Multi-round dialogue intention classification method and device

Abstract

The invention provides a multi-round dialogue intention classification method and device, and the method comprises the steps: defining and extracting a single-valued slot which is a core keyword contained in an FAQ intention; constructing intention clusters based on the single-valued slots, wherein the intention clusters of the same entity type and the intention clusters of the same single-valued slots are constructed; a single-valued slot inheriting slot mechanism is introduced into multiple rounds of dialogues, and different dynamic score adding strategies are performed on the intention clusters of the same entity type and the intention clusters with the same single-valued slot based on the recognition result that the hit intention has the single-valued slot; and performing intention classification again according to the intention score after score addition. The intention recognition accuracy and effectiveness can be improved.

Description

Multi-turn Dialogue Intent Classification Method and Device Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method and apparatus for classifying multi-turn dialogue intents. Background Technology

[0002] Multi-turn dialogue systems have been widely researched and applied in the field of Natural Language Processing (NLP), especially with the enhancement of their performance through deep learning models. The BERT model, proposed by Devlin et al. in 2018, aims to improve language understanding capabilities through bidirectional contextual understanding. Its pre-training process allows for fine-tuning in downstream tasks, adapting it to various language understanding tasks, including intent recognition and slot filling.

[0003] In recent years, research on multi-turn dialogues has deepened. For example, the multi-task BERT model proposed by Zhang et al. focuses on pattern-guided dialogue state tracking, improving the effectiveness of dialogue systems by processing multiple tasks simultaneously. Furthermore, Liu and Xu explored how to explicitly enrich word information to improve the accuracy of intent recognition and slot filling. Smith and Brown proposed the SLIM model, which achieves joint intent detection and slot filling through explicit slot-intent mapping, emphasizing its application in spoken language understanding.

[0004] In the field of intent recognition in multi-turn dialogues, although many studies have provided theoretical and practical foundations for improving system performance, some technical shortcomings still exist, especially in slot and intent recognition, which are summarized as follows.

[0005] Insufficient contextual understanding: Current models often fail to effectively capture and maintain long-term contextual information when processing multi-turn dialogues. As the number of dialogue turns increases, the model is prone to forgetting early information or confusing user intent, which directly affects the accuracy of intent recognition.

[0006] Insufficient semantic similarity discrimination ability: Existing models have limited discrimination ability when dealing with semantically similar expressions but different actual intentions. Although the multi-task BERT model proposed by Zhang et al. enhances dialogue state tracking, it still falls short when faced with complex intentions and cannot effectively handle the problem of intention recognition for specific entities.

[0007] Insufficient optimization in FAQ type intent recognition: Existing models often fail to effectively understand different aspects of the same topic when dealing with FAQ type questions. Although some research has attempted to improve recognition capabilities by enhancing contextual understanding, further optimization is still needed to improve the system's performance in complex scenarios. Summary of the Invention

[0008] In view of this, this application provides a method and apparatus for classifying multi-turn dialogue intents, the main purpose of which is to improve the accuracy and effectiveness of intent recognition.

[0009] According to one aspect of this application, a multi-turn dialogue intent classification method is provided, comprising: defining and extracting single-value slots, wherein the single-value slots are core keywords contained in FAQ intents; constructing intent clusters based on single-value slots, including constructing intent clusters of the same entity type and intent clusters with the same single-value slots; introducing a single-value slot inheritance mechanism in multi-turn dialogues, and applying different dynamic scoring strategies to intent clusters of the same entity type and intent clusters with the same single-value slots based on the recognition results of intents with single-value slots; and reclassifying intents based on the scored intents.

[0010] In one implementation, the method for extracting the single-value slot includes: identifying core keywords commonly contained in the standard FAQ questions and similar questions; verifying that the keyword has only a single value in the intent; and marking the keywords that meet the conditions as single-value slots.

[0011] In one implementation, the method for constructing the intent cluster of the same entity type includes: identifying intents with the same sentence structure; extracting replaceable entity keywords from different intents; and grouping intents with the same entity type but different specific values ​​into the same cluster.

