Method and device for determining interest point from multiple rounds of dialogues, equipment and medium

Abstract

The invention provides a method, a device, equipment and a medium for determining interest points from multiple rounds of dialogues, and the method comprises the following steps: in response to receiving a voice navigation command input by a non-first round of voice navigation dialogue, identifying an intention type corresponding to the voice navigation command; under the condition that the intention type is an intention continuing the previous text, determining a target slot value for identifying the interest point according to the non-first round of voice navigation dialogue and at least one round of historical voice navigation dialogue before the non-first round of voice navigation dialogue, and updating the current slot value of the slot position of the interest point to the target slot value; the updated interest point slot position is used for screening out the corresponding target interest point from the interest points of the map in response to the non-first-round voice navigation dialogue.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and particularly to a method, apparatus, device, and medium for determining points of interest from multi-turn conversations. Background Art

[0002] Voice navigation is a technology that uses voice commands and feedback to guide users for navigation, and is usually applied to automotive navigation systems and smartphone applications. It provides real-time driving routes and traffic information through GPS positioning to help users reach their destinations safely and conveniently. Users can set through voice commands to avoid manual operations and improve driving safety and efficiency.

[0003] In the application scenario where the dialogue system helps users search for points of interest according to the user's demand for points of interest for navigation, when the generalized point of interest output in the current turn of the conversation does not meet the expectation, there is a need to gradually refine the point of interest the user wants to go to through multi-turn conversation interaction to guide the dialogue system to output a point of interest that meets the expectation.

[0004] When the dialogue system in the related technology faces the user's voice navigation demand, it needs the user to input a complete voice navigation command each time in order to correctly execute the corresponding navigation task. For the user's demand of gradually refining the point of interest they want to go to through multi-turn conversation interaction to guide the dialogue system to output a point of interest that meets the expectation when the generalized point of interest output in the current turn of the conversation does not meet the expectation, the current dialogue system does not have a corresponding effective solution, resulting in the current dialogue system being not intelligent and convenient enough when processing voice navigation tasks. Summary of the Invention

[0005] To overcome the problems existing in the related technology, this specification provides a method, apparatus, device, and medium for determining points of interest from multi-turn conversations.

[0006] According to the first aspect of the embodiments of this specification, a method for determining points of interest from multi-turn conversations is provided. The method includes:

[0007] Respond to a voice navigation command received in a non-first-turn voice navigation conversation, and identify the type of intent corresponding to the voice navigation command;

[0008] When the type of intent is a continuation-of-previous-intent, determine a target slot value for identifying a point of interest according to the non-first-turn voice navigation conversation and at least one previous turn of historical voice navigation conversations, and update the current slot value of the point-of-interest slot to the target slot value; the updated point-of-interest slot is used to screen out corresponding target points of interest from the points of interest on the map in response to the non-first-turn voice navigation conversation.

[0009] According to a second aspect of the embodiments of the present specification, there is provided an apparatus for determining a point of interest from a multi-round conversation, the apparatus comprising:

[0010] An intent recognition module, configured to recognize an intent type corresponding to the voice navigation command in response to receiving a voice navigation command input in a non-first-round voice navigation conversation;

[0011] An interest point slot update module, configured to, when the intent type is an intent to continue the previous context, determine a target slot value for identifying an interest point according to the non-first-round voice navigation conversation and at least one previous round of historical voice navigation conversations, and update the current slot value of the interest point slot to the target slot value; the updated interest point slot is used to screen out corresponding target interest points from the interest points on the map in response to the non-first-round voice navigation conversation.

[0012] According to a third aspect of the embodiments of the present specification, there is provided an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, the steps of the method described in the first aspect are implemented.

[0013] According to a fourth aspect of the embodiments of the present specification, there is provided a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0014] The technical solutions provided by the embodiments of the present specification may include the following beneficial effects:

[0015] In the embodiments of the present specification, for a non-first-round voice navigation conversation in a multi-round conversation, the intent type corresponding to the voice navigation command input in the non-first-round voice navigation conversation is recognized. When the intent type is an intent to continue the previous context, a target slot value for identifying an interest point is determined according to the non-first-round voice navigation conversation and at least one previous round of historical voice navigation conversations, and the current slot value of the interest point slot is updated to the target slot value. The updated interest point slot is used to screen out corresponding target interest points from the interest points on the map in response to the non-first-round voice navigation conversation. It can be seen that in this embodiment, in response to the requirement that the user hopes to gradually refine the interest point he wants to go to through multi-round conversation interaction, a corresponding intent to continue the previous context is set. When the recognized intent type is an intent to continue the previous context, the target interest point output is made to meet the user's expectations by updating the interest point slot. In this embodiment, it is not necessary for the user to input a complete voice navigation instruction each time, but the form of multi-round conversation can be used to guide the dialogue system to output an interest point that meets the expectations, thereby improving the intelligence and convenience of the current dialogue system when processing voice navigation tasks.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this specification. Brief Description of the Drawings

[0017] The drawings herein are incorporated into and constitute a part of this specification, showing embodiments consistent with this specification, and are used together with the specification to explain the principles of this specification.

[0018] Figure 1 is a flowchart of a method for determining points of interest from multi-round conversations shown according to an exemplary embodiment of this specification.

[0019] Figure 2 is a schematic diagram of determining target points of interest from two-round conversations shown according to an exemplary embodiment of this specification.

[0020] Figure 3 is a schematic diagram of the structure of an electronic device shown according to an exemplary embodiment of this specification.

[0021] Figure 4 is a block diagram of a device for determining points of interest from multi-round conversations shown according to an exemplary embodiment of this specification. Detailed Embodiment

[0022] Voice navigation is a technology that uses voice commands and feedback to guide users for navigation, and is usually applied to automotive navigation systems and smartphone applications. It provides real-time driving routes and traffic information through GPS positioning, helping users reach their destinations safely and conveniently. Users can set through voice commands, avoiding manual operations and improving driving safety and efficiency.

