Intention recognition-based dialogue method, system and electronic device

Abstract

The application provides a conversation method and system based on intention recognition and an electronic device, and relates to the technical field of artificial intelligence conversation. The method comprises the following steps: obtaining to-be-processed data of the current round of conversation, wherein the to-be-processed data comprises a conversation sentence input by a user and structured data generated based on an interaction event; performing standardized processing on the to-be-processed data to obtain a first slot set; performing intention recognition and slot label prediction based on the to-be-processed data and the first slot set by using a pre-trained joint learning architecture to generate an initial conversation state of the current round of conversation; performing slot fusion on the initial conversation state according to a historical conversation state of a historical round to obtain a target conversation state; performing completeness detection on the target conversation state and generating a conversation instruction based on a detection result; and responding to the current round of conversation according to the conversation instruction. The present scheme can efficiently and accurately convert the fragmented expression of a user into a business request that can directly call a backend service, thereby improving the human-computer interaction experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence dialogue technology, and in particular to a dialogue method, system and electronic device based on intent recognition. Background Technology

[0002] In corporate travel management and intelligent itinerary assistant scenarios, users tend to use highly compressed, fragmented, and unstructured natural language to express complex needs that include multiple dimensions such as time, category, and location, in order to pursue interaction efficiency. However, existing interaction methods are difficult to adapt to such high-information-density input patterns. For example, traditional form-based interactions force users to follow fixed steps, which contradicts the flexibility of natural language; while traditional dialogue systems based on keyword matching or fixed word order often fail to accurately capture core elements when faced with disordered arrangement of key information flows, resulting in low recognition rates.

[0003] Therefore, current interaction methods make it difficult to transform fragmented, multi-dimensional information into structured business requests. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a dialogue method, system and electronic device based on intent recognition, which can efficiently and accurately convert fragmented user expressions into business requests that can directly call backend services, thereby improving the human-computer interaction experience.

[0005] In a first aspect, embodiments of the present invention provide a dialogue method based on intent recognition, the method comprising: Obtain the pending data for this round of dialogue, which includes: dialogue statements input by the user and structured data generated based on interaction events; The data to be processed is normalized to obtain a first set of slots; wherein, the normalization process includes at least: time parsing, extraction of domain entities and fusion of heterogeneous data; The pre-trained joint learning architecture performs intent recognition and slot label prediction based on the data to be processed and the first slot set, generating the initial dialogue state for this round of dialogue; The initial dialogue state is slot-fused based on the historical dialogue state of the historical rounds to obtain the target dialogue state; Perform a completeness check on the target dialogue state and generate dialogue instructions based on the check results; Respond to this round of dialogue according to the dialogue instructions.

[0006] Optionally, the normalization process of the data to be processed to obtain the first slot set includes: The time in the dialogue statement is parsed to obtain time data in a preset time format; The pre-trained entity extraction model extracts domain entities in random order based on the semantics of each domain entity in the dialogue statement to obtain a list of key information fragments. The time data, the list of key information fragments, and the structured data are merged into a first slot set.

[0007] Optionally, the pre-trained joint learning architecture performs intent recognition and slot label prediction based on the data to be processed and the first slot set to generate the initial dialogue state for this round of dialogue, including: The data to be processed and the first set of slots are input into a pre-trained joint learning architecture; wherein, the joint learning architecture includes: an encoder, which shares the intent task head and the slot task head of the encoder, and both the intent task head and the slot task head use a multi-head attention mechanism; The encoder extracts dialogue data features from the data to be processed and the first slot set; The intent task head, employing a multi-head attention mechanism, performs intent recognition on the dialogue data features to obtain the target intent label; The slot task head using a multi-head attention mechanism predicts the slot labels for the data features, resulting in a second set of slots and its corresponding slot label sequence. An initial dialogue state is generated based on the target intent tag, the first slot set, the second slot set, and the slot tag sequence.

[0008] Optionally, the step of performing slot fusion on the initial dialogue state based on the historical dialogue states of historical rounds to obtain the target dialogue state includes: If the timestamp of the time slot in the initial dialogue state does not match the timestamp of the time slot in the historical dialogue state, the dialogue state is updated by overwriting the timestamp of the time slot in the initial dialogue state. For the first slot missing in the initial dialogue state, the first slot and its slot label are inherited from the historical dialogue state; Based on the historical dialogue state, the pronouns in the initial dialogue state are resolved by dereferencing; After completing the above slot fusion, the target object state is obtained; wherein, the target object state includes: target awareness tag, current round slot set, and current round slot tag set.

[0009] Optionally, the step of performing a completeness check on the target dialogue state and generating dialogue instructions based on the check result includes: Invoke a preset business rule knowledge base; wherein, the business rule knowledge base is used to record related intent tags and mandatory slot sets; Obtain the set of target mandatory slots associated with the target intent tag in the target dialogue state from the business rule knowledge base; The completeness of the current slot set is checked based on the target forced slot set; If at least one mandatory slot is detected to be missing in the current slot set, the missing mandatory slot will be filled with slot labels based on the user profile database and corporate travel standard information. The slots that fail to fill are designated as the second slots, and the second slots are sorted according to the priority of business logic. Generate an inquiry command for each of the second slots based on the sorting results, to inquire about the slot label.

[0010] Optionally, the method further includes: If there is a slot with a semantically ambiguous slot label in the current round slot set in the target dialogue state, a clarification instruction is generated to clarify the slot label.

[0011] Optionally, responding to the current dialogue based on the dialogue instructions includes: The natural language model generates dialogue statements to respond to the current round of dialogue based on the dialogue instructions and preset dialogue templates.

[0012] Optionally, the method further includes: If the completeness check of the target dialogue state passes, a structured interface request is constructed for the target dialogue state to call the backend service, and the response to the interface request is displayed in the form of a rich media card.

[0013] Secondly, the present invention provides a dialogue system based on intent recognition, the system comprising: Dialogue data acquisition module: used to acquire the data to be processed in the current dialogue, including: dialogue statements input by the user and structured data generated based on interaction events; Data processing module: used to perform normalization processing on the data to be processed to obtain a first slot set; wherein, the normalization processing includes at least: time parsing, extraction of domain entities and fusion of heterogeneous data; Model processing module: used by a pre-trained joint learning architecture to perform intent recognition and slot label prediction based on the data to be processed and the first slot set, to generate the initial dialogue state of this round of dialogue; Slot fusion module: used to fuse the initial dialogue state with slots based on the historical dialogue states of historical rounds to obtain the target dialogue state; Instruction generation module: used to perform completeness detection on the target dialogue state and generate dialogue instructions based on the detection results; Dialogue Response Module: Used to respond to the current dialogue based on the dialogue instructions.