[0012] In one implementation, the method for constructing intent clusters with the same single-value slots includes: traversing all single-value slots in the intent library; clustering different intents that share the same single-value slot keyword; and establishing a corresponding intent mapping table for each single-value slot.

[0013] In one implementation, the dynamic scoring strategy for intent clusters with the same single-value slots includes: when the score of the Top 1 intent in the current round is lower than a threshold, retrieving the single-value slot inherited from the previous round; and applying a tiered scoring to all intents containing that single-value slot.

[0014] In one implementation, the score of the tiered bonus is positively correlated with one of the following factors: the confidence level of the single-value slot in the previous round; the time interval between the current round and the previous round; and the frequency of the slot's occurrence in historical dialogues.

[0015] In one implementation, the dynamic scoring strategy for intent clusters of the same entity type includes: when a single-value slot of the same entity type as the Top1 intent is detected in the user input, matching the intent of the corresponding entity in the intent cluster; and applying a fixed weight to the successfully matched intent.

[0016] In one implementation, the termination condition of the single-value slot inheritance mechanism includes one of the following: no points are awarded in two consecutive rounds of dialogue; a user-defined topic switching keyword is detected; or a new topic inquiry is initiated by the system.

[0017] According to one aspect of this application, a multi-turn dialogue intent classification device is provided, comprising: a single-value slot setting unit, used to define and extract single-value slots, wherein the single-value slots are core keywords contained in FAQ intents; an intent cluster construction unit, used to construct intent clusters based on single-value slots, including constructing intent clusters of the same entity type and intent clusters with the same single-value slots; a dynamic scoring unit, used to introduce a single-value slot inheritance mechanism in multi-turn dialogues, and based on the recognition result that the hit intent has a single-value slot, apply different dynamic scoring strategies to intent clusters of the same entity type and intent clusters with the same single-value slots; and an intent classification execution unit, used to reclassify the intents based on the scored intents.

[0018] By employing the above technical solution, this application provides a method for classifying intents in multi-turn dialogues. First, it defines and extracts single-value slots and constructs intent clusters based on these slots. Then, it introduces a single-value slot inheritance mechanism into the multi-turn dialogue. Based on the recognition results of intents with single-value slots, different dynamic scoring strategies are applied to intent clusters of the same entity type and intent clusters with the same single-value slots. Finally, the intents are reclassified based on the added intent scores. This application's solution has at least the following technical effects: 1. Combining the slot inheritance mechanism in multi-turn dialogues significantly improves the accuracy of intent recognition in complex multi-turn dialogue scenarios. Especially when user expressions are incomplete or highly dependent on context, the system demonstrates superior understanding and robustness, effectively solving the recognition difficulties of traditional methods in continuous dialogues. 2. The specific scoring strategy proposed in this application significantly improves the system's ability to distinguish between semantically similar but different intents, especially when processing intent clusters of the same entity type and FAQ-type intents. This improvement enables the system to more accurately capture subtle differences in user intents, providing more precise and coherent responses. 3. Overall, this application can significantly improve the intelligence and user experience of dialogue systems. In practical applications, the average number of dialogue turns per user is reduced, the task completion rate is increased, and the relevance and naturalness of system responses are significantly enhanced. Furthermore, this application has good scalability and adaptability, and can be effectively applied to dialogue systems in different fields, providing an innovative solution for improving the performance of various intelligent dialogue applications.

[0019] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 shows a flowchart of a multi-turn dialogue intent classification method provided by an embodiment of this application; Figure 2 shows a schematic diagram of an example of a multi-turn dialogue intent classification method provided by an embodiment of this application; Figure 3 shows a schematic diagram of the structure of a multi-turn dialogue intent classification device provided by an embodiment of this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present application can be combined with each other.

