Method and device for processing data in whole process based on large model

Abstract

Embodiments of the present disclosure disclose a large model-based full-process data intelligent processing method and device. A specific implementation of the method includes: performing element analysis on a multi-source business data set to generate a text element information group set; performing dialogue reference construction on the text element information group set to generate a dialogue reference information set; based on user dialogue text, performing the following dialogue process processing steps: performing dialogue state recognition on the user dialogue text to generate real-time dialogue state information; selecting target dialogue strategy information according to the real-time dialogue state information; generating dialogue reply information according to a large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and each multi-dimensional index information; pushing a message to a user end, and storing the real-time dialogue state information. The implementation can improve the accuracy of task-based dialogue question answering and reduce dialogue response time in a target business scenario.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and more specifically to a method and apparatus for intelligent processing of end-to-end data based on large models. Background Technology

[0002] As government information services gradually move towards intelligentization, information consultation services based on task-oriented dialogue systems are increasingly being applied to government platforms, enabling users to obtain relevant business information. Currently, when performing end-to-end data processing, task-oriented dialogue systems typically rely on keyword matching or knowledge base retrieval to parse and semantically search policy documents, and combine this with large language models for question answering.

[0003] However, when using the above method for end-to-end data processing in a task-oriented dialogue system, the following technical problems often arise: Using keyword matching or knowledge base retrieval makes it difficult to effectively parse and semantically understand multi-source unstructured policy data. As a result, when combining large language models to generate answers, the generated answers contain erroneous information and lack source evidence, often requiring additional verification. This leads to lower question-and-answer accuracy and longer response times. Summary of the Invention

[0004] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0005] Some embodiments of this disclosure propose a method and apparatus for intelligent processing of end-to-end data based on a large model, in order to solve the technical problems mentioned in the background section above.

[0006] Firstly, some embodiments of this disclosure provide a full-process intelligent data processing method based on a large model. The method includes: parsing the collected multi-source business dataset according to pre-constructed key element information to generate a set of text element information groups, wherein each of the collected multi-source business data is applied to a target business scenario; constructing dialogue references on the aforementioned text element information group set to generate a set of dialogue reference information, wherein each dialogue reference information corresponds to multi-dimensional index information; in response to determining that a user initiates a dialogue for the aforementioned target business scenario, based on the real-time received user dialogue text, executing the following dialogue process steps: according to historical... The system identifies the dialogue state of the received user dialogue text to generate real-time dialogue state information. Based on this real-time dialogue state information, it selects dialogue strategy information that meets the state matching conditions from a pre-built dialogue strategy information set as the target dialogue strategy information. Using a pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information, it generates dialogue response information for the user dialogue text. Based on the dialogue response information, it pushes messages to the user and stores the real-time dialogue state information as historical dialogue state information.

[0007] Secondly, some embodiments of this disclosure provide a full-process intelligent data processing device based on a large model. The device includes: an element parsing unit configured to parse the collected multi-source business dataset according to pre-constructed key element information to generate a set of text element information groups, wherein each of the collected multi-source business data is applied to a target business scenario; a dialogue reference construction unit configured to construct dialogue references on the above-mentioned text element information group set to generate a set of dialogue reference information, wherein each dialogue reference information corresponds to multi-dimensional index information; and a dialogue process processing unit configured to, in response to determining that the user initiates a dialogue in the above-mentioned target business scenario, execute based on the real-time received user dialogue text. The following dialogue processing steps are performed: Based on historical dialogue state information, the received user dialogue text is analyzed to identify the dialogue state and generate real-time dialogue state information; based on the real-time dialogue state information, dialogue strategy information that meets the state matching conditions is selected from a pre-built dialogue strategy information set as the target dialogue strategy information; based on a pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information, dialogue response information for the user dialogue text is generated; based on the dialogue response information, a message is pushed to the user terminal, and the real-time dialogue state information is stored as historical dialogue state information.

[0008] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect above.

[0009] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0010] The various embodiments of this disclosure have the following beneficial effects: the full-process data intelligent processing method based on a large model, as described in some embodiments of this disclosure, can improve the accuracy of task-oriented dialogue question answering and reduce dialogue response time in target business scenarios. Specifically, the reason for the low accuracy and high response time of related dialogues is that keyword matching or knowledge base retrieval makes it difficult to effectively parse and semantically understand multi-source unstructured policy data. This results in the generated answer containing erroneous information and lacking source evidence when combined with a large language model, often requiring additional verification, which in turn leads to lower question-and-answer accuracy and higher response time. Based on this, the full-process data intelligent processing method based on a large model, as described in some embodiments of this disclosure, firstly, performs element parsing on the collected multi-source business dataset based on pre-constructed key element information to generate a set of text element information groups. The collected multi-source business data is applied to the target business scenario. Thus, by parsing the key elements of the multi-source business data, unstructured business text can be converted into a set of text element information groups organized according to key element fields, resulting in structured text data. Then, dialogue references are constructed on the aforementioned set of text element information to generate a set of dialogue reference information. Each dialogue reference corresponds to multi-dimensional index information. This allows for the construction of dialogue reference information containing semantic and index information from the text element information, establishing a multi-dimensional index structure. This enables the semanticization and effective retrieval of business knowledge, improving the efficiency of business knowledge retrieval and the accuracy of semantic matching during dialogue. Subsequently, in response to determining that the user has initiated a dialogue for the aforementioned target business scenario, the following dialogue process steps are executed based on the real-time received user dialogue text: First, based on historical dialogue state information, the received user dialogue text is subjected to dialogue state recognition to generate real-time dialogue state information. Thus, by performing intent recognition and slot filling on the user dialogue text, and updating the state in conjunction with historical dialogue states, real-time dialogue state information reflecting the user's real-time needs is generated, thereby improving the accuracy of the dialogue system in recognizing user intent and its overall semantic understanding of the context in multi-turn dialogues. Second, based on the aforementioned real-time dialogue state information, dialogue strategy information that meets the state matching conditions is selected from the pre-constructed set of dialogue strategy information as the target dialogue strategy information. Therefore, by selecting dialogue strategies based on real-time dialogue status, the dialogue processing method can be dynamically determined according to user intent and parameter completeness, thereby reducing invalid retrieval and misunderstanding of context, and improving the processing efficiency and response speed of the dialogue process. The third step involves generating dialogue response information for the aforementioned user dialogue text based on the pre-trained large language model, the aforementioned target dialogue strategy information, the aforementioned real-time dialogue status information, the aforementioned dialogue reference information set, and the corresponding multi-dimensional index information.Therefore, by combining semantic knowledge from the dialogue reference information set with the generative capabilities of the large language model, the generated results can be inferred based on searchable business information, thereby reducing errors in generative responses and improving the accuracy of question-and-answer results. The fourth step involves pushing messages to the user based on the aforementioned dialogue response information and storing the real-time dialogue state information as historical dialogue state information. Thus, by constructing dialogue reference information containing semantic and index information from text element information and establishing a multi-dimensional index structure, rapid and effective retrieval of business data can be achieved, thereby improving the accuracy of responses generated using the large language model, reducing response errors, and shortening dialogue response time. Attached Figure Description