[0023] In the application scenario where the dialogue system helps users search for points of interest according to the user's navigation point-of-interest requirements, when the generalized points of interest output in the current round of conversation do not meet the expectations of the user, there is a need to gradually refine the points of interest the user wants to go to through multi-round conversation interactions to guide the dialogue system to output points of interest that meet the expectations. For example, the user inputs "Navigate to ** scenic spot" in the first round of conversation. At this time, the dialogue system can output a list of points of interest associated with "** scenic spot" and can remind the user to select the target location they want to go to. After the user selects, the corresponding navigation task will be triggered. However, the user may find that the expected location does not exist in the list of points of interest output in the first round of conversation. For example, the user wants to go to the parking lot near "** scenic spot", but it does not appear in the list of points of interest. At this time, the user wants to initiate a second round of conversation and supplement the input "the parking lot near it" in the second round of conversation, hoping that the dialogue system can associate the first round of conversation and can output a list of points of interest associated with "the parking lot near ** scenic spot" or directly initiate the navigation task to "the parking lot near ** scenic spot".

[0024] When a dialogue system in the related art faces the user's navigation requirement, the user must input a complete voice navigation command each time. For example, it can easily handle the first-round dialogue initiated by the above user. However, when continuing to face the second-round dialogue continuously initiated by the above user, since the current dialogue system does not pay attention to this requirement, that is, the user has the requirement to gradually refine the point of interest they want to go to through multiple rounds of dialogue, and correspondingly, the corresponding intention recognition has not been developed, it may still mis-recognize it as a navigation intention and attempt to extract the slot value of the point of interest from it. However, the slot value of the point of interest extracted may be "parking lot near the scenic spot", and this slot value of the point of interest is not associated with the previous round of dialogue, resulting in the inability to accurately determine the target point of interest expected by the user from the multiple rounds of dialogue of the user. Therefore, in multiple rounds of dialogue, for each round of dialogue, the user needs to input a complete voice navigation command. For example, in the second-round dialogue, the user needs to input "parking lot near the ** scenic spot", and the dialogue system can then completely extract the slot value of the point of interest from this voice navigation command.

[0025] It can be seen from this that for the situation where the generalized point of interest output by the user in the current round of dialogue does not meet the expectation, and the user hopes to gradually refine the point of interest they want to go to through multiple rounds of dialogue interaction to guide the dialogue system to output a point of interest that meets the expectation, the current dialogue system does not have a corresponding effective solution, resulting in the current dialogue system being not intelligent and convenient enough when processing voice navigation tasks.

[0026] In view of the above technical problems, this specification proposes a method for determining a point of interest from multiple rounds of dialogue to meet the user's requirement of guiding the dialogue system to output a point of interest that meets the expectation through multiple rounds of dialogue.

[0027] Next, the embodiments of this specification will be described in detail.

[0028] As Figure 1 shown, Figure 1 is a flowchart of a method for determining a point of interest from multiple rounds of dialogue shown in this specification according to an exemplary embodiment, including steps 101-102:

[0029] Step 101: In response to receiving a voice navigation command input in a non-first-round voice navigation dialogue, identify the type of intention corresponding to the voice navigation command.

[0030] Step 102: When the type of intention is a continuation-of-previous-intention, determine a target slot value for identifying a point of interest according to the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogue, and update the current slot value of the point of interest slot to the target slot value; the updated point of interest slot is used to screen out the corresponding target point of interest from the points of interest on the map in response to the non-first-round voice navigation dialogue.

[0031] This method can be applied to the dialogue systems of voice assistants in various map navigation software, voice assistants on various intelligent devices, in-vehicle voice assistants, etc.

[0032] A point of interest (POI) can be a geographical location entity marked with specific functions or services on a map. For example, hot pot restaurants, hospitals, and universities, etc. Of course, it can also be a specific geographical location, such as a road, a street, etc.

[0033] In the application scenario of point-of-interest navigation, users can express their needs associated with the point of interest, and the dialogue system helps them quickly find the target point of interest that meets the requirements. Specifically, in the dialogue system, users can express their needs associated with the point of interest through voice input or text input. For example, if a user wants to find the nearest gas station, they can start a conversation "Where is the nearest gas station?", and then the dialogue system can show a list of nearby gas stations, and the gas stations in this list can be sorted by distance. Another example, if a user wants to find a top-three hospital, they can start a conversation "I want to go to a top-three hospital", and then the dialogue system can show a list of top-three hospitals in the user's city.

[0034] Each round of dialogue in the multi-round dialogue in this solution may not be limited to being a dialogue for voice navigation tasks. For example, in an in-vehicle voice assistant, a user starts three rounds of conversations. The first round of conversation is "Open the window", the second round of conversation is "Navigate to [a certain] store", and the third round of conversation is "[It is the] store on [a certain] road". It can be seen that there can be other task-type conversations in the multi-round dialogue. Of course, other rounds of conversations other than voice navigation conversations may not be limited to being in the form of voice interaction, and can also be in the form of text input conversations. And it should be noted that the non-first-round voice navigation conversation in the multi-round dialogue in this solution can be any non-first-round voice navigation conversation in the multi-round dialogue.

[0035] Intent recognition can be to determine the hidden intention or purpose behind it by analyzing the text or voice input by the user, and classify it into a predefined intent type, so as to drive the execution of the subsequent task corresponding to this intent type. For example, assume that the user inputs "Book a flight ticket to Shanghai tomorrow" in this round of conversation, then the content input by this user can be classified into the flight ticket booking intent, and thus trigger corresponding actions, such as calling the flight ticket booking interface to execute the flight ticket booking task.