[0022] As analyzed above, while existing technologies in the field of intent recognition for multi-turn dialogues have provided theoretical and practical foundations for improving system performance, some technical shortcomings still exist, particularly in slot and intent recognition: 1. Insufficient contextual understanding: Current models often fail to effectively capture and maintain long-term contextual information when processing multi-turn dialogues. As the number of dialogue turns increases, the model easily forgets early information or confuses user intent, directly affecting the accuracy of intent recognition; 2. Insufficient semantic similarity discrimination ability: Existing models have limited discrimination ability when dealing with semantically similar expressions with different actual intents. Although the multi-task BERT model proposed by Zhang et al. enhances dialogue state tracking, it still falls short when facing complex intents and cannot effectively handle the intent recognition problem of specific entities; 3. Insufficient optimization for FAQ-type intent recognition: For FAQ-type questions, existing models often fail to effectively understand different aspects of the same topic. Although existing research has attempted to improve recognition ability by enhancing contextual understanding, further optimization is still needed to improve the system's performance in complex scenarios.

[0023] The main objective of this invention is to provide a multi-turn dialogue intent recognition method based on single-value slot inheritance, aiming to solve the following technical problems: 1. Improve the accuracy of intent recognition in multi-turn dialogue scenarios, especially when user expressions are incomplete or highly dependent on context; 2. Effectively handle intents containing specific entities, improving the system's ability to distinguish semantically similar but actually different expressions; 3. Accurately infer the user's true intent when the user asks follow-up questions continuously or the topic slightly shifts; 4. Provide a flexible intent scoring mechanism to adapt to different types of multi-turn dialogue scenarios; 5. Optimize the recognition of FAQ-type intents, improving the system's ability to understand different aspects of questions under the same topic.

[0024] Referring to Figure 1, a multi-turn dialogue intent classification method provided in an embodiment of this application includes the following steps S101-S104.

[0025] S101: Define and extract single-value slots, where the single-value slots are the core keywords contained in the FAQ intent.

[0026] Identifying core keywords from FAQ intents is crucial for understanding user intent. For example, in "Check Beijing Weather" and "Check Shanghai Weather," "weather" is the core keyword, and in each intent, "weather" corresponds to a single value scenario (i.e., the action of checking the weather in a specific location). Therefore, "weather" can be used as a single-value slot. By defining and extracting these single-value slots, clear anchor points are provided for subsequent intent clustering and score adjustments, improving the accuracy of intent recognition.

[0027] S102: Construct intent clusters based on single-value slots, including constructing intent clusters of the same entity type and intent clusters with the same single-value slots.

[0028] Building intent clusters based on single-value slots categorizes closely related intents and reduces interference from irrelevant intents. For example, intent clusters of the same entity type, such as "Query Beijing Temperature" and "Query Shanghai Temperature," both belong to the "City" entity and are grouped into the same cluster. Intent clusters with the same single-value slot, such as "Query Beijing Weather" and "What's the Weather Like in Beijing?", share the single-value slot "Weather" and are also grouped into the same cluster. Building intent clusters based on single-value slots makes intent classification more logical and facilitates subsequent dynamic adjustments.

[0029] S103: Introduce a single-value slot inheritance mechanism in multi-turn dialogues. Based on the recognition result of the single-value slot in the hit intent, apply different dynamic scoring strategies to intent clusters of the same entity type and intent clusters with the same single-value slot.

[0030] This application introduces a single-value slot inheritance mechanism and dynamic scoring, leveraging the contextual relevance of multi-turn dialogues to inherit single-value slot information from the previous turn. For example, if a user first asks "Beijing weather" (single-value slots "weather" and "Beijing"), and then asks "Shanghai?" in the next turn, the system inherits the "weather" slot and dynamically scores intents with the same entity type (city) such as "query Shanghai weather" and single-value slots containing "weather". Thus, by combining context, ambiguity can be reduced, and the coherence of intent recognition in multi-turn dialogues can be improved.

[0031] S104: Reclassify intent based on the intent score after the bonus points are added.

[0032] In this application, the scores can be re-ranked based on adjustments to select the most likely intent. For example, "check Shanghai weather" receives the highest score after the score adjustment, thus being identified as the intent for this round. This makes the classification results more consistent with the user's true intent and improves the fluency of the conversation.