[0011] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0012] Figure 1 This is a flowchart of some embodiments of the full-process intelligent data processing method based on a large model according to the present disclosure; Figure 2 This is a diagram illustrating the correspondence between various intent categories and user slot information; Figure 3 This is a schematic diagram of the dialogue strategy information set according to the full-process data intelligent processing method based on a large model disclosed herein; Figure 4 This is a schematic diagram of the structure of some embodiments of the full-process intelligent data processing device based on a large model according to the present disclosure; Figure 5 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0013] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0014] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0015] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0016] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0018] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] Figure 1 A flowchart 100 is shown, illustrating some embodiments of the large-model-based end-to-end data intelligent processing method according to this disclosure. This large-model-based end-to-end data intelligent processing method includes the following steps: Step 101: Based on the pre-constructed key element information, perform element parsing on the collected multi-source business dataset to generate a set of text element information groups.

[0020] In some embodiments, the execution entity (e.g., a computing device) of the full-process data intelligent processing method based on a large model can perform element parsing on the collected multi-source business dataset according to pre-constructed key element information to generate a set of text element information groups. The collected multi-source business data are applied to a target business scenario. The target business scenario can refer to a data service scenario where a user obtains relevant policy or business information through a user-end dialogue system. The target business scenario can be a dialogue scenario of policy consultation or policy Q&A. The key element information can be composed of various key element fields used to describe the core business information in the target business scenario. For example, in a policy consultation scenario, the key element information can be ("region", "object", "condition", "material", "time", "amount", "processing procedure", "contact information"). The multi-source business dataset can be a collection of business data collected from different data sources that can be applied to the target business scenario. For example, in a policy consultation or policy Q&A scenario, the multi-source business data can be HTML files, policy PDF files, and related images of policy announcement web pages or related web pages collected for one or more policies. The text element information in the aforementioned set of text element information can be specific information (i.e., field values) extracted from business text that corresponds to the corresponding key element fields. The aforementioned business text can be text information extracted from the aforementioned multi-source business dataset.

[0021] In practice, firstly, the executing entity can extract the text information contained in each multi-source business data set from the aforementioned multi-source business dataset to obtain the business text. Then, according to the key element fields contained in the aforementioned key element information, the extracted business text is matched against these fields (i.e., whether the corresponding business text contains the corresponding key element field and its value) to obtain the text element information for each element. Finally, the executing entity can group the text element information with the same corresponding key element fields into the same group to obtain a set of text element information groups.

[0022] It should be noted that the aforementioned computing devices can be either hardware or software. When the computing device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device. When the computing device is software, it can be installed on the hardware devices listed above. It can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are made here. It should be understood that the number of the aforementioned computing devices can be arbitrary, depending on the implementation requirements.

[0023] In some optional implementations of certain embodiments, the aforementioned execution entity may perform element parsing on the collected multi-source business dataset based on pre-built key element information to generate a set of text element information groups: The first step is to perform text extraction processing on the aforementioned multi-source business dataset to generate a business text sequence. Each business text in this sequence corresponds to text location information. In practice, the execution entity can use DOM parsing to extract webpage text from HTML files, PDF text extraction to extract PDF text from PDF files, and OCR recognition to extract image text from image files. This extracts each segment of text information contained in the multi-source business data of different file types as business text. The extracted webpage text, PDF text, and image text are then concatenated and sorted according to paragraphs to obtain the business text sequence. The text location information corresponding to each business text represents its position or origin within the corresponding multi-source business data. For example, if the business text originates from a PDF file, the corresponding text location information could be the PDF name and page number. If the business text originates from an HTML file, the corresponding text location information could be the URL of the HTML file. If the business text originates from an image, the corresponding text location information could be the image identifier information representing the image.

[0024] The second step involves extracting elements from the aforementioned business text sequence based on the key element information to generate individual text element information. In practice, the executing entity can perform key element field matching on the business text sequence according to each key element field included in the key element information to generate individual text element information.

[0025] The third step involves classifying the generated text element information based on the key element fields included in the aforementioned key element information, resulting in a set of text element information groups. Each text element information group corresponds to one key element field. In practice, the executing entity can group text element information with the same corresponding key element field into the same group to obtain the text element information group set.

[0026] In some optional implementations of certain embodiments, the aforementioned execution entity may perform element extraction processing on the aforementioned business text sequence based on the aforementioned key element information through the following steps to generate various text element information: The first step is to perform the following element processing steps for each key element field included in the above key element information: The first sub-step involves performing element matching on the aforementioned business text sequence based on the key element fields to obtain candidate text sequences. In practice, the executing entity can identify candidate text sequences from the business text sequences that contain keywords corresponding to the key element fields. For example, for the key element field "time," business texts can be matched using keywords such as "application time," "deadline time," and "acceptance time."