[0036] The intention of continuing the above text can be generally understood as the intention that in the scenario of navigation points of interest, when the generalized point of interest output in the previous round of conversation does not meet the expectations of the user, the user hopes to gradually refine the point of interest they want to go to through multi-round conversation interaction to guide the conversation system to output the point of interest that meets the expectations. For example, if the user's first-round conversation input is "Navigate to the store on ** Road", and the second-round conversation input is "The one on ** Road", then the intention type of the second-round conversation input can be identified as the intention of continuing the above text. For the user input of "Open the sunroof", the intention type of this input can be identified as the "Open window intention". Of course, in addition to the intention of continuing the above text in this solution, the conversation system may also simultaneously perform other intention recognitions on the user's input. This solution does not impose any restrictions on the methods of other intention recognitions and how to execute tasks after recognition.

[0037] In a conversation system, a slot can be a predefined variable used to extract key information related to the intention from the user's input. Such as "time", "location", and "service type", etc. The slot value is the specific data in the user's input corresponding to the slot, that is, the actual content filled into the slot. For example, in the intention type of "flight booking", the slot of "departure city" can be filled with "Beijing". Correspondingly, the point-of-interest slot can be a slot used to fill in the point of interest. For example, if the point-of-interest information extracted from the user input is "restaurant", the point-of-interest slot can be specifically filled with "restaurant"; if the point-of-interest information extracted from the user input is "hospital", the point-of-interest slot can be specifically filled with "hospital".

[0038] In one embodiment, for non-first-round voice navigation conversations in multi-round conversations, the non-first-round voice navigation conversation can be any non-first-round voice navigation conversation in the multi-round conversations, and the following execution logics for each non-first-round voice navigation conversation are the same. Next, only a certain non-first-round voice navigation conversation in the multi-round conversations will be introduced:

[0039] In response to receiving the voice navigation command input in the non-first-round voice navigation conversation, identify the intention type corresponding to the voice navigation command.

[0040] In the case where the intention type is the intention of continuing the above text, determine the target slot value used to identify the point of interest according to the non-first-round voice navigation conversation and at least one previous round of historical voice navigation conversations, and update the current slot value of the point-of-interest slot to the target slot value. The updated point-of-interest slot is used to filter out the corresponding target point of interest from the points of interest on the map in response to the non-first-round voice navigation conversation.

[0041] It should be noted that if the intent type of the voice navigation command recognized for the non-first-round voice navigation dialogue input is not the intent to continue the previous context, corresponding processing can be performed according to other recognized intents, and this specification does not impose any restrictions on this.

[0042] Among them, for each round of voice navigation dialogue in the multi-round voice navigation dialogue, the same point of interest slot can be processed. The point of interest slot before update can be filled with a slot value for identifying the point of interest determined according to the at least one round of historical voice navigation dialogue. For example, assume that the user continuously initiates three rounds of voice navigation dialogues. For the first round of voice navigation dialogue, the first point of interest slot value in the point of interest slot can be determined according to the first round of voice navigation dialogue, and this point of interest slot is used to filter out the corresponding target point of interest from the points of interest on the map in response to the first round of voice navigation dialogue. For the second round of voice navigation dialogue, if the intent type of the voice navigation command recognized for the second round of voice navigation dialogue input is the intent to continue the previous context, the second point of interest slot value for identifying the point of interest can be determined according to the second round of voice navigation dialogue and the first round of voice navigation dialogue, and the first point of interest slot value in the point of interest slot is updated to the second point of interest slot value. The updated point of interest slot is used to filter out the corresponding target point of interest from the points of interest on the map in response to the second round of voice navigation dialogue. For the third round of voice navigation dialogue, if the intent type of the voice navigation command recognized for the third round of voice navigation dialogue input is the intent to continue the previous context, the third point of interest slot value for identifying the point of interest can be determined according to the third round of voice navigation dialogue and the first and second rounds of voice navigation dialogue, and the second point of interest slot value in the point of interest slot is updated to the third point of interest slot value. The updated point of interest slot is used to filter out the corresponding target point of interest from the points of interest on the map in response to the third round of voice navigation dialogue. It can be seen that in each non-first-round voice navigation dialogue of the multi-round dialogue, if the intent type of the voice navigation command recognized for the current round of voice navigation dialogue input is the intent to continue the previous context, the target slot value for identifying the point of interest can be determined according to the current round of voice navigation dialogue and at least one round of historical voice navigation dialogue before it, and the current slot value in the point of interest slot is updated to the target slot value. The updated point of interest slot is used to filter out the corresponding target point of interest from the points of interest on the map in response to the current round of voice navigation dialogue.

[0043] In this embodiment, for non-first-round voice navigation dialogues in multi-round dialogues, the intent type corresponding to the voice navigation command input in the non-first-round voice navigation dialogue is recognized. When the intent type is the intent to continue the previous context, the target slot value for identifying the point of interest is determined based on the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogues, and the current slot value of the point of interest slot is updated to the target slot value. The updated point of interest slot is used to filter out the corresponding target point of interest from the points of interest on the map in response to the non-first-round voice navigation dialogue. It can be seen that in this embodiment, in response to the user's need to gradually refine the point of interest they want to go to through multi-round dialogue interaction, the corresponding intent to continue the previous context is set. When the recognized intent type is the intent to continue the previous context, the output target point of interest meets the user's expectations by updating the point of interest slot. In this embodiment, it is not necessary for the user to input a complete voice navigation instruction each time, but rather the form of multi-round dialogues can be used to guide the dialogue system to output the point of interest that meets the expectations, thereby improving the intelligence and convenience of the current dialogue system when processing voice navigation tasks.

[0044] In one embodiment, in response to receiving the voice navigation command input in the non-first-round voice navigation dialogue, when it is recognized that the intent type of the voice navigation command is the intent to continue the previous context, the target slot value for identifying the point of interest can be determined based on the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogues, and the current slot value of the point of interest slot is updated to the target slot value. The associated slot value associated with the point of interest can also be extracted from the voice navigation command and filled into the corresponding associated slot; and / or, the historical slot associated with the point of interest determined in the at least one previous round of historical voice navigation dialogues is obtained. In addition to the updated point of interest slot, the associated slot and / or the historical slot can also be used to filter out the corresponding target point of interest from the points of interest on the map in response to the non-first-round voice navigation dialogue.