[0033] In one implementation, the method for extracting single-value slots may specifically include steps (1)-(3). (1) Identify the core keywords commonly contained in the standard and similar questions of the FAQ intent. Find common key information from the standard and similar questions, such as "how to unsubscribe from membership" (standard question) and "method of unsubscribing from membership" (similar question), the common keywords are "unsubscribe" and "member". Thus, the core elements of the intent are locked, laying the foundation for subsequent slot extraction. (2) Verify that the keyword has only a single value in the intent. Verify that the keyword has only a single value in the intent, that is, ensure that the keyword has only one value dimension in the intent. For example, "member" in the above intent only represents the single concept of "member service" and has no other meaning. If the keyword has multiple values ​​(such as "business" in "handling business" can refer to multiple businesses), it does not meet the requirements. Thus, the uniqueness of the single-value slot is guaranteed, and confusion is avoided. (3) Mark the keywords that meet the conditions as single-value slots. Officially determine the keywords that meet the conditions as single-value slots. For example, "unsubscribe" and "member" are marked as single-value slots. By clarifying the slot information, a basis is provided for subsequent clustering and scoring.

[0034] In one implementation, the method for constructing intent clusters of the same entity type includes steps (1)-(3). (1) Identify intents with the same sentence structure. Find intents with consistent structure through sentence analysis, such as "query the address of A" and "query the address of B", both of which are "query + [entity] + address". The effect is to filter out intents with similar expression logic. (2) Extract replaceable entity keywords from different intents. Extracting replaceable entity keywords means finding the entity parts that can be changed in the sentence structure. In the above example, "A" and "B" are replaceable entity keywords (such as "company" and "restaurant"). Thus, the variables in the intent are clarified, providing a basis for clustering. (3) Group intents with the same entity type but different specific values ​​into the same cluster. Grouping into the same cluster means classifying intents with the same entity type but different values. For example, "query the address of company A" and "query the address of company B" both have the entity type "company", so they are grouped into the same cluster. Thus, the aggregation of intents of the same type is realized, which is convenient for subsequent scoring based on entity type.

[0035] In one implementation, the method for constructing intent clusters with the same single-value slot includes steps (1)-(3). (1) Traverse all single-value slots in the intent library. Traversing single-value slots means scanning all single-value slots in the intent library to ensure no omissions. For example, traversing all labeled single-value slots such as "weather", "address", and "member" provides complete slot information for clustering. (2) Cluster different intents that share the same single-value slot keyword. Clustering intents that share the same single-value slot means grouping different intents containing the same single-value slot into one category. For example, "query Beijing weather", "what is the weather like in Beijing", and "tomorrow's weather in Beijing" all contain the "weather" slot and are clustered into one category. Thus, the association between slots and intents is established, which facilitates quick location when adding points later. (3) Establish a corresponding intent mapping table for each single-value slot. Establishing an intent mapping table means creating a corresponding intent list for each single-value slot. For example, the mapping table for the "weather" slot contains all the above-mentioned related intents. This improves query efficiency and allows for quick identification of the target intent when dynamically awarding points.

[0036] In one implementation, the dynamic scoring strategy for intent clusters with identical single-value slots includes: when the score of the top intent in the current round is below a threshold, retrieving the inherited single-value slot from the previous round; and applying a tiered scoring system to all intents containing that single-value slot. Specifically, determining the score of the top intent in the current round involves triggering the scoring mechanism when the credibility of the highest-scoring intent is insufficient (below the threshold). For example, if the highest-scoring intent in the current round, "Query Shanghai Traffic," has a score of 0.6 (threshold 0.7), the strategy is triggered. This avoids misjudgments due to insufficient scores. Retrieving the inherited single-value slot from the previous round involves calling the slot information retained from the previous round. For example, if the intent in the previous round was "Query Beijing Weather," the inherited slot is "Weather." This utilizes contextual information to enhance the relevance of intent recognition in the current round. The tiered scoring system awards points to related intents based on their degree of relevance. For example, intents containing the "Weather" slot, such as "Query Shanghai Weather" and "How is the Weather in Shanghai?", are scored by 0.2 and 0.15 respectively, based on their degree of relevance. This increases the score of relevant intents, making them more likely to be selected and improving accuracy.