[0027] The second sub-step involves extracting and labeling features for each candidate text within the aforementioned candidate text sequence to generate text feature information. This generated text feature information includes feature location information. In practice, for each candidate text within the aforementioned candidate text sequence, the executing entity first performs word segmentation on the candidate text using a natural language processing tool to obtain a candidate text word segmentation sequence. This natural language processing tool can be either Jieba or Yaha. Then, a pre-built sequence labeling model can be used to perform BIO sequence labeling on each candidate text word segment in the candidate text word segmentation sequence, thereby marking the text fragments corresponding to the current key feature field. Finally, the marked field value fragments are extracted and a correspondence is established with the current key feature field to generate text feature information. Each text feature information can include the key feature field name and its corresponding field value. Feature location information can be the character position of the corresponding key feature field value in the corresponding candidate text.

[0028] The sequence labeling model described above can be composed of a text encoding layer, a context modeling layer, and a label decoding layer connected sequentially. The text encoding layer can employ a BERT model encoder to convert the input text into a contextual semantic representation. The context modeling layer can use a BiLSTM network to enhance sequence positional dependencies. The label decoding layer can employ a CRF (Conditional Random Field) layer to perform sequence labeling on each token in the sequence. For any token in the candidate text segmentation sequence, its label can be divided into B-key feature field labels and I-key feature field labels, representing the start and end positions of the corresponding key feature field value, respectively. For example, the B-Time label can represent the start position of the key feature field "time".

[0029] Step 102: Construct dialogue references from the set of text element information groups to generate a set of dialogue reference information.

[0030] In some embodiments, the aforementioned execution entity can construct a dialogue reference from the aforementioned set of text element information groups to generate a dialogue reference information set. Each dialogue reference information corresponds to multi-dimensional index information. The dialogue reference information in the aforementioned dialogue reference information set can be text information that can be directly used in the dialogue process of the aforementioned target business scenario. In practice, firstly, the aforementioned execution entity can concatenate the text location information and element location information corresponding to each text element information to obtain text element location information. For example, the text location information is “File1-P2-3”, indicating that the corresponding business text originates from the 3rd paragraph on page 2 of a PDF file named “File1”. The element location information is “[10, 25]”, indicating that the position of the field value of the corresponding key element field is the 10th to 25th characters of the corresponding candidate text or business text. The concatenated text element location information can be “File1-P2-3-[10, 25]”. Then, the aforementioned execution entity can perform semantic embedding on the aforementioned business text subsequence using a pre-trained semantic encoding model to generate the corresponding business text semantic vector. The aforementioned semantic encoding model can be a sentence vector model or a text embedding model. As an example, the sentence vector model mentioned above could be a Sentence-BERT model, a SimCSE model, or a RoBERTa model. The text embedding model mentioned above could be an E5 text embedding model (e.g., multilingual-e5-base) or a BGE (BAAI General Embedding) model. Finally, the execution entity can encapsulate the text element information, the business text subsequence, the business text semantic vector, the corresponding text element location information, and the corresponding key element fields into dialogue reference information, and determine the business text semantic vector as a semantic index, the corresponding key element fields as field indexes, and the field indexes and semantic indexes as multi-dimensional index information.

[0031] In some optional implementations of certain embodiments, the aforementioned execution entity may construct a dialogue reference set by performing the following steps on the aforementioned set of text element information groups: The first step is to perform the following construction steps for each text element in the above set of text element information groups: The first sub-step involves generating text element location information based on the text location information and element location information corresponding to the aforementioned text element information. In practice, the executing entity can first concatenate the text location information and element location information corresponding to the aforementioned text element information to generate text element location information. Here, the text location information represents the position of the business text containing the text element information in the multi-source business data, and the element location information represents the position of the text element information in the corresponding business text. Thus, text element location information representing the position of the corresponding text element information in the multi-source business data can be obtained.

[0032] The second sub-step involves determining the business text sub-sequence associated with the aforementioned text element information based on the business text sequence, the text element location information, and the context window. In practice, the executing entity uses the position represented by the text element location information as the center and extracts relevant business texts within the business text sequence according to the range of the context window to obtain the business text sub-sequence associated with the text element information. The context window can be represented as the number of paragraphs of the business text to be extracted. Thus, by introducing the context window, the executing entity can not only retain the field values ​​themselves but also retain key element fields and the semantic descriptions before and after their field values, giving the subsequently generated reference information a more complete contextual semantics.

[0033] The third sub-step involves semantically encoding the aforementioned business text subsequences to obtain business text semantic vectors. In practice, the executing entity can use the aforementioned semantic encoding model to semantically embed the aforementioned business text subsequences to generate corresponding business text semantic vectors.

[0034] The fourth sub-step involves constructing dialogue business information based on the aforementioned text element information, the aforementioned business text subsequence, the aforementioned business text semantic vector, the corresponding text element location information, and the corresponding key element fields. In practice, the executing entity can uniformly encapsulate the field values ​​of the key element fields represented by the aforementioned text element information, the context text represented by the aforementioned business text subsequence, the business text semantic vector, the text element location information, and the key element fields to construct dialogue business information. This yields dialogue business information that explicitly includes information related to the key element fields.

[0035] The second step involves filtering the aforementioned business text sequences based on each candidate text sequence to generate candidate text subsequences. In practice, besides the dialogue business information directly extracted from key element fields, the aforementioned business text sequences may still contain business texts that are not covered by key element fields but have business semantic value (e.g., they can be cited as supplementary explanations, restrictions, exception handling instructions, processing prompts, or remarks during actual question-and-answer processes). Therefore, firstly, the executing entity can also remove business texts that are repeated in the aforementioned business text sequences and candidate text sequences, and divide the sequences according to whether the text paragraphs are continuous (i.e., consecutive texts in a paragraph are considered as one candidate text subsequence, and non-consecutive texts in a paragraph with a character count greater than or equal to a preset character threshold are considered as a separate candidate text subsequence), thus obtaining each candidate text subsequence.