[0014] Thirdly, embodiments of the present invention also provide an electronic device, which includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the steps of the method provided in the first aspect.

[0015] This invention provides a dialogue method, system, and electronic device based on intent recognition. First, by processing data containing dialogue statements and structured data, and combining standardized processing such as time parsing, domain entity extraction, and heterogeneous data fusion, fragmented information originally scattered in natural language text and interaction events can be uniformly transformed into a first set of slots that is understandable and easy for machines to process. Then, a joint learning architecture is used to simultaneously perform intent recognition and slot label prediction, avoiding information loss due to error propagation in traditional pipelined processing. By utilizing the dual features of the data to be processed and the first set of slots, the joint learning architecture can more accurately understand the user's true intent in a specific context and directly output an initial dialogue state with the target intent label, thereby significantly improving the efficiency and accuracy of the conversion from unstructured input to a semi-structured state. Secondly, by introducing historical dialogue states to perform slot fusion on the initial dialogue state, better information inheritance and completion across rounds can be achieved, effectively alleviating the problem of incomplete or ambiguous information in single-round dialogues, allowing the target dialogue state to accumulate and integrate fragmented information from multiple rounds of interaction. The resulting target dialogue state not only includes the explicit requirements of the current round but also incorporates implicit constraints from historical interactions, improving accuracy and completeness. Based on this, dialogue instructions generated from the completeness check results and responses guide users to complete all the information required for their business. Overall, this solution efficiently and accurately transforms fragmented user expressions into business requests that can directly invoke backend services, thereby improving the completeness and accuracy of dialogue information conversion in human-computer dialogue scenarios and enhancing the human-computer interaction experience.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an intent-based dialogue method provided in an embodiment of the present invention; Figure 2 A flowchart of step S103 in a dialogue method based on intent recognition provided in an embodiment of the present invention; Figure 3 A flowchart of step S104 in a dialogue method based on intent recognition provided in an embodiment of the present invention; Figure 4 A flowchart of step S105 in a dialogue method based on intent recognition provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a dialogue system based on intent recognition provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0020] icon: 100 - Dialogue Data Acquisition Module; 200 - Data Processing Module; 300 - Model Processing Module; 400 - Slot Fusion Module; 500 - Command Generation Module; 600 - Dialogue Response Module; 101 - Processor; 102 - Memory; 103 - Bus; 104 - Communication interface. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In corporate travel management and intelligent itinerary assistant scenarios, existing interaction methods struggle to adapt to user input patterns. Traditional form-based interactions force users to follow fixed steps, contradicting the flexibility of natural language. Traditional dialogue systems based on keyword matching or fixed word order often fail to accurately capture core elements when faced with disordered flow of key information, resulting in low recognition rates.

[0023] For intelligent dialogue and slot filling technologies, the following methods are mainly used: Sequence labeling models (such as BiLSTM-CRF) are used to extract named entities, but they often treat intent recognition and slot filling as independent tasks, lacking joint optimization, which leads to semantic conflicts between intent and slots in complex sentences.

[0024] Static slot configuration predefines all slots required for all intents. When a critical slot is missing, it can only mechanically query in a fixed order, lacking flexibility and context awareness.

[0025] Dialogue state tracking lacks the ability to deeply correlate and dynamically maintain cross-dimensional information (such as time, location, and category), making it difficult to handle user modifications or referencing of information in a certain dimension during multiple rounds of dialogue. For example, if a user first says "book a hotel in Beijing" and then says "I want an executive room," the system cannot effectively inherit the location slot for "Beijing."

[0026] Therefore, current interaction methods suffer from rigidity and inefficiency in processing multimodal inputs. When key information is missing, the accuracy of intent recognition is poor, and it is unable to efficiently generate dynamic questioning logic that conforms to human intuition. This hinders the efficient and intelligent transformation from fragmented, multi-dimensional information in natural language to structured business requests.

[0027] In this context, embodiments of the present invention provide a dialogue method, system, and electronic device based on intent recognition. This solution can be applied to many scenarios such as natural language processing, intelligent dialogue systems, human-computer interaction, and travel booking.

[0028] To facilitate understanding of this embodiment, a dialogue method based on intent recognition disclosed in this embodiment of the invention will first be introduced, such as... Figure 1 As shown, the method includes: Step S101: Obtain the data to be processed in this round of dialogue. The data to be processed includes: dialogue statements input by the user and structured data generated based on interaction events.

[0029] In this embodiment, the user-inputted dialogue statements in the data to be processed are natural language text. The user's input can be monitored in real time through dialogue interfaces such as web chat windows, app voice input boxes, or API interfaces to obtain the dialogue statements in natural language text form. For example, in a travel booking application, the user inputs the text: "Book me a flight to Beijing tomorrow." The obtained dialogue statement is the string: "Book me a flight to Beijing tomorrow."

[0030] In this embodiment, the structured data generated based on interactive events is obtained using the following example. Front-end interfaces of devices such as mobile phones and computers contain various interactive components, such as date pickers, location pickers, and product type radio buttons. When users click, swipe, or select on these components, the front-end generates structured data with a specific schema, rather than natural language text.

[0031] Based on this, this embodiment can listen to user operation events and serialize them into standard JSON format data packets. These data packets can carry deterministic information representing the user's intent without requiring complex semantic understanding.

[0032] One example scenario is that, while the user is inputting the above dialogue statement, or in the previous round of dialogue, the user selected "2025-03-28" by clicking the calendar control on the interface. In this scenario, the JSON data is structured data generated based on the interaction event.

[0033] By aggregating dialogue statements and structured data, the data to be processed in this round of dialogue can be formed, thereby capturing both explicit text commands and implicit interface operations from the user.

[0034] Step S102: Normalize the data to be processed to obtain the first slot set; wherein, the normalization process includes at least: time parsing, extraction of domain entities and fusion of heterogeneous data.

[0035] After acquiring the data to be processed, in order to eliminate the randomness of user input, such as colloquial expressions, ambiguous time descriptions, and format differences between different modal data, this embodiment can perform standardization processing on the data to be processed, so as to uniformly convert unstructured, natural language dialogue statements and heterogeneous structured data into a standardized slot data format.