[0045] It should be noted that the associated slot associated with the point of interest and the historical slot associated with the point of interest can both be regarded as the slots associated with the point of interest. The only difference lies in the acquisition timing. The former is acquired in the current round of voice navigation dialogue, while the latter is acquired in the historical voice navigation dialogue. The associated dimension of the slot associated with the point of interest can be the attribute information of the point of interest, such as business hours, facility conditions, service types, ratings, etc., or the user's preference information for the point of interest, etc. Correspondingly, the slot associated with the point of interest can be filled with slot values such as business hours, facility conditions, service types, ratings, and user preferences. Of course, there are also other associated dimensions, which will not be elaborated one by one in this specification.

[0046] For example, in the first round of the voice navigation conversation, the user inputs "Navigate to ** Restaurant, and the restaurant needs to have a parking lot", and in the second round of the voice navigation conversation, the user supplements the input with "It needs to be near ** Road and its business hours are until 12 o'clock at night". For the first round of the voice navigation conversation, the slot value that can be filled in the point of interest slot is "** Restaurant", and "parking lot" is filled in the facility condition slot associated with the point of interest slot. For the second round of the voice navigation conversation, it can be recognized that the intention type of the voice navigation command input in the second round of the voice navigation conversation is the intention to continue the previous text. The target slot value used to identify the point of interest can be determined based on the first round and the second round of the voice navigation conversations as "** Restaurant on ** Road", and the slot value "** Restaurant" filled in the point of interest slot is updated to "** Restaurant on ** Road". Also, "12 o'clock at night" is extracted from the voice navigation command input in the second round of the voice navigation conversation and filled in the business hours slot associated with the point of interest. Then, in the second round of the voice navigation conversation, in addition to the point of interest slot, the business hours slot extracted in this round and the facility condition slot extracted in the previous round are both used to filter out the corresponding target points of interest from the points of interest on the map in response to the voice navigation conversation in this round. At this time, the filtered target points of interest can be ** Restaurants that meet the conditions of being located on ** Road, having a parking lot, and having business hours until 12 o'clock at night.

[0047] In this embodiment, on the basis of implementing the inheritance of the slots of the generalized points of interest input by the user in the previous round of the voice navigation conversation, the requirement of inheriting the slots associated with the points of interest input by the user in the previous round of the voice navigation conversation is also implemented, so as to integrate the fragmented information about the points of interest input in multiple rounds of the voice navigation conversation and maintain the coherence of the voice navigation conversation.

[0048] In one embodiment, when identifying the intention type corresponding to the voice navigation command, the voice of the voice navigation command can be converted into a voice navigation text; the first deep learning model is used to identify the intention type corresponding to the voice navigation text.

[0049] In one embodiment, the first deep learning model can be a classification model such as BERT, Robustly Optimized BERT Approach (RoBERTa), Text Convolutional Neural Network (TEXTCNN), etc. The model framework of the first deep learning model is not limited in this specification. Of course, in order to train the first deep learning model to have the ability to recognize the intention of continuing the above text, this specification provides a first continuation dataset for training the first deep learning model. Any sample in the first continuation dataset includes a refined intention text and a label; the content of the refined intention text is used to express the user's intention to refine the features of the interest point, and the label is used to classify the intention type.

[0050] Among them, the refined intention text can be obtained by summarizing the words used by the user in the navigation scenario of the interest point that needs to continue the above text, generalizing these words into corresponding sentence patterns, and expanding the sentence patterns into refined intention texts. The label can be represented by 0 to i, where i is a positive integer not less than 1. Among them, 0 indicates that the intention type is to continue the above text intention, and 1 to i are other intention types. In other words, the first deep learning model of this solution is not limited to being able to recognize the intention of continuing the above text, but can also recognize other intentions, such as the intention of booking a flight, the intention of chatting, the intention of opening a window, etc.

[0051] In one embodiment, this specification provides several common navigation scenarios of interest points that need to continue the above text:

[0052] The first scenario is that the user continues to refine the store name entered in the historical round of voice navigation dialogue into the branch name of the store. For example, the user can enter "Navigate to ** store" in the first round of voice navigation dialogue. At this time, the dialogue system can output a list of interest points of each branch of the ** store in the city. When the user finds that the list of interest points output in the first round of dialogue exists or does not exist the expected store, the user can enter "I want to go to the one of ** branch" in the second round of voice navigation dialogue. When the branch name entered by the user in the second round of voice navigation dialogue is in the list of interest points output in the first round of navigation dialogue, the navigation task can be directly initiated; if not, the target store can be re-determined according to the inputs of these two rounds of voice navigation dialogue.

[0053] The second scenario is that the user refines the point of interest (POI) entered in the historical round of voice navigation dialogue. The POI is defaultly located in the user's city and is further refined to the city expected by the user. For example, the user can enter "Navigate to ** place" in the first round of voice navigation dialogue. If the POI is distributed in multiple cities, the dialogue system may return the "** place" in the user's city. The user can enter "It's the place in ** city" in the second round of voice navigation dialogue. The dialogue system can re-determine the target POI as "** place in ** city" by combining "** place" and "** city".

[0054] The third scenario is that the user refines the POI entered in the historical round of voice navigation dialogue to a POI on a certain street / road. For example, the user can enter "Navigate to ** store" in the first round of voice navigation dialogue. At this time, the dialogue system can output a list of POIs of each branch of the ** store in the city. Whether the user finds the expected store in the list of POIs output in the first round of dialogue or not, the user can enter "I want to go to the one on ** road" in the second round of voice navigation dialogue. When the branch name entered by the user in the second round of voice navigation dialogue is in the list of POIs output in the first round of navigation dialogue, the navigation task can be directly initiated; if not, the target store can be re-determined based on the inputs of these two rounds of voice navigation dialogue.