[0036] The third step involves generating dialogue candidate information based on the aforementioned candidate text subsequences. In practice, for each candidate text subsequence, the executing entity first performs semantic embedding on the candidate text subsequence using the semantic encoding model to generate candidate text semantic vectors. Then, the executing entity can uniformly encapsulate the candidate text subsequences, the corresponding candidate text semantic vectors, and the corresponding text location information to construct dialogue candidate information that may contain candidate semantics.

[0037] The fourth step involves constructing multi-dimensional indexes for each of the aforementioned dialogue candidate information and each dialogue business information to generate various multi-dimensional index information. In practice, firstly, the aforementioned execution entity can use a keyword extraction algorithm to extract keywords from the business text subsequences included in each dialogue candidate information and the candidate text subsequences included in each dialogue candidate information, obtaining each business text keyword and each candidate text keyword. For example, the keyword extraction algorithm can be the TF-IDF (Term Frequency–Inverse Document Frequency) algorithm or the TextRank keyword extraction algorithm. Then, for each dialogue business information, the aforementioned execution entity can determine the extracted corresponding business text keywords as keyword indexes, the business text semantic vectors contained in the dialogue business information as semantic indexes, and the key element fields contained in the dialogue business information as field indexes to generate corresponding multi-dimensional index information. Finally, for each dialogue candidate information, the aforementioned execution entity can determine the extracted corresponding candidate text keywords as keyword indexes, the candidate text semantic vectors contained in the dialogue candidate information as semantic indexes, and the preset fields (e.g., "supplementary explanation") as field indexes to generate corresponding multi-dimensional index information.

[0038] The fifth step involves identifying the generated dialogue service information and the aforementioned dialogue candidate information as dialogue reference information, thus obtaining a set of dialogue reference information. In practice, the executing entity can identify both the generated dialogue service information and the dialogue candidate information as dialogue reference information to obtain a set of dialogue reference information.

[0039] It should be noted that after extracting elements from multi-source business data, relying solely on structured element information as the retrieval basis can easily overlook the semantic context of elements and non-field-type descriptive information. This leads to lower retrieval efficiency, incomplete reference information, and semantic matching bias in the semantic retrieval and knowledge referencing process of the dialogue system, thereby reducing the accuracy of dialogue responses and increasing retrieval response time. Therefore, this application associates text element information with business text subsequences representing their context, performs semantic encoding on the business text subsequences to generate semantic vectors, and constructs a multi-dimensional index structure containing dialogue business information and dialogue candidate information, thus forming a unified set of dialogue reference information from structured element information and unstructured semantic information. Through a multi-dimensional retrieval mechanism of field indexing, keyword indexing, and semantic indexing, field matching and semantic similarity calculation can be performed simultaneously during user dialogue, thereby improving the retrieval efficiency and matching accuracy of dialogue reference information, and reducing retrieval response time and erroneous reference rate.

[0040] Step 103: In response to determining that the user has initiated a dialogue for the target business scenario, based on the real-time received user dialogue text, the following dialogue process steps are executed: Step 1031: Based on historical dialogue state information, perform dialogue state recognition on the received user dialogue text to generate real-time dialogue state information.

[0041] In some embodiments, the aforementioned execution entity can identify the dialogue state of the received user dialogue text based on historical dialogue state information to generate real-time dialogue state information. In practice, the aforementioned execution entity can determine that the user has initiated a dialogue by responding to the dialogue text sent by the user terminal regarding the aforementioned target business scenario (i.e., policy consultation or policy Q&A). The aforementioned historical dialogue state information can be relevant information recorded in the previous round of dialogue. The aforementioned previous round of dialogue can be the most recent dialogue that occurred within the past 30 seconds, based on the sending time of the user's dialogue text.

[0042] In practice, firstly, the aforementioned execution entity can perform word segmentation, intent recognition, and text entity extraction on the received user dialogue text to generate user intent information and text entity information as dialogue state information. Then, the aforementioned execution entity can merge the various entity information included in the historical dialogue state information with the various entity information included in the current dialogue state information to update the current dialogue state information and obtain real-time dialogue state information.

[0043] In some optional implementations of certain embodiments, the aforementioned execution entity may generate real-time dialogue state information by performing dialogue state identification on the received user dialogue text based on historical dialogue state information through the following steps: The first step is to perform text segmentation on the aforementioned user dialogue text to obtain a text segmentation sequence. In practice, the executing entity can use the aforementioned natural language processing tool to perform text segmentation on the aforementioned user dialogue text to obtain a text segmentation sequence. For example, the aforementioned user dialogue text could be "What materials are needed for the application for high-tech enterprise status in XX city?", and the resulting text segmentation sequence after segmentation could be (XX city / high-tech enterprise / application / needs / what / materials).

[0044] The second step involves performing intent recognition on the aforementioned text segmentation sequence to generate user dialogue intent information. In practice, the executing entity can use a pre-trained intent recognition model to process the intent of the text segmentation sequence. For example, this intent recognition model could be a BERT classification model, a RoBERTa classification model, or other deep neural network models. The intent recognition model can output one or more predefined user intent categories and generate a corresponding confidence value for each category. The executing entity can then select the user intent category with the highest confidence value as the user dialogue intent information. For example, the user intent categories output by the intent recognition model could include, but are not limited to: deadline query type, materials list query type, reward amount query type, processing procedure query type, qualification assessment type, and specific policy consultation type.