[0036] In this embodiment, the normalization process for the data to be processed may include steps such as time parsing, extraction of domain entities, and fusion of heterogeneous data, as shown below.

[0037] Time parsing steps. Parse the time in the dialogue statements to obtain time data in a preset time format.

[0038] In one specific embodiment, time segments in dialogue statements can be identified, and relative or fuzzy times can be mapped to time data in a standard time format based on the current system time and a preset time knowledge base. The time format can be, for example, a format for year, month, and day: YYYY-MM-DD, or a format accurate to the second: YYYY-MM-DD HH:MM.

[0039] For example, if a user enters the statement: "I want to go to Beijing tomorrow afternoon," the time segment is identified as "tomorrow." Combining this with the current system time "2025-03-27," the specific date corresponding to "tomorrow" is determined to be "2025-03-28." Therefore, time data in a preset time format is generated: {"time_type": "departure_time", "value": "2025-03-28}.

[0040] The steps for extracting domain entities: A pre-trained entity extraction model extracts domain entities in a random order based on the semantics of each entity in the dialogue, resulting in a list of key information fragments.

[0041] The entity extraction model in this embodiment generally adopts the Transformer-NER model, specifically a BERT-based sequence labeling model. This entity extraction model has been trained on a large corpus of travel-related language, possesses semantic understanding capabilities, and can identify the inherent meaning of entities and extract entities based on the identified semantics. Since the entity extraction model extracts entities based on semantic features rather than positional features, the extraction process does not need to rely on the fixed position or order of words in a sentence, thus effectively supporting entity extraction from disordered or non-standard word order text.

[0042] In practice, a pre-trained entity extraction model can be used to perform semantic analysis on the dialogue statements. Based on the identified semantics, domain entities such as DATE_TIME, LOCATION, PRODUCT_TYPE (e.g., hotel or airfare), and ROOM_TYPE can be extracted. The output of the entity extraction model generally includes the entity type of the domain entity (e.g., location, product type) and its corresponding text fragments. These output contents form a list of key information fragments.

[0043] For example, the user input dialogue is a fragmented phrase: "Beijing, Hotel, Executive Suite". Although the statement lacks a clear grammatical structure and the information is randomly piled up, the entity extraction model can still identify that "Beijing" belongs to the LOCATION entity, "Hotel" belongs to the PRODUCT_TYPE entity, and "Executive Suite" belongs to the ROOM_TYPE entity. Subsequently, the information of these extracted named entities is organized into a list of key information fragments: [{"entity": "Beijing", "type": "LOCATION"}, {"entity": "Hotel", "type": "PRODUCT_TYPE"}, {"entity": "Executive Suite", "type": "ROOM_TYPE"}].

[0044] The steps for merging heterogeneous data include: merging time data, a list of key information fragments, and structured data into a first set of slots.

[0045] In this embodiment, a pre-created empty slot set container can be obtained; the list of key information fragments corresponding to time data and domain entity data will be filled into the slot set container. Subsequently, the structured data is traversed, and for slots that already exist in the structured data (such as dates selected by clicking on the calendar), the values ​​in the structured data are used first, or they are logically validated and merged with the time data and list of key information fragments parsed from the dialogue statement before being filled into the slot set container, finally generating the first slot set.

[0046] The above embodiments enable the processing of fragmented inputs, complex inputs containing ambiguous temporal descriptions, and inputs mixed with interface operations. In particular, by fusing high-confidence structured data with natural language processing results, the generated first set of slots can be ensured to have extremely high accuracy, laying a solid data foundation for subsequent intent recognition and slot filling.

[0047] Step S103: The pre-trained joint learning architecture performs intent recognition and slot label prediction based on the data to be processed and the first slot set to generate the initial dialogue state for this round of dialogue.

[0048] To accurately capture users' multi-dimensional intentions and track the dialogue state in real time, this embodiment adopts a pre-trained joint learning architecture that couples intention recognition (sequence classification task) and slot label prediction (sequence labeling task) together for training and inference, thereby generating the initial dialogue state of the current round of dialogue.

[0049] Reference Figure 2, in this embodiment, the joint learning architecture performs intent recognition and slot label prediction based on the data to be processed and the first slot set, and generates the initial dialogue state of this round of dialogue, which may include the following steps.

[0050] Step S301: Input the data to be processed and the first slot set into the pre-trained joint learning architecture; wherein, the joint learning architecture includes: an encoder, an intent task head and a slot task head sharing the encoder, and both the intent task head and the slot task head use the multi-head attention mechanism.

[0051] Specifically, the joint learning architecture may include the following core components: a bottom-layer encoder and two top-layer parallel task heads, namely an intent task head and a slot task head. To enhance the focusing ability of the joint learning architecture on key dialogue data features, both the intent task head and the slot task head integrate the multi-head attention mechanism, thereby establishing a bidirectional information path between intent and slot, enabling the judgment of intent to dynamically refer to the extraction result of the slot, and vice versa.

[0052] In this embodiment, the data to be processed can be concatenated with the first slot set, or connected by a specific separator (such as [SEP]) to form an input sequence rich in context information; this sequence contains both the user's original expression and the confirmed slot information.

[0053] Input the processed data to be processed after the above concatenation into the encoder in the joint learning architecture.

[0054] Step S302: The encoder extracts dialogue data features from the data to be processed and the first slot set.

[0055] In this embodiment, the encoder (such as a Transformer-based encoder) scans the data to be processed through a multi-layer neural network. The encoder understands the meaning of the words themselves and captures the dependencies between words. For example, there is an association relationship between "tomorrow" and "flight", and there is a business logic relationship between "Shanghai" and "Beijing" in the first slot set.

[0056] Convert discrete words (such as "check", "Shanghai", "adult ticket") into computable high-dimensional vector features. These high-dimensional vector features contain both the semantics of the text and the existing states in the first slot set, forming dialogue data features rich in context information.

[0057] Step S303: The intent task head using the multi-head attention mechanism performs intent recognition on the dialogue data features to obtain the target intent label; In this embodiment, the intent task head first performs a global scan of the dialogue data features through a multi-head attention mechanism to capture intent features that play a key role in judging intent, such as verbs that can represent action behaviors and core objects that represent business entities.