[0037] In one implementation, the tiered scoring is positively correlated with one of the following factors: the confidence level of the previous round's single-value slot; the time interval between the current round and the previous round; and the frequency of the slot's occurrence in historical dialogues. Higher confidence levels result in more points. For example, if the confidence level for the "weather" slot in the previous round was 0.9 (high), then an additional 0.2 points would be awarded for intents containing "weather" in this round; if the confidence level was 0.6 (low), an additional 0.1 points would be awarded. Introducing confidence levels makes the scoring more reasonable in aligning with the slot's reliability. Shorter time intervals between the current and previous rounds result in more points. For example, a 1-minute interval awards 0.1 more points to the relevant intent than a 10-minute interval. This reflects the immediate relevance of the dialogue; closer intervals indicate greater contextual influence. The more frequently the slot appears in historical dialogues, the more points are awarded. For example, a "weather" slot appearing 3 times in historical dialogues awards 0.15 more points than appearing once. By highlighting slots that users frequently pay attention to, the priority of corresponding intentions can be increased.

[0038] In one implementation, a dynamic scoring strategy is executed for intent clusters of the same entity type. This includes: when a single-value slot of the same entity type as the Top 1 intent is detected in the user input, the intent corresponding to the entity is matched within the intent cluster; and a fixed weight is added to the successfully matched intent. Detecting single-value slots of the same entity type involves identifying slots in the user input that match the entity type of the current Top 1 intent. For example, if the current Top 1 intent is "Query Beijing temperature" (entity type "city"), and the user inputs "What about Shanghai?", where "Shanghai" is an entity type of "city," this captures the user's continued need for the same type of entity. Matching the corresponding entity intent involves finding intents containing that entity within the same entity type cluster. For example, in the "City temperature query" cluster, "Query Shanghai temperature" is matched. This quickly locates the specific intent matching the user input. The fixed weight adds a fixed score (e.g., 0.3) to the successfully matched intent. For example, "Query Shanghai temperature" originally scored 0.5, and after the weighting, it becomes 0.8. This significantly improves the score of the target intent, ensuring its correct identification.

[0039] In one implementation, the slot inheritance mechanism includes: retaining single-value slots identified in the previous round of dialogue; considering inherited slot information when identifying intent in the current round; and releasing inherited slots if no points are awarded in the current round. Retaining single-value slots from the previous round involves temporarily storing them as a reference for the current round. For example, if the previous round's slot was "Member," this information is retained in the current round. This ensures contextual continuity and avoids information loss. Considering inherited slots in the current round involves analyzing inherited slots when identifying the current intent. For example, if a user says "How to cancel," the intent to "cancel membership" is prioritized based on the inherited "Member" slot. This utilizes historical information to assist in current identification, improving accuracy. Releasing slots if no points are awarded means that if no points are awarded in the current round, the slot is considered irrelevant and is no longer retained. For example, if the previous round's slot was "Member," and the current round's intent is "Check the weather" without any points awarded, the "Member" slot is released. This promptly removes invalid information and avoids interfering with subsequent dialogue.

[0040] In one implementation, the termination condition for the single-value slot inheritance mechanism includes one of the following: no points are awarded in two consecutive rounds of dialogue; a user explicitly indicates a topic-switching keyword is detected; or the system proactively initiates a new topic inquiry. Two consecutive rounds without awarding points means that no points are awarded based on the inherited slot in two consecutive rounds, indicating that the slot is meaningless, and inheritance terminates. For example, after inheriting the "weather" slot, if two rounds of dialogue do not involve weather-related intentions and no points are awarded, then inheritance terminates. This maintains system efficiency by automatically filtering outdated information. Detecting topic-switching keywords means that the user explicitly indicates a topic change (e.g., "Let's not talk about this anymore," "Let's change the topic"), and inheritance terminates. This allows for rapid response to user topic changes and avoids incorrect associations. The system proactively initiates a new topic means that the system proactively inquires about new content (e.g., "Do you need to query other information?"), and the previous slot inheritance terminates. This ensures that the new topic led by the system is not interfered with by the old slot, maintaining smooth dialogue.

[0041] The following example illustrates the multi-turn dialogue intent recognition method based on single-value slots and intent clusters proposed in this application. Referring to Figure 2, this example includes the following key steps.