[0045] The third step involves filling in the pre-built user slot information based on the aforementioned text segmentation sequence, thus updating the user slot information. This user slot information can consist of predefined slots representing necessary entities for dialogue within the user's conversation text, with each slot initially containing an empty value. For example, the user slot information could be ("Region": empty, "Policy Name": empty, "Company Name": empty, "Amount": empty, "Company Type": empty, "Time": empty). Empty values ​​represent the corresponding entities to be filled.

[0046] In practice, firstly, the aforementioned execution entity can use the sequence labeling model to identify the tags of each term in the text segmentation sequence, thereby recognizing entity information related to the aforementioned business scenario. The model will output corresponding slot tags for each term in the text sequence, such as region slots, policy name slots, or enterprise type slots. Then, the execution entity can fill the corresponding terms into a predefined user slot information structure according to the recognition results, thereby updating the user slot information.

[0047] The fourth step is to generate missing dialogue slot information based on the user dialogue intent information and the updated user slot information. For example... Figure 2 As shown, this includes predefined user intent categories 201 and pre-constructed user slot information 202. Each user intent category 201 corresponds to a slot in the user slot information 202, specifically represented by connecting lines. For example, the first user intent category corresponds to slots 1, 2, 5, and 6, indicating that the first user intent category represented by the user dialogue text requires entity information corresponding to slots 1, 2, 5, and 6 to be realized. In practice, the executing entity can determine the missing slots corresponding to the user intent categories in the user slot information as missing dialogue slot information, based on the user intent categories and corresponding slots contained in the user dialogue intent information.

[0048] The fifth step is to determine the above-mentioned user dialogue intent information, updated user slot information, and above-mentioned missing dialogue slot information as dialogue state information.

[0049] The sixth step involves updating the historical dialogue state information to obtain real-time dialogue state information. In practice, the executing entity can fill empty slots identified in the current round with the slots already filled in the historical dialogue state information (i.e., fill slots that are empty in the current round but not empty in previous rounds), thereby forming real-time dialogue state information containing multi-round dialogue context information.

[0050] Step 1032: Based on the real-time dialogue state information, select dialogue strategy information that meets the state matching conditions from the pre-built dialogue strategy information set as the target dialogue strategy information.

[0051] In some embodiments, the execution entity can select dialogue strategy information that meets the state matching conditions from a pre-built dialogue strategy information set based on the real-time dialogue state information. The dialogue strategy information in the aforementioned dialogue strategy information set can be a string representing the corresponding dialogue strategy. The aforementioned state matching conditions can be used to determine whether the confidence level corresponding to the user intent category contained in the user dialogue intent information meets the confidence level condition, and to select the corresponding dialogue strategy information according to whether the dialogue missing slot information included in the aforementioned real-time dialogue state information is empty. The aforementioned dialogue strategy information set is as follows: Figure 3 As shown, S01 to S12 represent different dialogue strategies. For example, S01 to S07 represent follow-up questioning strategies, respectively asking for entity information corresponding to slots 1 to 7. S08 represents a conditional retrieval strategy, indicating that the current dialogue already has the necessary entity information, allowing retrieval of relevant dialogue reference information and generation of a dialogue response. Furthermore, S09 to S12 can be custom dialogue strategies, including but not limited to intent clarification, outputting a processing entry point, refusing to answer and explaining the reason, and ending the dialogue. In practice, the executing entity can determine whether the current dialogue lacks relevant slot information based on the missing slot information included in the real-time dialogue status information. If slots are missing, the corresponding follow-up questioning strategy is selected. If no slots are missing, a conditional retrieval strategy can be selected.

[0052] In some optional implementations of certain embodiments, the aforementioned execution entity may select dialogue policy information that satisfies the dialogue state conditions as the target dialogue policy information from a pre-built set of dialogue policy information based on the aforementioned real-time dialogue state information through the following steps: The first step, in response to determining that the user dialogue intent information included in the above real-time dialogue state information meets the confidence condition, is to perform the following steps: The first sub-step involves, in response to determining that the missing dialogue slot information included in the aforementioned real-time dialogue state information is not empty, selecting dialogue strategy information that satisfies the slot follow-up condition from the aforementioned dialogue strategy information set as the target dialogue strategy information, based on the missing dialogue slot information included in the aforementioned real-time dialogue state information. The aforementioned confidence condition can be that the confidence level corresponding to the user intent category included in the aforementioned user dialogue intent information is greater than or equal to a preset confidence threshold (e.g., 0.6). In practice, in response to determining that the missing dialogue slot information included in the aforementioned real-time dialogue state information is not empty, i.e., at least one necessary entity information is missing in the current dialogue, the executing entity can select dialogue strategy information representing the missing slot as the target dialogue strategy information from the aforementioned dialogue strategy information set, based on the missing slots represented by the missing dialogue slot information included in the aforementioned real-time dialogue state information.

[0053] The second sub-step involves, in response to determining that the dialogue missing slot information included in the aforementioned real-time dialogue state information is empty, selecting dialogue strategy information representing conditional retrieval from the aforementioned dialogue strategy information set as the target dialogue strategy information. Here, the aforementioned conditional retrieval dialogue strategy indicates that the current dialogue already possesses the necessary entity information, enabling the retrieval of relevant dialogue reference information and the generation of dialogue responses.

[0054] The second step involves selecting dialogue strategy information representing intent clarification from the aforementioned dialogue strategy information set, in response to the determination that the user dialogue intent information included in the aforementioned real-time dialogue status information does not meet the confidence threshold. Therefore, when the confidence level corresponding to the user intent category included in the user intent information is less than the aforementioned preset confidence threshold, it indicates that the dialogue intent contained in the current user dialogue text is ambiguous, and intent clarification is required to determine the user's true dialogue intent.

[0055] Step 1033: Generate dialogue response information for the user's dialogue text based on the pre-trained large language model, target dialogue strategy information, real-time dialogue state information, dialogue reference information set, and corresponding multi-dimensional index information.

[0056] In some embodiments, the execution entity may generate dialogue response information for the user dialogue text based on a pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information.