[0058] Subsequently, to further eliminate semantic ambiguity, the intent task head, based on a multi-head attention mechanism, calculates the semantic relevance between the aforementioned intent features and the slot features output by the slot task head. Specifically, the multi-head attention mechanism allows the intent task head to dynamically monitor the results of slot extraction, using high-confidence slot information (such as identified entities like "hotel" and "executive suite") as contextual enhancement signals and fusing them into the representation space of the intent features. Based on these enhanced intent features after fusing slot semantics, the main intent of the user's current dialogue can be determined more accurately, thereby outputting the target intent label.

[0059] The intent recognition mechanism in this embodiment overcomes the limitations of traditional methods that rely solely on text sequence features. By introducing slot features as guidance, it achieves collaborative inference of intent assisted by slots. This joint mapping approach effectively alleviates the challenge of intent recognition when user input is ambiguous or the sentence order is disordered, significantly enhancing intent recognition capabilities and improving its accuracy.

[0060] Step S304: The slot task head using the multi-head attention mechanism predicts the slot labels of the data features to obtain the second slot set and its corresponding slot label sequence.

[0061] In this embodiment, the slot task head also employs a multi-head attention mechanism to perform fine-grained semantic analysis on the dialogue data features output by the encoder. Specifically, this slot task head not only focuses on the local contextual features of each word in the dialogue data features, but also dynamically introduces intent features from the intent task head as global semantic constraints through a multi-head attention mechanism.

[0062] For example, when the intent task head determines that the target intent label of the current dialogue is "flight booking", this intent feature will guide the slot task head to focus on slot types related to flight business (such as departure point, destination, departure time, cabin class, etc.) and suppress slot predictions that are not related to the current intent (such as "check-in date" and "room type" in hotel booking).

[0063] Based on this, the slot task head analyzes the features of the dialogue data word by word, predicting the corresponding slot label for each word, such as "Beijing" corresponding to "B-DESTINATION" and "tomorrow" corresponding to "B-DEPARTURE_TIME". Finally, the slot labels of all words are integrated to obtain the second set of slots added or corrected in this round of dialogue and their corresponding slot label sequences.

[0064] This joint prediction mechanism for intent-guided slots in this embodiment can effectively avoid the slot filling deviation problem caused by intent recognition errors in traditional pipeline models, and significantly improve the accuracy and business relevance of slot label prediction.

[0065] Step S305: Generate the initial dialogue state based on the target intent tag, the first slot set, the second slot set, and the slot tag sequence.

[0066] In this embodiment, the target intent tag can be used as the intent field of the initial dialogue state. Subsequently, the first slot set and the second slot set are merged and deduplicated. For slots with conflicts (e.g., "Beijing" in the first slot set and "Shanghai" in the second slot set), a decision is made according to a preset priority strategy. This priority strategy may be: latest input first or structured data first.

[0067] The final generated initial dialogue state can be a structured JSON object, which integrates UI interaction events, historical slots, and semantic understanding results of the current dialogue round, providing an accurate data foundation for subsequent dialogue strategy formulation.

[0068] Step S104: Slot fusion of the initial dialogue state based on the historical dialogue state of the historical rounds to obtain the target dialogue state.

[0069] To maintain the contextual coherence of multi-turn dialogues and improve information integrity, this embodiment introduces slot fusion of the initial dialogue state with the historical dialogue state. The slot fusion may include a series of operations such as overriding, inheritance, and resolution, integrating the initial dialogue state with the historical dialogue state to finally generate the target dialogue state for the current round of dialogue.

[0070] Optional, refer to Figure 3 An example of obtaining the target dialogue state by slot fusion of the initial dialogue state based on the historical dialogue state of the historical rounds may include the following steps.

[0071] Step S401: If the time tag of the time slot in the initial dialogue state does not match the time tag of the time slot in the historical dialogue state, update the dialogue state by overwriting the time tag of the time slot in the initial dialogue state.

[0072] In this embodiment, a semantic comparison of time tags is performed between the historical dialogue state and the initial dialogue state. If the time slot in the initial dialogue state does not match the time slot in the historical dialogue state (e.g., the historical dialogue state shows "this Friday," while the initial dialogue state shows "replaced with next Friday"), a time conflict occurs. In this case, an overwrite operation is performed, and the time tag extracted from the initial dialogue state is used to update the time tag of the corresponding time slot in the historical dialogue state.

[0073] Step S402: For the first slot that is missing in the initial dialogue state, inherit the first slot and its slot label from the historical dialogue state.

[0074] In this embodiment, an integrity check is performed on the initial dialogue state to identify the first slot that is necessary in the current business scenario but missing in this round of dialogue. For example, if the initial dialogue state only contains room type (such as "twin room"), it is understandable that in a business scenario of querying room type, there must be a first slot to limit the room type, such as destination and hotel type. In this case, the aforementioned first slot and its slot label are retrieved and inherited from the historical dialogue state. This inheritance mechanism avoids users repeatedly providing the same information in multiple rounds of dialogue, ensuring that the booking operation can use the user's previous high-value choices.

[0075] Step S403: Perform referential resolution on pronouns in the initial dialogue state based on the historical dialogue state. These pronouns include: it, that house, this, etc.

[0076] In this embodiment, the referential slots in the initial state are first identified. Then, in the slot set of historical dialogue states, the most likely antecedent is matched based on semantic relevance, temporal distance, and business logic. The referential is replaced with its specific entity value, completing the referential resolution, and the referential slots are filled with accurate slot labels.

[0077] Step S404: After completing the above slot fusion, the target object state is obtained; wherein, the target object state includes: target consciousness tag, current round slot set, and current round slot tag set.

[0078] It is understandable that after completing a series of slot fusion processes such as overriding, inheritance, and resolution, we can obtain multiple slots that the user actually mentioned in this round of conversation and that have accurate slot tags. These slots constitute the current round of slot set. Correspondingly, the slot tags corresponding to each slot in the current round of slot set constitute the current round of slot tag set.

[0079] Building upon this, the target awareness tag, the current round slot set, and the current round slot tag set are encapsulated to generate the target dialogue state for this round of conversation. This target dialogue state can be a structured data object, typically in JSON format.

[0080] At this point, this embodiment has completed a full round of dialogue understanding and state tracking. The target dialogue state will be passed to subsequent steps for completeness detection and instruction generation, etc., to generate the final dialogue response content.

[0081] Step S105: Perform a completeness check on the target dialogue state and generate dialogue instructions based on the check results.