[0055] The fourth scenario is that the user refines the POIs such as ** community, ** building, etc. entered in the historical round of voice navigation dialogue to a certain door / phase / parking lot / terminal building. For example, the user can enter "I want to go to ** community" in the first round of voice navigation dialogue. At this time, the dialogue system can output places in the city with the name of "** community". The user can also refine the community address in the second round of voice navigation dialogue. For example, continue to enter "I want to go to the third phase" in the second round of navigation dialogue. At this time, the dialogue system can re-determine the target POI as "** community, third phase" by combining "** community" and "third phase".

[0056] The fifth scenario is that the user refines the POI entered in the historical round of voice navigation dialogue to a place near the POI, such as a nearby store, a nearby parking lot, etc. For example, the user can enter "I want to go to ** road" in the first round of voice navigation dialogue. If the user wants to further refine the parking lot near ** road, the user can enter "A place where it's convenient to park nearby" in the second round of voice navigation dialogue. At this time, the dialogue system can re-determine the target POI as "** road ** parking lot" by combining "** road" and "** parking lot".

[0057] Of course, there are other navigation scenarios that need to continue the POI mentioned above, which will not be elaborated here.

[0058] In one embodiment, when identifying the intent type corresponding to a voice navigation command, the voice navigation text may be matched with a preset sentence pattern. If the match is successful, the intent type corresponding to the voice navigation command is identified as the intent to continue the previous text. If the match fails, the task corresponding to the identified intent type of the voice navigation command is executed. For example, when a user inputs "Open the window", the intent of this task instruction is not to continue the previous text, but to open the window, and the task corresponding to the intent of opening the window can be executed.

[0059] Among them, the structure of the preset sentence pattern can be composed of a prefix word + a point of interest + a suffix word. For the prefix word, it can be [I want to go to, I want to go, Navigate to, Navigate to, ...]. For the suffix word, it can be "that, that store, the place to park", and the point of interest can be extracted through a slot recognition model. For the obtained voice navigation text, the slot value of the point of interest can be extracted from it. If it cannot be extracted, the match fails. If the slot value of the point of interest is extracted, it is then continued to determine whether the description of the voice navigation text conforms to the structure of the preset sentence pattern. If it conforms, the match is successful; if it does not conform, the match fails.

[0060] In another embodiment, the voice navigation text is matched with a preset sentence pattern. If the match is successful, the intent type corresponding to the voice navigation command is identified as the intent to continue the previous text; if the match fails, the intent type corresponding to the voice navigation text is identified through a first deep learning model.

[0061] In this embodiment, the method of using rule matching to identify intents is faster and more accurate, but it is difficult to handle uncommon new sentence patterns. Although the deep learning model is less efficient and slower than the method of rule matching, it has stronger generalization ability and can handle uncommon new sentence patterns. By using the complementary method of rule matching and the deep learning model in this embodiment, the speed and accuracy of identifying intents can be comprehensively improved.

[0062] In one embodiment, when the intent type is the intent to continue the previous text, when determining the target slot value for identifying the point of interest according to the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogues, it can be determined whether the type of the previous round of dialogue of the non-first-round voice navigation dialogue is a navigation dialogue. When the type of the previous round of dialogue is a navigation dialogue and the intent type is the intent to continue the previous text, the target slot value for identifying the point of interest is determined according to the non-first-round voice navigation dialogue and the previous round of dialogue. [[ID=...]]

[0063] For example, assume that for a multi-round conversation, the user inputs "Open the window" in the i-th round of the conversation and "Navigate to ** Road" in the (i + 1)-th round of the conversation, where i is a positive integer not less than 1. Of course, there may be other conversations before the i-th round of the conversation. For the (i + 1)-th round of the conversation, it can be determined that the previous round of the conversation is not a navigation conversation, so the task corresponding to continuing the above-mentioned intention does not need to be executed.

[0064] Assume that the user inputs "Navigate to ** Store" in the i-th round of the conversation and "The one on ** Road" in the (i + 1)-th round of the conversation. For the (i + 1)-th round of the conversation, it can be determined that the previous round of the conversation is a navigation conversation, and the intention type corresponding to the voice navigation instruction input in this round of the conversation is the intention of continuing the above text. Then, the target slot value for identifying the point of interest can be determined as "** Store on ** Road" based on this round of the conversation and the previous round of the conversation.

[0065] It should be noted that in this embodiment, it is not necessary to consider whether there is a historical voice navigation conversation of the navigation conversation type before the i-th round of the conversation. Instead, the target slot value for identifying the point of interest can be determined by combining this round of the conversation and the previous round of the conversation, so as to ensure the continuity of the inheritance of the point of interest slot.

[0066] In this embodiment, by identifying whether the previous round of the conversation is a voice navigation conversation, it is possible to avoid erroneously inheriting the content of the previous non-voice navigation conversation to this round of the voice navigation conversation. Moreover, it is also possible to remedy the situation where the previous round of the conversation is a non-voice navigation conversation and the intention of this round of the voice navigation conversation is mis-identified as the intention of continuing the above text, because the first-round voice navigation conversation generally should not be identified as the intention of continuing the above text.

[0067] In one embodiment, when determining the target slot value for identifying the point of interest according to the non-first-round voice navigation conversation and at least one previous round of historical voice navigation conversations, the voice navigation text corresponding to the voice navigation command and the voice navigation texts input in at least one previous round of historical voice navigation conversations can be input into the second deep learning model, and the target slot value output by the second deep learning model can be obtained.

[0068] For example, if the user initiates three consecutive voice navigation conversations, for the third voice navigation conversation, at least one previous round of historical voice navigation conversations can be the first voice navigation conversation and the second voice navigation conversation, or it can be the previous voice navigation conversation of this round, that is, the second voice navigation conversation.