[0042] S21: Single-value slot definition and extraction. 1) A keyword contained in the intent (usually a FAQ intent). This keyword is a core keyword in the corresponding wording (standard questions and similar questions) of the FAQ intent, and the related expressions all revolve around this keyword. This keyword is generally a certain type of entity, but it only exists in one value in the intent. 2) Based on the business scenario, extract the single-value slot from the intent library according to the above definition.

[0043] S22: Intent cluster construction. In this embodiment of the application, two types of intent clusters can be constructed based on single-value slots: 1) Intent clusters of the same entity type: intent sets with the same sentence structure but different keywords; 2) Intent clusters of the same single-value slots: different intent sets with the same single-value slots.

[0044] S23: Bonus scoring strategy for intent clusters with identical single-value slots. Intent bonus scoring principles may include: 1) If a single-value slot was identified in the previous round, that slot may be inherited in this round; 2) If the top intent score in this round of dialogue does not reach the threshold, all intents in the intent cluster corresponding to the single-value slot will receive bonus points. After bonus scoring, processing continues according to existing NLU logic; 3) Bonus scores can be dynamically adjusted based on dialogue history. After bonus scoring, processing continues according to the original NLU logic; 4) If no bonus is awarded in the second round, the inherited slot is released, meaning the slot inheritance from the previous round is no longer retained; if a bonus is awarded, the slot is retained for the next round of dialogue. This multi-round dialogue intent bonus scoring method aims to improve the system's ability to identify entity-related intents in multi-round dialogues through slot inheritance and bonus scoring mechanisms. The goal is to award bonuses to relevant intents by inheriting key information from slots when scores are low, enabling the system to more accurately identify the user's true needs and ultimately improve the overall response accuracy.

[0045] S24: Bonus scoring strategy for intent clusters of the same entity type. For single-value FAQ intents with the same question format, the system extracts different entity members and assigns them to the same entity, forming intent clusters of the same entity type. During NLU recognition, the specific bonus scoring strategy is as follows: 1) When the user's input matches an intent (i.e., the Top 1 intent), the system first determines if the intent has a single-value slot; 2) If the Top 1 intent has a single-value slot, the system checks if there is an intent cluster of the same entity type for that slot. If so, and the user's input contains a single-value slot of the same entity, the system finds the intent corresponding to the single-value slot in the user's input within the intent cluster and awards points. This bonus scoring logic only applies within the current round and does not affect subsequent dialogues.

[0046] Therefore, this example improves the accuracy of intent recognition in single-turn dialogues by configuring associated entities for single-value FAQ intents with the same question format and determining the user's true needs through entity member matching. The system assigns points to relevant intents based on the entity member information in the user's input.

[0047] In summary, this application provides a method for classifying intents in multi-turn dialogues. First, it defines and extracts single-value slots and constructs intent clusters based on these slots. Then, it introduces a single-value slot inheritance mechanism into the multi-turn dialogue. Based on the recognition results of intents containing single-value slots, different dynamic scoring strategies are applied to intent clusters of the same entity type and intent clusters with the same single-value slots. Finally, the intents are reclassified based on the scores after the scoring. This application's solution has at least the following technical effects.

[0048] 1. By combining slot inheritance mechanisms in multi-turn dialogues, the accuracy of intent recognition in complex multi-turn dialogue scenarios can be significantly improved. Especially when user expressions are incomplete or highly dependent on context, the system demonstrates superior understanding and robustness, effectively solving the recognition difficulties of traditional methods in continuous dialogues.

[0049] 2. The specific scoring strategy proposed in this application can significantly improve the system's ability to distinguish between semantically similar but different intentions, especially when dealing with intent clusters of the same entity type and FAQ-type intents. This improvement enables the system to more accurately capture subtle differences in user intent and provide more precise and coherent responses.

[0050] 3. Overall, this application can significantly improve the intelligence and user experience of dialogue systems. In practical applications, the average number of dialogue turns per user is reduced, the task completion rate is increased, and the relevance and naturalness of system responses are significantly enhanced. Furthermore, this application has good scalability and adaptability, and can be effectively applied to dialogue systems in different fields, providing an innovative solution for improving the performance of various intelligent dialogue applications.