[0057] In practice, firstly, in response to determining the target dialogue strategy information representation condition retrieval (i.e., the necessary entities to fulfill the user's intent), the aforementioned executing entity performs text retrieval on the aforementioned dialogue reference information set using the aforementioned real-time dialogue state information and various multi-dimensional index information to obtain the target dialogue reference information. Finally, the aforementioned executing entity can construct large language model prompts using the target dialogue reference information and input the large language model prompts into the aforementioned large language model to generate dialogue response information.

[0058] In some optional implementations of certain embodiments, the aforementioned execution entity can generate dialogue response information for the aforementioned user dialogue text through the following steps: based on a pre-trained large language model, the aforementioned target dialogue strategy information, the aforementioned real-time dialogue state information, the aforementioned dialogue reference information set, and the corresponding multi-dimensional index information. The first step, in response to the retrieval of the target dialogue strategy information representation conditions, is to perform the following response generation steps: The first sub-step involves performing a multi-dimensional search on the aforementioned dialogue reference information set based on the real-time dialogue status information and the corresponding multi-dimensional index information, to obtain the target dialogue reference information set.

[0059] The second sub-step involves constructing the large language model prompt text based on the aforementioned target reference information set and real-time dialogue state information. In practice, the executing entity can use the aforementioned target reference information set as evidence text, and fill the pre-constructed large language prompt word template with user slot information and user dialogue intent information, including the factual dialogue state information, to construct the large language model prompt text. The aforementioned large language model prompt word template may include user dialogue text, real-time dialogue state information, target dialogue reference information set, and output constraint instructions for the large language model. The aforementioned output constraint instructions can be used to limit the large language model to generate response content only based on the aforementioned target dialogue reference information set, avoiding the generation of responses that are irrelevant to the reference information or lack basis. For example, the aforementioned large language prompt word template could be: "The current user's response is [user dialogue text], their requirement is [user dialogue intent information], the user's basic information is [user slot information], and the relevant policy information is as follows: [target reference information set]. Please provide a correct dialogue response that meets the user's needs based on the provided relevant policy information and the user's basic information." In addition, the implementing entity can add response style control information to the prompt text, such as requiring a concise response, listing the basis for the conditions, or indicating the source of the reference.

[0060] The third sub-step involves inputting the aforementioned large language model prompt text into the large language model to generate dialogue response information for the aforementioned user dialogue text. This dialogue response information can be the generated response text. For example, the aforementioned large language model can be an LLaMA series model (e.g., LLaMA2, LLaMA3), a Qwen model, or a ChatGPT model.

[0061] In some optional implementations of certain embodiments, the aforementioned execution entity may perform multi-dimensional retrieval of the aforementioned dialogue reference information set based on the aforementioned real-time dialogue state information and the corresponding multi-dimensional index information through the following steps to obtain the target dialogue reference information set: The first step is to construct text retrieval request information based on the aforementioned real-time dialogue status information. This text retrieval request includes filtering conditions. In practice, the executing entity first constructs text retrieval request information corresponding to the current dialogue target, based on the user dialogue intent information and user slot information included in the real-time dialogue status information. This text retrieval request information can include a description of the retrieval target and filtering conditions used to limit the retrieval scope. The filtering conditions can consist of slots corresponding to user intent categories. Then, the executing entity converts the real-time dialogue status information into standardized text retrieval request information. For example, if the real-time dialogue status information is ("Intent Category": query application time, "Regional Slot": Suzhou, "Policy Type Slot": High-tech Enterprise), then the constructed text retrieval request information would be: ("Retrieval Target": query high-tech enterprise application and deadline, "Filter Conditions": Region = Suzhou, Policy Type = High-tech Enterprise).

[0062] The second step is to construct a text query statement based on the aforementioned text retrieval request information. In practice, the executing entity can convert the user intent category, slots, and filtering conditions in the text retrieval request information into natural language statements. Specifically, a pre-defined query statement generation template can be used to concatenate the user intent and the entity information corresponding to each slot to generate a semantically clear text query statement. The pre-defined query statement generation template can be a pre-defined natural language template that lacks relevant entity information in the dialogue text. The text query statement can be obtained by filling in the entity information corresponding to the relevant slots. For example, if the text retrieval request information is: ("Retrieval Target": query for high-tech enterprise application and deadline, "Filter Conditions": region = Suzhou, policy type = high-tech enterprise), the generated text query statement can be "Suzhou high-tech enterprise application deadline".

[0063] The third step is to extract keywords from the above text query statement to generate text query keywords. In practice, the executing entity can use the keyword extraction algorithm described above to extract keywords from the above text query statement to generate at least one text query keyword.

[0064] The fourth step is to perform semantic encoding on the above text query statement to generate a text query semantic vector. In practice, the executing entity can use the above text embedding model to perform semantic encoding on the above text query statement to obtain the text query semantic vector.

[0065] The fifth step involves retrieving the multi-dimensional index information corresponding to the aforementioned dialogue citation information set based on the filtering conditions, text query semantic vector, and query keywords included in the text retrieval request, thus obtaining a candidate dialogue citation information set. In practice, the executing entity can first use the filtering conditions to filter in the field index to remove dialogue citation information unrelated to the fields represented by the current filtering conditions. Then, the executing entity can perform keyword matching retrieval in each keyword index using the text query keywords to recall dialogue citation information related to the core query words (i.e., keyword overlap greater than or equal to 60%). Afterward, the executing entity can perform semantic similarity retrieval in the vector index using the text query semantic vector and cosine similarity to recall dialogue citation information semantically similar to the query (greater than or equal to 80%). Finally, the executing entity can merge and deduplicate the dialogue citation information obtained from the field retrieval results, keyword retrieval results, and semantic retrieval results to obtain a candidate dialogue citation information set.