[0069] In addition to inputting the voice navigation text corresponding to the voice navigation command and the voice navigation texts input in at least one round of historical voice navigation conversations before it into the second deep learning model, the current slot value filled in the point of interest slot can also be jointly input into the second deep learning model. Of course, in order to obtain the target slot value, the input data can also be other content, and this specification does not impose any restrictions on the content of the input data.

[0070] In one embodiment, the second deep learning model can adopt the frameworks of various large language models, and this specification does not limit the specific framework of the second deep learning model. The input data of any sample in the second continuation dataset for training the second deep learning model includes multi-round voice navigation texts, and the output data includes the slot values used to identify points of interest extracted from the multi-round voice navigation texts.

[0071] Among them, the multi-round voice navigation texts can be multiple context-related texts with points of interest that the user hopes to gradually refine the point of interest they want to go to through multi-round dialogue interactions to guide the dialogue system to output a point of interest that meets expectations when the generalized point of interest output in the current round of dialogue does not meet expectations. For example, the multi-round voice navigation texts can be "Navigate to ** store" and "The one on ** road", then the slot value used to identify the point of interest extracted from this multi-round voice navigation text is "The ** store on ** road". Another example is that on the basis of the previous example, the user inputs a new voice navigation text "The parking lot near it", then the slot value used to identify the point of interest extracted from the latest multi-round voice navigation text is updated to "The parking lot near the ** store on ** road". The multi-round voice navigation texts can be summarized from the common phrases of users in various navigation scenarios, and this specification does not impose any restrictions on the method of obtaining the multi-round voice navigation texts.

[0072] In one embodiment, when determining the target slot value used to identify the point of interest according to the non-first-round voice navigation conversation and at least one round of historical voice navigation conversations before it, a first slot value can be obtained, and the first slot value is the slot value used to identify the point of interest determined by the at least one round of historical voice navigation conversations. A second slot value used to identify the point of interest is extracted from the voice navigation instruction, and the target slot value is determined according to the first slot value and the second slot value.

[0073] Specifically, a second slot value used to identify the point of interest can be extracted from the voice navigation instruction through a slot extraction model. The slot extraction technology for extracting points of interest is essentially the same as the slot extraction technology for extracting other content, and this specification does not impose any restrictions on the slot extraction model.

[0074] When determining the target slot value based on the first slot value and the second slot value, the first slot value and the second slot value can be concatenated to obtain the target slot value. For example, if the first slot value is "** Building" and the second slot value is "Phase III", the target slot value can be obtained by concatenating the first slot value and the second slot value, that is, "** Building Phase III". Among them, the concatenation rule can consider whether the concatenation order conforms to the naming convention and whether the concatenated name is ambiguous, etc.

[0075] In this embodiment, by extracting the first slot value and the second slot value and concatenating them according to the rules, the target interest point can be quickly determined, thereby improving the response speed of the dialogue system.

[0076] In one embodiment, there may be a situation where the determination of the target slot value based on the first slot value and the second slot value fails. The reason may be that the slot extraction model fails to extract the second slot value, or extracts the wrong second slot value, resulting in the inability to concatenate the first slot value and the second slot value into the target slot value according to the rules. This specification proposes that in the case where the determination of the target slot value based on the first slot value and the second slot value fails, the voice navigation text corresponding to the voice navigation command and the voice navigation texts input in at least one previous round of historical voice navigation conversations are input into the second deep learning model, and the target slot value output by the second deep learning model is obtained.

[0077] In this embodiment, the speed of determining the target slot value by the rule is faster and can quickly respond to the dialogue system. However, due to reasons such as the inflexibility of the slot extraction model or the concatenation rule, the target slot value cannot be determined. Therefore, in the case where the determination by the rule fails, the second deep learning model can be used to deeply analyze multiple rounds of voice navigation texts and extract the target slot value from them. By combining the rule and the deep learning model, the response speed can be guaranteed while the accuracy of determining the target slot value is also improved.

[0078] In one embodiment, in order to verify the correctness of the target slot value output by the second deep learning model during actual use, the parameters of the second deep learning model are continuously optimized based on the actual output results, thereby continuously improving the accuracy of the second deep learning model in determining the target slot value. In addition to outputting the target slot value, the second deep learning model can also output the merged text, which is determined according to the actual input data, that is, determined according to the voice navigation text corresponding to the voice navigation command and the voice navigation texts input in at least one previous round of historical voice navigation conversations. The merged text and the target slot value output by the second deep learning model can be obtained and saved, and the merged text and the target slot value can be used to optimize the second deep learning model.

[0079] For example, when running online, the actual slot value can be extracted from the merged text output by the second deep learning model. For example, the merged text can be input into a module for extracting point-of-interest slot values from speech navigation text, and the actual slot value output by the module can be obtained. If the actual slot value is the same as the target slot value, it indicates that the output result of the second deep learning model is correct; if the actual slot value is different from the target slot value, it indicates that the output result of the second deep learning model is incorrect. Error cases where the actual slot value is different from the target slot value can be saved. The error cases include the data actually input to the second deep learning model, the actual slot value corresponding to the data, the target slot value, and the merged text. The error cases are used to analyze offline the reason for the incorrect output of the second deep learning model or to re-fine-tune the parameters of the second deep learning model based on the error cases.

[0080] Of course, it is also possible to directly obtain and save the merged text output by the second deep learning model, the target slot value, and the data input to the second deep learning model corresponding to the merged text. Offline, verify whether the actual slot value extracted from the merged text is the same as the target slot value. If not, use the input data and the actual slot value as training data to re-fine-tune the parameters of the second deep learning model.

[0081] In this embodiment, the second deep learning model can be continuously improved through the actual usage data of online users, thereby improving the adaptability of the second deep learning model to the actual scenario.

