options in sequence.

With this calculation result, the user's most frequently used intention tasks can be taken out as shortcut keys, and the number of conversations between the user and the system can also be reduced.

The above-mentioned embodiments only illustrate the principles, features and effects of the present invention, and are not intended to limit the scope of implementation of the present invention. Anyone who is familiar with the art can comment on the above without departing from the spirit and scope of the present invention. Modifications and changes to the implementation form. Any equivalent changes and modifications made by using the content disclosed in the present invention should still be covered by the scope of the patent application. Therefore, the protection scope of the present invention should be as listed in the scope of the patent application.

100‧‧‧Starting dialogue module

200‧‧‧Text Input Module

300‧‧‧Text preprocessing module

400‧‧‧Intent Recognition Module

500‧‧‧Intent parameter recognition module

600‧‧‧Dialog Management Module

700‧‧‧System response content creation module

800‧‧‧Reply and Improvement Module

900‧‧‧Shortcut key creation and update module

1000‧‧‧Dialog System

Claims (15)

A dialogue system combining intent and shortcut keys includes: an intent recognition module that converts a text pre-processing result into an intent feature; an intent parameter recognition module that generates an intent parameter based on the text pre-processing result and the intent feature Features; a dialog management module that generates dialog states and intent tasks based on the intent feature and the intent parameter feature and based on the user-specific shortcut key table of shortcut key information, where, when a shortcut key input by the user is detected, The dialog management module searches for the intended task corresponding to the shortcut key through the user-specific shortcut key table of the shortcut key information; and a system reply content creation module based on the user-specific shortcut key table based on the shortcut key information The generated dialogue state and the intention task obtain the system reply content from a reply content database. For example, in the dialog system described in item 1 of the scope of patent application, the system's reply content creation module generates a sentence for continuing the dialog, for the sentence to ask for further inquiry to obtain the dialog information. The dialog system described in the first item of the patent application, wherein the intention parameter feature includes at least one of the following: an intention parameter category, an intention parameter name, and an intention parameter representative name, or the intention feature includes at least one of the following : Intent attribute, intent domain category, intent domain, and intent name. Such as the dialog system described in item 1 of the scope of patent application, wherein the shortcut key information includes the user’s current dialog information and the user’s exclusive Shortcut key table. The dialogue system described in item 1 of the scope of patent application, wherein the intention task is a task based on the combination of the intention feature and the intention parameter feature. For example, the dialog system described in item 1 of the scope of patent application further includes a text pre-processing module, which processes a text input content to obtain the text pre-processing result. For example, the dialog system described in item 6 of the scope of patent application, wherein the text pre-processing module includes: a sentence normalization module, which filters and encodes specific symbols and language in the text input content to Generate a normalized text content; a sentence segmentation module, which separates the normalized text content by word units, and then generates a word-separated vocabulary result; and a vocabulary vectorization module, which segments the word The result of the latter word is converted into a vector representation as the result of the pre-processing of the word. For example, the dialog system described in item 6 of the scope of patent application further includes a text input module for the user to input the text input content. A dialogue method combining intent and shortcut keys includes the following steps: converting a text pre-processing result into an intent feature; generating an intent parameter feature based on the text pre-processing result and the intent feature; according to the intent feature and the intent parameter feature and based on User-specific shortcut key list for shortcut key information to generate dialog status and intention tasks When it is detected that a user enters a shortcut key, the user-specific shortcut key table of the shortcut key information is used to find the intended task corresponding to the shortcut key; and the user-specific shortcut key based on the shortcut key information The dialog state and the intention task generated by the table obtain the system reply content from a reply content database. For example, the dialogue method described in item 9 of the scope of patent application further includes generating a sentence to continue the dialogue according to the content of the system's reply, for the sentence to ask for further inquiry to obtain dialogue information. The dialogue method according to claim 9, wherein the intention parameter feature includes at least one of the following: an intention parameter category, an intention parameter name, and an intention parameter representative name, or the intention feature includes at least one of the following : Intent attribute, intent domain category, intent domain, and intent name. For example, in the dialog method described in item 9 of the scope of patent application, the shortcut key information includes the current dialog information of the user and the user-specific shortcut key list. The dialogue method described in item 9 of the scope of patent application, wherein the intention task is a task based on the combination of the intention feature and the intention parameter feature. The dialogue method described in item 9 of the scope of patent application further includes processing a text input content to obtain the text pre-processing result before converting a text pre-processing result into an intent feature. Such as the dialog method described in item 14 of the scope of patent application, wherein the needle The step of processing the text input content to obtain the text pre-processing result includes: filtering and coding conversion of specific symbols and languages in the text input content to generate a normalized text content; the normalized text content Separate words in units of words to generate a word-breaking vocabulary result; and converting the word-breaking word result into a vector representation as the pre-processing result of the word.

Images