[0066] The sixth step involves rearranging the candidate dialogue reference information set to obtain the target dialogue reference information set. In practice, for each candidate dialogue reference information in the set, the executing entity can determine the weighted sum of semantic similarity score, keyword matching degree, and field matching degree to obtain the corresponding reliability coefficient. Then, according to the determined reliability coefficient, the candidate dialogue reference information set is sorted, and the top k candidate dialogue reference information is selected as the target dialogue reference information.

[0067] It should be noted that in the knowledge retrieval process of a dialogue system, relying solely on single keyword matching or simple vector retrieval methods often fails to simultaneously consider structured condition constraints and semantic similarity matching. This can easily lead to an excessively broad range of search results or insufficient relevance, resulting in redundant candidate citation information, low retrieval accuracy, and semantic matching deviations. Consequently, this affects the accuracy of the information cited by the dialogue system when generating responses and increases the response time required for the retrieval and filtering process. Therefore, this application constructs text retrieval request information using real-time dialogue state information and further generates text query statements. Simultaneously, it utilizes structured filtering conditions, keyword matching, and semantic vector similarity to recall candidate citation information. After obtaining the set of candidate dialogue citation information, a reordering process is then used to rank the candidate citation information by relevance, thereby obtaining the target dialogue citation information set. This improves the retrieval efficiency and relevance of dialogue citation information and reduces unnecessary computational overhead and response time during the retrieval process.

[0068] Step 1034: Based on the dialogue response information, push a message to the user terminal and store the real-time dialogue status information as historical dialogue status information.

[0069] In some embodiments, the aforementioned executing entity can push messages to the user terminal based on the aforementioned dialogue response information, and store the aforementioned real-time dialogue status information as historical dialogue status information. In practice, the aforementioned executing entity can use a preset sensitive word replacement set to detect sensitive words in the aforementioned dialogue response information, replace existing sensitive words, push the replaced response text to the user terminal, and store the real-time dialogue status information as historical dialogue status information. The aforementioned sensitive word replacement set can be composed of various replacement sensitive word groups, each containing a sensitive word and a corresponding replacement word. For example, a replacement sensitive word group may contain the sensitive word "politics" and the corresponding replacement word "policy".

[0070] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a full-process intelligent data processing device based on a large model. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this large-model-based end-to-end data intelligent processing device can be specifically applied to various electronic devices.

[0071] like Figure 4 As shown, some embodiments of the large-model-based end-to-end data intelligent processing device 400 include: an element parsing unit 401, a dialogue reference construction unit 402, and a dialogue process processing unit 403. The element parsing unit 401 is configured to parse the collected multi-source business dataset according to pre-constructed key element information to generate a set of text element information groups, wherein each collected multi-source business data is applied to a target business scenario; the dialogue reference construction unit 402 is configured to construct dialogue references on the aforementioned set of text element information groups to generate a set of dialogue reference information, wherein each dialogue reference information corresponds to multi-dimensional index information; the dialogue process processing unit 403 is configured to, in response to determining that a user initiates a dialogue for a target business scenario, execute the following dialogue process processing steps based on the real-time received user dialogue text: according to historical... The process involves identifying the dialogue state of the received user dialogue text to generate real-time dialogue state information. Based on this real-time dialogue state information, a dialogue strategy that meets the state matching conditions is selected from a pre-built set of dialogue strategy information as the target dialogue strategy information. Then, based on a pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information, a dialogue response is generated for the user dialogue text. Finally, based on the dialogue response information, a message is pushed to the user, and the real-time dialogue state information is stored as historical dialogue state information.

[0072] It is understandable that the various units and references described in the large-model-based end-to-end data intelligent processing device 400 are related. Figure 1 The steps described in the method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the large-model-based end-to-end intelligent data processing device 400 and the units contained therein, and will not be repeated here.

[0073] The following is for reference. Figure 5 It shows a schematic diagram of the structure of an electronic device (e.g., a computing device) 500 suitable for implementing some embodiments of the present disclosure. Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0074] like Figure 5 As shown, the electronic device 500 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory 502 or a program loaded from a storage device 508 into a random access memory 503. The random access memory 503 also stores various programs and data required for the operation of the electronic device 500. The processing unit 501, the read-only memory 502, and the random access memory 503 are interconnected via a bus 504. An input / output interface 505 is also connected to the bus 504.

[0075] Typically, the following devices can be connected to the input / output interface 505: input devices 506 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 507 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 508 including, for example, magnetic tape, hard disk, etc.; and communication devices 509. Communication device 509 allows electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 An electronic device 500 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 5 Each box shown can represent a device or multiple devices as needed.

[0076] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 509, or installed from a storage device 508, or installed from a read-only memory 502. When the computer program is executed by the processing device 501, it performs the functions defined above in the methods of some embodiments of this disclosure.

[0077] It should be noted that, in some embodiments of this disclosure, the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0078] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0079] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: perform element parsing on the collected multi-source business dataset based on pre-constructed key element information to generate a set of text element information groups, wherein each of the collected multi-source business data is applied to a target business scenario; construct dialogue references on the aforementioned set of text element information groups to generate a set of dialogue reference information, wherein each dialogue reference information corresponds to multi-dimensional index information; and, in response to determining that a user terminal has initiated a dialogue for the target business scenario, execute the following dialogue flow processing steps based on the real-time received user dialogue text: root Based on historical dialogue state information, the received user dialogue text is subjected to dialogue state recognition to generate real-time dialogue state information. Based on the real-time dialogue state information, dialogue strategy information that meets the state matching conditions is selected from a pre-constructed dialogue strategy information set as the target dialogue strategy information. Based on the pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information, dialogue response information for the user dialogue text is generated. Based on the dialogue response information, a message is pushed to the user terminal, and the real-time dialogue state information is stored as historical dialogue state information.