[0082] In one embodiment, Figure 2 is a schematic diagram showing a method for determining a target point of interest from a two-round conversation according to an exemplary embodiment of this specification. As Figure 2 shown, this specification introduces a method for determining a point of interest from a multi-round conversation in combination with an actual application scenario. Assume that the multi-round conversation is specifically two rounds, and the inputs in the first round of conversation are divided into two types of conversation tasks: Q1: "I want to go to ** store" and Q1: "Turn on the air conditioner". The former is a navigation task, while the latter is a vehicle function control task; the input in the second round of conversation, that is, the current round of conversation, is Q2: "The one on ** road".

[0083] Assume that the user inputs "Turn on the air conditioner" in the first round of conversation. It can be determined that the type of this round of conversation is not a navigation conversation. Then, the point-of-interest slot value can be extracted from the input of the current conversation and filled into the point-of-interest slot. The finally output point-of-interest slot is [** road].

[0084] Assume that the user inputs "I want to go to ** store" in the first round of conversation. It can be determined that the type of this round of conversation is a navigation conversation. Then, it can be further determined whether the input intention of the current conversation is to continue the intention of the previous text. If not, the point of interest slot value can be extracted from the input of the current conversation and filled into the point of interest slot; if so, the target slot value used to identify the point of interest can be determined in a rule-based manner.

[0085] When determining the target slot value used to identify the point of interest in a rule-based manner, the first slot value can be obtained. The first slot value can be the slot value used to identify the point of interest determined in the first round of conversation, that is, ** store. The second slot value used to identify the point of interest, that is, ** road, can be extracted from the input of this round of conversation. And the target slot value, that is, ** store ** road, can be determined according to the first slot value and the second slot value. If the determination of the target slot value is successful, the current slot value in the current point of interest slot can be updated to the target slot value. If the determination of the target slot value fails, the target slot value can be determined through the second deep learning model, and the current slot value in the current point of interest slot can be updated to the target slot value.

[0086] Corresponding to the embodiments of the foregoing method, this specification also provides embodiments of a device and a terminal to which the device is applied.

[0087] Figure 3 It is a schematic structural diagram of an electronic device shown according to an exemplary embodiment of this specification. As Figure 3 shown, at the hardware level, the electronic device 300 includes a processor 302, an internal bus 304, a network interface 306, a memory 308, and a non-volatile memory 310. Of course, there may also be other hardware required for other services. One or more embodiments of this specification can be implemented in a software manner. For example, the processor 302 reads the corresponding computer program from the non-volatile memory 310 into the memory 308 and then runs it. Of course, in addition to the software implementation manner, one or more embodiments of this specification do not exclude other implementation manners, such as a logic device or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logic module, and can also be hardware or a logic device.

[0088] Figure 4 It is a block diagram of a device for determining a point of interest from multi-round conversations shown according to an exemplary embodiment of this specification. As Figure 4 shown, the device can be applied to the electronic device 300 shown as Figure 3 shown to implement the technical solution of this specification. The device includes:

[0089] An intention recognition module 402, configured to recognize an intention type corresponding to the voice navigation command in response to receiving a voice navigation command input in a non-first-round voice navigation conversation;

[0090] The point of interest slot update module 404 is configured to, when the intent type is the intent of continuing the previous text, determine a target slot value for identifying a point of interest according to the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogues, and update the current slot value of the point of interest slot to the target slot value; the updated point of interest slot is used to filter out corresponding target points of interest from the points of interest on the map in response to the non-first-round voice navigation dialogue.

[0091] Optionally, the point of interest slot update module 404 is further configured to extract an associated slot value associated with the point of interest from the voice navigation command and fill it into a corresponding associated slot; and / or obtain a historical slot associated with the point of interest determined in the at least one previous round of historical voice navigation dialogues; wherein, in addition to the updated point of interest slot, the associated slot and / or the historical slot are also used to filter out corresponding target points of interest from the points of interest on the map in response to the non-first-round voice navigation dialogue.

[0092] Optionally, the intent recognition module 402 is specifically configured to convert the voice of the voice navigation command into a voice navigation text; identify an intent type corresponding to the voice navigation text through a first deep learning model; or match the voice navigation text with a preset sentence pattern, and if the match is successful, identify the intent type corresponding to the voice navigation command as the intent of continuing the previous text.

[0093] Optionally, the point of interest slot update module 404 is further configured to, when matching the voice navigation text with a preset sentence pattern, if the match fails, identify an intent type corresponding to the voice navigation text through a first deep learning model.

[0094] Optionally, any sample in the first continuation dataset for training the first deep learning model includes a refined intent text and a label; the content of the refined intent text is used to express the user's intent to refine the description of the features of the point of interest, and the label is used to classify the intent type.

[0095] Optionally, the intent recognition module 402 is further configured to determine whether the type of the previous round of dialogue of the non-first-round voice navigation dialogue is a navigation dialogue. The point of interest slot update module 404 is specifically configured to, when the intent type is the intent of continuing the previous text and the type of the previous round of dialogue is a navigation dialogue, determine a target slot value for identifying the point of interest according to the non-first-round voice navigation dialogue and the previous round of dialogue.

[0096] Optionally, the point of interest slot updating module 404 is specifically configured to obtain a first slot value, where the first slot value is a slot value used to identify a point of interest determined by the at least one round of historical voice navigation conversations; extract a second slot value used to identify a point of interest from the voice navigation instruction; determine a target slot value according to the first slot value and the second slot value; or input the voice navigation text corresponding to the voice navigation command and the voice navigation texts input in at least one round of historical voice navigation conversations before it into a second deep learning model, and obtain the target slot value output by the second deep learning model.

[0097] Optionally, the input data of any sample in the second continuation dataset for training the second deep learning model includes multiple rounds of voice navigation texts, and the output data includes the slot values used to identify points of interest extracted from the multiple rounds of voice navigation texts.