[0051] Corresponding to the above method, this application also provides a multi-turn dialogue intent classification device, as shown in Figure 3, including: a single-value slot setting unit 301, used to define and extract single-value slots, wherein the single-value slots are core keywords contained in the FAQ intent; an intent cluster construction unit 302, used to construct intent clusters based on single-value slots, including constructing intent clusters of the same entity type and intent clusters with the same single-value slots; a dynamic scoring unit 303, used to introduce a single-value slot inheritance mechanism in multi-turn dialogue, and based on the recognition result that the hit intent has a single-value slot, apply different dynamic scoring strategies to intent clusters of the same entity type and intent clusters with the same single-value slots; and an intent classification execution unit 304, used to reclassify the intent based on the scored intent scores.

[0052] For details on the implementation of the multi-turn dialogue intent classification device, please refer to the description of the aforementioned method embodiments, which will not be repeated here.

[0053] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0054] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0055] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0056] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0057] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0058] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0059] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0060] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for classifying multi-turn dialogue intent, characterized in that, include: Define and extract single-value slots, where the single-value slots are the core keywords contained in the FAQ intent; Intent clusters are constructed based on single-value slots, including constructing intent clusters of the same entity type and intent clusters with the same single-value slots; a single-value slot inheritance mechanism is introduced in multi-turn dialogues, and based on the recognition results of the hit intent having a single-value slot, different dynamic scoring strategies are applied to intent clusters of the same entity type and intent clusters with the same single-value slots; and the intent is reclassified based on the added intent scores.

2. The method according to claim 1, characterized in that, The method for extracting single-value slots includes: identifying core keywords commonly contained in the standard FAQ questions and similar questions; verifying that the keyword has only a single value in the intent; and marking keywords that meet the conditions as single-value slots.

3. The method according to claim 1, characterized in that, The method for constructing intent clusters of the same entity type includes: identifying intents with the same sentence structure; extracting replaceable entity keywords from different intents; and grouping intents with the same entity type but different specific values ​​into the same cluster.

4. The method according to claim 1, characterized in that, The method for constructing intent clusters with the same single-value slots includes: traversing all single-value slots in the intent library; clustering different intents that share the same single-value slot keyword; and establishing a corresponding intent mapping table for each single-value slot.

5. The method according to claim 1, characterized in that, The dynamic scoring strategy for intent clusters with the same single-value slots includes: when the score of the Top 1 intent in the current round is lower than the threshold, retrieving the single-value slot inherited from the previous round; and applying a tiered scoring to all intents containing that single-value slot.

6. The method according to claim 5, characterized in that, The tiered scoring is positively correlated with one of the following factors: the confidence level of the single-value slot in the previous round; the time interval between the current round and the previous round; and the frequency of the slot's occurrence in historical dialogues.

7. The method according to claim 1, characterized in that, The dynamic scoring strategy for intent clusters of the same entity type includes: when a single-value slot of the same entity type as the Top1 intent is detected in the user input, the intent of the corresponding entity is matched in the intent cluster; and a fixed weight is added to the intent that is successfully matched.

8. The method according to claim 1 or 7, characterized in that, The termination conditions of the single-value slot inheritance mechanism include one of the following: no points are awarded in two consecutive rounds of dialogue; a user-defined topic switching keyword is detected; or a new topic inquiry is initiated by the system.

9. A multi-turn dialogue intent classification device, characterized in that, include: A single-value slot setting unit is used to define and extract single-value slots, wherein the single-value slots are the core keywords contained in the FAQ intent; The intent cluster construction unit is used to construct intent clusters based on single-value slots, including constructing intent clusters of the same entity type and intent clusters with the same single-value slots; the dynamic scoring unit is used to introduce a single-value slot inheritance mechanism in multi-turn dialogues, and based on the recognition result that the hit intent has a single-value slot, apply different dynamic scoring strategies to intent clusters of the same entity type and intent clusters with the same single-value slots respectively. The intent classification execution unit is used to reclassify intents based on the added intent scores.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.

Images