[0082] To ensure that all necessary business information has been collected before the final action (such as calling an API or generating a response) is performed, this embodiment introduces a preset business rule knowledge base to perform a completeness check on the target dialogue state. This check mainly detects whether the mandatory slots associated with the target awareness tag in this round are complete, and generates dialogue instructions based on the detection results.

[0083] like Figure 4 As shown, this embodiment may include steps S501-S507.

[0084] Step S501: Call the preset business rule knowledge base; wherein, the business rule knowledge base is used to record the set of intent tags and forced slots with related relationships.

[0085] The aforementioned business rule knowledge base can be a structured database or configuration file used to record the mapping relationship between each intent tag (such as booking a hotel, checking the weather, transferring money, etc.) and its corresponding mandatory slot set. The mandatory slot set contains the core slots necessary for executing the business logic corresponding to its intent.

[0086] Step S502: Obtain the set of target mandatory slots associated with the target intent tag in the target dialogue state from the business rule knowledge base.

[0087] This embodiment extracts the target intent tag from the target dialogue state, and uses the target intent tag as an index to retrieve and obtain the set of target mandatory slots associated with it from the business rule knowledge base.

[0088] Step S503: Perform a completeness check on the current slot set based on the target forced slot set.

[0089] Step S504: If at least one mandatory slot is missing in the current slot set, then fill the missing mandatory slot with slot labels according to the user profile database and corporate travel standard information.

[0090] This embodiment can iterate through the target mandatory slot set and compare each slot with a valid value in the current round of the target dialogue state's slot set. If some mandatory slots in the target mandatory slot set are not present in the current round of the slot set, that is, if mandatory slots are missing in the current round of the slot set, then the missing mandatory slots can be automatically filled with slot labels.

[0091] For example, if the target intent tag is hotel booking, the associated target mandatory slot set is: {destination_city (destination city), checkin_date (check-in date), checkout_date (check-out date), room_type (room type)}. The slot set for this round is: {destination_city: "Shanghai", checkin_date: "2023-11-05", room_type: "King Room"}.

[0092] Upon comparison, it was discovered that the `checkout_date` (departure date) was missing from the target slot set in the current slot set. Therefore, it was determined that the slot information in the current slot set was incomplete. In this situation, directly asking the user for supplementary information would interrupt their train of thought, potentially disrupting their flow and degrading the user experience.

[0093] To address this issue, this embodiment does not simply expose the missing mandatory slots directly to the user when performing a completeness check. Instead, it prioritizes attempting to automatically fill the missing mandatory slots before triggering the follow-up questioning process (i.e., subsequent step S505).

[0094] This automated data entry process relies on external data sources, primarily user profile databases and corporate travel standards information. The user profile database stores users' historical behavioral preferences, such as whether a user has chosen four-star hotels in 80% of their past hotel bookings or prefers window seats. Corporate travel standards information represents company-level policy constraints, such as a daily limit of 500 RMB for employee business trip accommodations, and a hotel rating not exceeding four stars.

[0095] In this case, based on the user profile database and corporate travel standard information, the missing mandatory slots are filled with slot labels, which may include the following embodiments.

[0096] Based on the type of the missing mandatory slot, retrieve the relevant target data from the two data sources mentioned above.

[0097] If, after searching, neither the user profile database nor the corporate travel standard information contains target data associated with the missing mandatory slot, it is determined that there is currently insufficient contextual basis to support the automatic filling of the mandatory slot. In this case, to ensure the accuracy of business data and avoid service deviations caused by blind guessing, this embodiment terminates automatic filling and triggers subsequent step S505 to guide the user to explicitly supplement the missing information, thereby ensuring the completeness of the target mandatory slot set.

[0098] If the target data can be successfully retrieved, then verify whether the personal preferences represented by the user profile data in the target data conflict with the policy constraints in the company's travel standards.

[0099] For example, if a user profile shows a preference for five-star hotels, but the company's travel standards clearly stipulate a maximum of four-star accommodation, then a conflict is identified. In this case, it is impossible to determine which rule should be followed, therefore the autofill fails, and the subsequent step S505 is immediately triggered, which involves sending a clear clarification request to the user to resolve the rule conflict.

[0100] If there is no contradiction between the two, for example: the user profile database shows that the user prefers four-star hotels, while the corporate travel standard information shows that the travel standard is five-star hotels; it can be seen that the standard of five-star hotels can cover the user's preference for four-star hotels, so the two do not conflict.

[0101] Without conflict, slot labels are generated for missing mandatory slots based on the target data. Specifically, the retrieved valid target data is converted into slot labels that conform to the target dialogue state format requirements and written into the target dialogue state, thereby completing the automatic filling of the missing mandatory slot.

[0102] For example, if the missing mandatory slot is the hotel star rating, the slot label "hotel_star_rating: 4" will be automatically generated based on the target data of the user profile database that prefers four-star hotels, and then added to the target dialogue state.

[0103] This embodiment automatically fills in missing mandatory slot labels, effectively reducing redundant follow-up questions during human-computer interaction. Users no longer need to declare each piece of business information individually; instead, the remaining slot labels are automatically filled in based on historical preferences and company standards. This intelligent slot-filling method not only significantly shortens the task completion path and reduces user operating costs but also improves the naturalness and fluency of human-computer interaction, greatly enhancing the processing efficiency of multi-turn dialogue tasks.

[0104] Step S505: Use the failed-to-fill slots as the second slots, and sort the second slots according to the priority of business logic.

[0105] In practical applications, for multiple missing mandatory slots in the target mandatory slot set, only some slots may be successfully filled automatically, while the remaining slots may fail to be filled due to lack of data support or rule restrictions. Therefore, this embodiment uses the mandatory slots that failed to be filled as the second slot. For the second slot, relevant information still needs to be obtained from the user in the current target dialogue state.

[0106] To avoid sending users a barrage of follow-up questions, this embodiment pre-defines business logic priorities. These priorities are ordered based on the dependencies or importance of slots during business execution. For example, in a hotel booking scenario, the business logic priorities are set as follows: Date > Location > Room Type. That is, determining check-in and check-out dates typically takes precedence over determining the specific room type.

[0107] The second slot is reordered according to the priority of business logic. For example, if the set of second slots contains {room_type, checkin_date}, after sorting, checkin_date (which belongs to the date dimension) will be placed before room_type (which belongs to the room type dimension).

[0108] Step S506: Generate an inquiry instruction for each second slot according to the sorting result to inquire about the slot label.