[0080] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0081] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0082] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0083] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A full-process intelligent data processing method based on a large model, characterized in that, include: Based on the pre-constructed key element information, the collected multi-source business dataset is parsed to generate a set of text element information groups, in which each of the collected multi-source business data is applied to the target business scenario. Dialogue references are constructed on the set of text element information groups to generate a set of dialogue reference information, wherein each dialogue reference information corresponds to multi-dimensional index information; In response to determining that the user has initiated a dialogue for the target business scenario, the following dialogue processing steps are executed based on the real-time received user dialogue text: Based on historical dialogue state information, the received user dialogue text is used to identify the dialogue state in order to generate real-time dialogue state information. Based on the real-time dialogue state information, select dialogue strategy information that meets the state matching conditions from the pre-built dialogue strategy information set as the target dialogue strategy information. Based on the pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information, dialogue response information is generated for the user's dialogue text. Based on the dialogue response information, a message is pushed to the user's terminal, and the real-time dialogue status information is stored as historical dialogue status information.

2. The method according to claim 1, characterized in that, The step of parsing the collected multi-source business dataset based on pre-constructed key element information to generate a set of text element information groups includes: The multi-source business dataset is subjected to text extraction processing to generate a business text sequence, wherein each business text in the business text sequence corresponds to text location information; Based on the key element information, the business text sequence is processed by element extraction to generate information on each text element. Based on the key element fields included in the key element information, the generated text element information is classified to obtain a set of text element information groups, wherein each text element information group corresponds to a key element field.

3. The method according to claim 2, characterized in that, The step of extracting elements from the business text sequence based on the key element information to generate information on each text element includes: For each key element field included in the key element information, perform the following element processing steps: Based on the key element fields, element matching is performed on the business text sequence to obtain a candidate text sequence; For each candidate text included in the candidate text sequence, feature annotation and extraction are performed on the candidate text to generate text feature information, wherein the generated text feature information corresponds to feature location information.

4. The method according to claim 1, characterized in that, The step of identifying the dialogue state of the received user dialogue text based on historical dialogue state information to generate real-time dialogue state information includes: The user dialogue text is processed by text segmentation to obtain a text segmentation sequence; Intent recognition is performed on the text segmentation sequence to generate user dialogue intent information; Based on the text segmentation sequence, fill the pre-built user slot information with slots and update the user slot information. Based on the user dialogue intent information and the updated user slot information, generate dialogue missing slot information; The user dialogue intent information, the updated user slot information, and the missing slot information in the dialogue are determined as dialogue state information; Based on the historical dialogue status information, the dialogue status information is updated to obtain real-time dialogue status information.

5. The method according to claim 1, characterized in that, The step of selecting dialogue strategy information that meets the state matching conditions from a pre-built set of dialogue strategy information based on the real-time dialogue state information as the target dialogue strategy information includes: In response to determining that the user dialogue intent information included in the real-time dialogue state information meets the confidence condition, the following steps are performed: In response to determining that the dialogue missing slot information included in the real-time dialogue state information is not empty, based on the dialogue missing slot information included in the real-time dialogue state information, dialogue strategy information that satisfies the slot probing condition is selected from the dialogue strategy information set as the target dialogue strategy information. In response to determining that the dialogue missing slot information included in the real-time dialogue state information is empty, dialogue strategy information representing the conditions is selected from the dialogue strategy information set as the target dialogue strategy information. In response to determining that the user dialogue intent information included in the real-time dialogue state information does not meet the confidence condition, dialogue strategy information representing intent clarification is selected from the dialogue strategy information set as the target dialogue strategy information.

6. The method according to claim 1, characterized in that, The step of generating dialogue response information for the user's dialogue text based on the pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information includes: In response to determining the target dialogue strategy information representation conditions for retrieval, the following response generation steps are performed: Based on the real-time dialogue status information and the corresponding multi-dimensional index information, the dialogue reference information set is searched in multiple dimensions to obtain the target dialogue reference information set. Based on the target reference information set, construct a large language model prompt text; The prompt text of the large language model is input into the large language model to generate dialogue response information for the user's dialogue text.

7. The method according to claim 6, characterized in that, The step of performing a multi-dimensional retrieval of the dialogue reference information set based on the real-time dialogue state information and the corresponding multi-dimensional index information to obtain the target dialogue reference information set includes: Based on the real-time dialogue status information, a text retrieval request is constructed, wherein the text retrieval request includes filtering condition information; Based on the text retrieval request information, construct a text query statement; Keyword extraction is performed on the text query statement to generate text query keywords; The text query statement is semantically encoded to generate a text query semantic vector; Based on the filtering conditions, text query semantic vector, and query keywords included in the text retrieval request, the multi-dimensional index information corresponding to the dialogue reference information set is retrieved to obtain the candidate dialogue reference information set. The candidate dialogue reference information set is rearranged to obtain the target dialogue reference information set.

8. A full-process intelligent data processing device based on a large model, characterized in that, include: The element parsing unit is configured to parse the collected multi-source business dataset based on pre-built key element information to generate a set of text element information groups, wherein each of the collected multi-source business data is applied to the target business scenario. The dialogue reference construction unit is configured to construct dialogue references on the set of text element information groups to generate a set of dialogue reference information, wherein each dialogue reference information corresponds to multi-dimensional index information. The dialogue process processing unit is configured to, in response to determining that a user initiates a dialogue for the target business scenario, perform the following dialogue process processing steps based on the real-time received user dialogue text: Based on historical dialogue state information, identify the dialogue state of the received user dialogue text to generate real-time dialogue state information; based on the real-time dialogue state information, select dialogue strategy information that meets the state matching conditions from a pre-built dialogue strategy information set as the target dialogue strategy information; based on a pre-trained large language model, the target dialogue strategy information, the real-time dialogue state information, the dialogue reference information set, and the corresponding multi-dimensional index information, generate dialogue response information for the user dialogue text; based on the dialogue response information, push a message to the user terminal, and store the real-time dialogue state information as historical dialogue state information.

9. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 7.

10. A computer-readable medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images