[0098] Optionally, when it fails to determine the target slot value according to the first slot value and the second slot value, the point of interest slot updating module 404 is specifically configured to input the voice navigation text corresponding to the voice navigation command and the voice navigation texts input in at least one round of historical voice navigation conversations before it into the second deep learning model, and obtain the target slot value output by the second deep learning model.

[0099] Optionally, the second deep learning model further outputs a merged text, where the merged text is determined according to the voice navigation text corresponding to the voice navigation command and the voice navigation texts input in at least one round of historical voice navigation conversations before it. The apparatus further includes a data saving module, configured to obtain and save the merged text and the target slot value output by the second deep learning model, where the merged text and the target slot value are used to optimize the second deep learning model.

[0100] For the implementation processes of the functions and effects of each module in the above apparatus, please refer to the implementation processes of the corresponding steps in the above method for details, which will not be elaborated here.

[0101] For the apparatus embodiments, since they basically correspond to the method embodiments, please refer to the partial descriptions of the method embodiments for the relevant parts. The apparatus embodiments described above are only illustrative. The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place, or may be distributed to multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this specification. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0102] The present specification also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of any of the foregoing methods for determining an interest point from a multi-turn conversation provided by the present application are implemented.

[0103] Specifically, computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks.

[0104] The present specification also provides a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of any of the foregoing methods for determining an interest point from a multi-turn conversation are implemented.

Claims

1. A method for determining points of interest from multiple rounds of conversations, characterized in that, The method includes: In response to receiving a voice navigation command for a non-first-round voice navigation dialogue input, identifying an intent type corresponding to the voice navigation command; In the case where the intent type is an intent to continue the previous text, determining a target slot value for identifying a point of interest according to the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogues, and updating the current slot value of the point of interest slot to the target slot value; the updated point of interest slot is used to filter out corresponding target points of interest from the points of interest on the map in response to the non-first-round voice navigation dialogue.

2. The method according to claim 1, wherein After the step of responding to receiving a voice navigation command for a non-first-round voice navigation dialogue input, the method further includes: Extracting an associated slot value associated with a point of interest from the voice navigation command and filling it into a corresponding associated slot; and / or, obtaining a historical slot associated with a point of interest determined in the at least one previous round of historical voice navigation dialogues; Wherein, the associated slot and / or the historical slot is used to filter out corresponding target points of interest from the points of interest on the map in response to the non-first-round voice navigation dialogue.

3. The method according to claim 1, wherein The identifying an intent type corresponding to the voice navigation command includes: Converting the voice of the voice navigation command into voice navigation text; Identifying an intent type corresponding to the voice navigation text through a first deep learning model; or, matching the voice navigation text with a preset sentence pattern, and if the match is successful, identifying the intent type corresponding to the voice navigation command as an intent to continue the previous text.

4. The method according to claim 3, wherein The method further includes: In the case of matching the voice navigation text with a preset sentence pattern, if the match fails, identifying an intent type corresponding to the voice navigation text through a first deep learning model.

5. The method according to claim 3 or 4, characterized in that, Any sample in the first continuation dataset for training the first deep learning model includes a refined intent text and a label; the content of the refined intent text is used to express the intent of the user to have a refined description of the features of a point of interest, and the label is used to classify the intent type.

6. The method according to claim 1, wherein: The method further includes: determining whether the type of the previous dialogue of the non-first-round voice navigation dialogue is a navigation dialogue; In the case where the intent type is an intent to continue the previous text, determining a target slot value for identifying a point of interest according to the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogues includes: in the case where the intent type is an intent to continue the previous text and the type of the previous dialogue is a navigation dialogue, determining a target slot value for identifying a point of interest according to the non-first-round voice navigation dialogue and the previous dialogue.

7. The method according to claim 1, wherein The determining a target slot value for identifying a point of interest according to the non-first-round voice navigation dialogue and at least one previous round of historical voice navigation dialogues includes: Obtaining a first slot value, where the first slot value is a slot value determined by the at least one previous round of historical voice navigation dialogues for identifying a point of interest; extracting a second slot value for identifying a point of interest from the voice navigation instruction; determining a target slot value according to the first slot value and the second slot value; Or, Input the speech navigation text corresponding to the speech navigation command and the speech navigation texts input in at least one previous round of historical speech navigation conversations into a second deep learning model, and obtain the target slot value output by the second deep learning model.

8. The method according to claim 7, wherein The input data of any sample in the second continuation dataset for training the second deep learning model includes multi-round speech navigation texts, and the output data includes the slot values for identifying points of interest extracted from the multi-round speech navigation texts.

9. The method according to claim 7, wherein The method further includes: In the case where determining the target slot value based on the first slot value and the second slot value fails, input the speech navigation text corresponding to the speech navigation command and the speech navigation texts input in at least one previous round of historical speech navigation conversations into the second deep learning model, and obtain the target slot value output by the second deep learning model.

10. The method according to any one of claims 7-9, characterized in that The second deep learning model also outputs a merged text, which is determined based on the speech navigation text corresponding to the speech navigation command and the speech navigation texts input in at least one previous round of historical speech navigation conversations. The method further includes: Obtain and save the merged text and the target slot value output by the second deep learning model, and the merged text and the target slot value are used to optimize the second deep learning model.

11. An apparatus for determining points of interest from multiple rounds of conversations, characterized in that, The device includes: An intent recognition module, configured to recognize the intent type corresponding to the speech navigation command in response to receiving a speech navigation command input in a non-first-round speech navigation conversation; An interest point slot update module, configured to, when the intent type is a continue previous intent, determine a target slot value for identifying a point of interest according to the non-first-round speech navigation conversation and at least one previous round of historical speech navigation conversations, and update the current slot value of the interest point slot to the target slot value; the updated interest point slot is used to filter out corresponding target points of interest from the points of interest on the map in response to the non-first-round speech navigation conversation.

12. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method according to any one of claims 1-10.

Images