[0109] Specifically, based on the sorting results, query commands for querying slot labels are generated sequentially for each second slot. For example, for the first-ranked slot, `checkin_date`, the following query command is generated: `Request(checkin_date)`. The generated query commands are then passed to the natural language model and translated into specific user responses.

[0110] The above embodiments employ a hybrid approach that prioritizes automatically filling slot labels and then generates query commands for the remaining slots according to priority. This approach not only utilizes known data sources to reduce interaction rounds but also ensures that query commands are generated in the order that best aligns with business logic when data is missing. This results in a clear and smooth dialogue process.

[0111] Step S507: In the current round of slot set in the target dialogue state, if there is a slot with semantically ambiguous slot label, then generate a clarification instruction to clarify the slot label.

[0112] To address the common semantic ambiguity problem in natural language understanding, this embodiment can identify ambiguous expressions in user input and eliminate semantic ambiguity by generating specific clarification instructions, thereby ensuring the accuracy of slot filling.

[0113] Specifically, semantic parsing can be performed on the dialogue statements input by the user. If there are expressions with unclear semantic boundaries or ambiguous meanings in the parsing results, the slot and slot label corresponding to the semantically ambiguous expression are determined in the slot set in this round.

[0114] For example, a user enters "Book me a meeting room tomorrow morning." While the time entity "tomorrow morning" is recognized, the concept of "morning" is ambiguous in the specific time slot mapping. It could refer to the time period "08:00-10:00," or "09:00-11:00," or even specifically "08:30." In this case, it is determined that the time expression is ambiguous, meaning that a single text input corresponds to multiple possible slot labels.

[0115] Furthermore, examine the current set of slots in the target dialogue state to obtain the ambiguous slots and ambiguous slot labels corresponding to the aforementioned semantically ambiguous expressions. It is necessary to clarify the ambiguous slots and ambiguous slot labels.

[0116] For the identified ambiguous slots and ambiguous slot labels, this embodiment can generate a clarification instruction, denoted as Clarify(time_ambiguity), to clarify the slot label. This clarification instruction does not directly execute the reservation operation, but rather triggers the clarification process of the dialogue manager. Specifically, based on the Clarify(time_ambiguity) instruction, a targeted follow-up question is called from a preset dialogue library or dynamically generated: "We found a time slot for tomorrow morning. Do you mean 8:00 AM or 9:00 AM specifically?"

[0117] This mechanism enables the proactive identification and processing of ambiguous and unclear user intents, transforming vague natural language into precise structured data, thereby avoiding business execution errors caused by guessing user intents.

[0118] Step S106: Respond to this round of dialogue according to the dialogue instructions.

[0119] This embodiment includes: generating dialogue statements to respond to the current round of dialogue based on dialogue instructions and preset dialogue templates using a natural language model.

[0120] In one scenario, the dialogue instruction is the inquiry instruction as described in the preceding embodiments. Specifically, the natural language model parses the dialogue instruction as an inquiry instruction and obtains the slot of the inquiry slot label carried by the inquiry instruction. The natural language model extracts the semantic information of the slot and retrieves the inquiry template corresponding to the semantic information from a preset dialogue template.

[0121] For example, in response to the Request(checkin_date) command, the slot is identified as "check-in date," and the dialogue template "What are your specific plans for [action]?" is invoked. The "[action]" in the template is then filled in with "checking into the hotel" based on the context, ultimately generating the dialogue statement that responds to this round of conversation: "What specific date do you plan to check into the hotel?". This template-based generation method ensures the standardization and accuracy of key information inquiries.

[0122] In another scenario, the dialogue instruction is the clarification instruction as described in the preceding embodiments. Specifically, the natural language model parses the dialogue instruction as a clarification instruction and obtains the slot of the clarification instruction carrying the slot label to be clarified. The natural language model extracts the semantic ambiguity type of the slot and retrieves the clarification template corresponding to the semantic ambiguity type from the preset dialogue templates.

[0123] For example, in response to the Clarify (destination) command, when it is detected that the user's input "Nanjing" may refer to "Nanjing City" or "Nanjing Road", this embodiment will fill these two candidate entities as variables into the clarification sentence template and generate a dialogue statement to respond to this round of dialogue: "Do you mean Nanjing City or Nanjing Road when you say 'Nanjing'?"

[0124] In this embodiment, a natural language model is used to transform dialogue commands into natural, fluent responses that conform to human language habits. These natural language responses not only comply with business logic but also provide a good interactive experience, thereby guiding users to complete subsequent information supplementation or confirmation.

[0125] According to the above embodiments, if the completeness check of the target dialogue state passes, a structured interface request is constructed for the target dialogue state to call the backend service, and the response result to the interface request is displayed in the form of a rich media card.

[0126] Specifically, if all mandatory slots in the target dialogue state are filled and unambiguous, the completeness check of the target dialogue state is considered passed. In this case, the task execution and result feedback phase can proceed based on the complete target dialogue state. This process mainly includes: structured request construction, backend service invocation, and rich media card rendering.

[0127] Structured request construction involves encapsulating discrete target dialogue states into standard structured interface requests based on predefined API protocols (such as JSON Schema).

[0128] Backend service calls include: sending the API request to the corresponding backend service, such as a hotel inventory management system, airline system, or third-party OTA aggregation platform, through an API gateway. Upon receiving the API request, the backend service executes specific business logic, such as inventory query, price calculation, and discount matching, and returns response data.

[0129] To enhance the user experience, this embodiment does not directly return response data in plain text format. Instead, it utilizes rich media card technology to visualize and render the response data. The response data typically includes a list of business results and detailed attributes, such as hotel name, image URL, price, rating, distance, and bookability. This embodiment parses and cleans the returned response data, extracts key fields for front-end display, and maps them to rich media cards.

[0130] The rich media card is pushed to the user's client interface. Users can intuitively browse search results in the current dialog box without jumping to another page, and directly initiate the next round of interaction (such as triggering a payment process) by clicking the button on the card, thus achieving a seamless closed loop from dialogue to service.

[0131] In summary, the intent-based dialogue method provided in this disclosure includes: acquiring data to be processed for the current round of dialogue, the data to be processed including: dialogue statements input by the user and structured data generated based on interaction events; performing normalization processing on the data to be processed to obtain a first slot set; wherein, the normalization processing includes at least: time parsing, extraction of domain entities, and fusion of heterogeneous data; generating an initial dialogue state for the current round of dialogue by performing intent recognition and slot label prediction based on the data to be processed and the first slot set using a pre-trained joint learning architecture; performing slot fusion on the initial dialogue state according to the historical dialogue states of previous rounds to obtain a target dialogue state; performing completeness detection on the target dialogue state and generating dialogue instructions based on the detection results; and responding to the current round of dialogue according to the dialogue instructions.

[0132] In this technical solution, firstly, by using unprocessed data containing dialogue statements and structured data, and combining normalization processes such as time parsing, domain entity extraction, and heterogeneous data fusion, fragmented information originally scattered in natural language text and interaction events can be uniformly transformed into a first set of slots that are understandable and easy for machines to process. Then, a joint learning architecture is used to simultaneously perform intent recognition and slot label prediction, avoiding information loss caused by error propagation in traditional pipelined processing. By utilizing the dual features of the unprocessed data and the first set of slots, the joint learning architecture can more accurately understand the user's true intent in a specific context and directly output the initial dialogue state with the target intent label, thereby significantly improving the efficiency and accuracy of the conversion from unstructured input to a semi-structured state. Secondly, by introducing historical dialogue states to perform slot fusion on the initial dialogue state, it is possible to better achieve cross-turn information inheritance and completion, effectively alleviating the problem of incomplete or ambiguous information in single-turn dialogues, allowing the target dialogue state to accumulate and integrate fragmented information from multiple turns of interaction. The resulting target dialogue state not only includes the explicit requirements of the current round but also incorporates implicit constraints from historical interactions, improving accuracy and completeness. Based on this, dialogue instructions generated from the completeness check results and responses guide users to complete all the information required for their business. Overall, this solution efficiently and accurately transforms fragmented user expressions into business requests that can directly invoke backend services, thereby improving the completeness and accuracy of dialogue information conversion in human-computer dialogue scenarios and enhancing the human-computer interaction experience.

[0133] More specifically, this solution can achieve the following effects: (1) The interaction efficiency is greatly improved: Through dynamic slot filling and automatic completion mechanism based on known data sources, unnecessary interaction rounds are significantly reduced. Users only need to input core information to complete the task in the shortest time.

[0134] (2) High robustness and accuracy: The joint learning architecture is used to process intent and slots, which solves the problem of insufficient understanding of fragmented and non-standardized inputs in traditional solutions and improves the recognition accuracy of key information for travel booking.

[0135] (3) Intelligent dialogue experience: It realizes the follow-up questioning logic similar to a real assistant, and can proactively and intelligently initiate follow-up questions based on the importance of missing information and the user's historical preferences, rather than mechanically asking the user to fill in all the information.

[0136] (4) Powerful context maintenance capability: Through dialogue state tracking, it can stably track and update slot information in multiple dimensions such as time, location, and room type, ensuring that users can modify any dimension of information at any time during multiple rounds of dialogue without worrying about losing context.

[0137] (5) Strong business scalability: Business rules and mandatory slot definitions are independent of the underlying model. When the business or standards of travel change, the system can adapt by simply updating the knowledge base, which has good maintainability and scalability.

[0138] Corresponding to the above embodiments of the dialogue method based on intent recognition, this invention also provides a dialogue system based on intent recognition, such as... Figure 5 As shown, the system includes: Dialogue data acquisition module 100: used to acquire the data to be processed in the current dialogue, the data to be processed includes: dialogue statements input by the user and structured data generated based on interaction events; Data processing module 200: used to perform normalization processing on the data to be processed to obtain a first slot set; wherein, the normalization processing includes at least: time parsing, extraction of domain entities and fusion of heterogeneous data; Model processing module 300: used to perform intent recognition and slot label prediction based on the data to be processed and the first slot set by a pre-trained joint learning architecture, and generate the initial dialogue state of this round of dialogue; Slot fusion module 400: used to perform slot fusion on the initial dialogue state according to the historical dialogue state of the historical rounds to obtain the target dialogue state; Instruction generation module 500: used to perform completeness detection on the target dialogue state and generate dialogue instructions based on the detection results; Dialogue response module 600: used to respond to the current dialogue according to the dialogue instructions.

[0139] As demonstrated by the intent-based dialogue system described above, this system transforms the data to be processed into a first set of slots that the machine can understand through normalization. By combining intent recognition and slot prediction simultaneously with a joint learning architecture, it avoids the error accumulation of traditional pipelined processing, significantly improving the efficiency and accuracy of converting unstructured input into a semi-structured state. Furthermore, the introduction of historical dialogue states for slot fusion effectively solves the problems of cross-turn information inheritance and ambiguous referencing, ensuring the integrity of the target dialogue state. Finally, dialogue instructions generated based on completeness detection guide the user to complete key information, efficiently and accurately converting fragmented user expressions into directly invoked structured business requests, thereby improving the accuracy of information conversion and the business loop capability in human-computer dialogue scenarios.

[0140] The intent-based dialogue system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned intent-based dialogue method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned intent-based dialogue method embodiment.

[0141] This embodiment also provides an electronic device, the structural schematic diagram of which is shown below. Figure 6 As shown, the device includes a processor 101 and a memory 102; wherein the memory 102 is used to store one or more computer instructions, which are executed by the processor to implement the steps of the above-described intention-based dialogue method.

[0142] Figure 6 The electronic device shown also includes a bus 103 and a communication interface 104, with the processor 101, communication interface 104 and memory 102 connected via the bus 103.

[0143] The memory 102 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. The bus 103 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0144] The communication interface 104 is used to connect to at least one user terminal and other network units through a network interface, and to send encapsulated IPv4 packets or IPv4 packets to the user terminal through the network interface.

[0145] Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. The processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 102. The processor 101 reads the information in memory 102 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0146] In practical implementation, this electronic device has functions such as data input, equipment selection, calculation and processing, list generation, data storage and transmission, adapting to the operational needs of residential photovoltaic power station site surveys. Its processor can adopt a high-performance, low-power processor (such as Snapdragon 870, Dimensity 9200, etc.), with a main frequency of no less than 2.4GHz, supporting multi-threaded processing to ensure rapid response in complex calculations such as equipment selection and material calculations; it is equipped with no less than 8GB of LPDDR5 RAM to ensure smooth multitasking; and it has built-in no less than 256GB of UFS 3.1 storage space for storing equipment library data, project data, material lists, and cost details.

[0147] This electronic device can be equipped with a 6.7-inch AMOLED high-definition touchscreen (2400×1080 resolution), supporting handwriting input and multi-touch, facilitating quick input of parameters such as roof dimensions on-site; it integrates a high-precision GPS positioning module to automatically record the geographical location information of the survey site and link it to project data; it is equipped with a 13-megapixel rear camera to support roof photo measurement function, automatically extracting roof dimensions through image recognition technology to reduce manual measurement errors; it supports Bluetooth 5.2, Wi-Fi 6 and 5G communication modules to realize equipment library data updates, project data synchronization and material list sharing, ultimately realizing the data input and interaction process.

[0148] This electronic device features a built-in 5000mAh high-capacity lithium battery, supports 65W fast charging, and can meet the needs of on-site operation for more than 8 hours on a single charge. The terminal dimensions are controlled within 160mm × 75mm × 8.5mm, and the weight does not exceed 200g. It is equipped with a non-slip, wear-resistant shell and a foldable stand, making it easy to handle by hand or place on any flat surface. Furthermore, the device features a USB-C interface for data transfer and charging; a reserved NFC interface for quick pairing and data reading with power station equipment; and a 3.5mm headphone jack for convenient voice navigation in noisy environments.

[0149] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, devices, and methods can be implemented in other ways. The system 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. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0150] 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.

[0151] In addition, the functional units in the various embodiments of the present invention 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.

[0152] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion 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, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0153] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dialogue method based on intent recognition, characterized in that, The method includes: Obtain the pending data for this round of dialogue, which includes: dialogue statements input by the user and structured data generated based on interaction events; The data to be processed is normalized to obtain a first set of slots; wherein, the normalization process includes at least: time parsing, extraction of domain entities and fusion of heterogeneous data; The pre-trained joint learning architecture performs intent recognition and slot label prediction based on the data to be processed and the first slot set, generating the initial dialogue state of this round of dialogue; The initial dialogue state is slot-fused based on the historical dialogue state of the historical rounds to obtain the target dialogue state; Perform a completeness check on the target dialogue state and generate dialogue instructions based on the check results; Respond to this round of dialogue according to the dialogue instructions.

2. The dialogue method based on intent recognition according to claim 1, characterized in that, The process of normalizing the data to be processed to obtain a first set of slots includes: The time in the dialogue statement is parsed to obtain time data in a preset time format; The pre-trained entity extraction model extracts domain entities in random order based on the semantics of each domain entity in the dialogue statement to obtain a list of key information fragments. The time data, the list of key information fragments, and the structured data are merged into a first slot set.

3. The dialogue method based on intent recognition according to claim 1, characterized in that, The pre-trained joint learning architecture performs intent recognition and slot label prediction based on the data to be processed and the first slot set, generating the initial dialogue state for this round of dialogue, including: The data to be processed and the first set of slots are input into a pre-trained joint learning architecture; wherein, the joint learning architecture includes: an encoder, which shares the intent task head and the slot task head of the encoder, and both the intent task head and the slot task head use a multi-head attention mechanism; The encoder extracts dialogue data features from the data to be processed and the first slot set; The intent task head, employing a multi-head attention mechanism, performs intent recognition on the dialogue data features to obtain the target intent label; The slot task head using a multi-head attention mechanism predicts the slot labels for the data features, resulting in a second set of slots and its corresponding slot label sequence. An initial dialogue state is generated based on the target intent tag, the first slot set, the second slot set, and the slot tag sequence.

4. The dialogue method based on intent recognition according to claim 1, characterized in that, The step of slot fusion of the initial dialogue state based on the historical dialogue states of historical rounds to obtain the target dialogue state includes: If the timestamp of the time slot in the initial dialogue state does not match the timestamp of the time slot in the historical dialogue state, the dialogue state is updated by overwriting the timestamp of the time slot in the initial dialogue state. For the first slot missing in the initial dialogue state, the first slot and its slot label are inherited from the historical dialogue state; Based on the historical dialogue state, the pronouns in the initial dialogue state are resolved by dereferencing; After completing the above slot fusion, the target object state is obtained; wherein, the target object state includes: target awareness tag, current round slot set, and current round slot tag set.

5. The dialogue method based on intent recognition according to claim 4, characterized in that, The step of performing a completeness check on the target dialogue state and generating dialogue instructions based on the check result includes: Invoke a preset business rule knowledge base; wherein, the business rule knowledge base is used to record related intent tags and mandatory slot sets; Obtain the set of target mandatory slots associated with the target intent tag in the target dialogue state from the business rule knowledge base; The completeness of the current slot set is checked based on the target forced slot set; If at least one mandatory slot is detected to be missing in the current slot set, the missing mandatory slot will be filled with slot labels based on the user profile database and corporate travel standard information. The slots that fail to fill are designated as the second slots, and the second slots are sorted according to the priority of business logic. Generate an inquiry command for each of the second slots based on the sorting results, to inquire about the slot label.

6. The dialogue method based on intent recognition according to claim 1 or 5, characterized in that, The method further includes: If there is a slot with a semantically ambiguous slot label in the current round slot set in the target dialogue state, a clarification instruction is generated to clarify the slot label.

7. The dialogue method based on intent recognition according to claim 6, characterized in that, The step of responding to the current dialogue according to the dialogue instructions includes: The natural language model generates dialogue statements to respond to the current round of dialogue based on the dialogue instructions and preset dialogue templates.

8. The dialogue method based on intent recognition according to claim 1, characterized in that, The method further includes: If the completeness check of the target dialogue state passes, a structured interface request is constructed for the target dialogue state to call the backend service, and the response to the interface request is displayed in the form of a rich media card.

9. A dialogue system based on intent recognition, characterized in that, The system includes: Dialogue data acquisition module: used to acquire the data to be processed in the current dialogue, including: dialogue statements input by the user and structured data generated based on interaction events; Data processing module: used to perform normalization processing on the data to be processed to obtain a first slot set; wherein, the normalization processing includes at least: time parsing, extraction of domain entities and fusion of heterogeneous data; Model processing module: used by a pre-trained joint learning architecture to perform intent recognition and slot label prediction based on the data to be processed and the first slot set, to generate the initial dialogue state of this round of dialogue; Slot fusion module: used to fuse the initial dialogue state with slots based on the historical dialogue states of historical rounds to obtain the target dialogue state; Instruction generation module: used to perform completeness detection on the target dialogue state and generate dialogue instructions based on the detection results; Dialogue Response Module: Used to respond to the current dialogue based on the dialogue instructions.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the steps of the intent-based dialogue method according to any one of claims 1 to 8.

Images