Product intention recognition method and system and computer readable storage medium

Abstract

The invention provides a product intention recognition method and system and a computer readable storage medium, in the method, after an inquiry request is responded, a historical inquiry case is matched from a preset case knowledge base and is injected into a preset cue word framework as an ICL example, and then first product intention analysis and knowledge base routing are completed based on a preset small parameter model; the problem that a large-parameter model is poor in response speed is avoided by means of the light-weight characteristic of a small-parameter model, the real-time interaction requirement of a passenger scene is met, the capability limitation of the small-parameter model during independent reasoning is made up through injection of an ICL example, and the product intention in oral inquiry can be accurately recognized. Meanwhile, on the basis of determining the first target inquiry product, the inquiry request is routed to the corresponding product knowledge base, so that the pressure of a subsequent question and answer retrieval link can be further reduced, the response efficiency of an overall question and answer link is guaranteed, and application limitation caused by lack of real-time performance of a large-parameter model is avoided.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a product intent recognition method, system, and computer-readable storage medium. Background Technology

[0002] Product intent recognition refers to accurately identifying the specific product mentioned in a user's query in an intelligent question-answering scenario, thereby ensuring the accuracy of the response. Currently, in intelligent question-answering scenarios, product intent recognition solutions for users often suffer from challenges. Because user descriptions of product-related questions are often colloquial, diverse, and ambiguous, traditional intent recognition models struggle to accurately identify the specific product type inquired about. This further leads to the problem that when the model performs intelligent question-answering based on the knowledge base corresponding to the query product, it may route to the wrong product knowledge base, resulting in low product intent recognition accuracy. While current solutions utilize large-parameter language models to address this issue, these models have long response times, cannot meet real-time customer needs, and are not widely applicable. Summary of the Invention

[0003] To address the aforementioned issues, this application provides a product intent recognition method, system, and computer-readable storage medium.

[0004] The embodiments of this application disclose the following technical solutions:

[0005] In a first aspect, embodiments of this application provide a product intent recognition method, including:

[0006] In response to an inquiry request, historical inquiry cases matching the inquiry request are determined from a preset case knowledge base.

[0007] The historical query cases are injected as ICL examples into the preset prompt word framework to obtain the reasoning prompt text;

[0008] Based on a preset small parameter model, the first product intent is analyzed through the inference prompt text to obtain the first target inquiry product, and the inquiry request is routed to the first product knowledge base corresponding to the first target inquiry product.

[0009] In one possible implementation, after obtaining the first target query product, the method further includes:

[0010] Based on the preset large parameter model, the inference result quality inspection is performed on the first target query product output by the preset small parameter model to obtain the inference quality inspection result.

[0011] Based on the reasoning quality inspection results and the first target query product, the preset case knowledge base or preset training dataset is updated; the preset training dataset is the training data set of the preset small parameter model.

[0012] The parameter scale of the preset large parameter model is larger than that of the preset small parameter model.

[0013] In one possible implementation, updating the preset case knowledge base or preset training dataset based on the inference quality inspection results and the first target query product includes:

[0014] If the reasoning quality inspection result indicates that the intention reasoning is correct, the query request, the first target query product, and the routing record for the query request are identified as update cases, and the update cases are hot-updated to the preset case knowledge base.

[0015] If the reasoning quality inspection result indicates that the intention reasoning is incorrect, the second product intention analysis is performed using the preset large parameter model and the reasoning prompt text to obtain the second target inquiry product and determine the second product knowledge base corresponding to the second target inquiry product.

[0016] The first target query product, the query request, the second target query product, and the routing records between the second target query product and the second product knowledge base are updated to the preset training dataset to obtain the updated training dataset.

[0017] In one possible implementation, the method further includes updating the preset training dataset with the first target query product, the query request, the second target query product, and the routing records between the second target query product and the second product knowledge base to obtain an updated training dataset.

[0018] Based on the updated training dataset, the preset small parameter model is trained using SFT.

[0019] In one possible implementation, the preset small parameter model is a model with a parameter scale of no more than 10B.

[0020] Secondly, embodiments of this application provide a product intent recognition system, including:

[0021] The case determination module is used to respond to an inquiry request and determine historical inquiry cases that match the inquiry request from a preset case knowledge base.

[0022] The prompt construction module is used to inject the historical query cases as ICL examples into the preset prompt word framework to obtain inference prompt text;

[0023] The intent analysis module is used to perform first product intent analysis based on a preset small parameter model and the inference prompt text to obtain the first target inquiry product, and to route the inquiry request to the first product knowledge base corresponding to the first target inquiry product.

[0024] In one possible implementation, the system further includes: an inference quality control module; the inference quality control module is specifically used for:

[0025] Based on the preset large parameter model, the inference result quality inspection is performed on the first target query product output by the preset small parameter model to obtain the inference quality inspection result.

[0026] Based on the reasoning quality inspection results and the first target query product, the preset case knowledge base or preset training dataset is updated; the preset training dataset is the training data set of the preset small parameter model.

[0027] The parameter scale of the preset large parameter model is larger than that of the preset small parameter model.

[0028] In one possible implementation, the inference quality control module includes: an update unit, which is specifically used for:

[0029] If the reasoning quality inspection result indicates that the intention reasoning is correct, the query request, the first target query product, and the routing record for the query request are identified as update cases, and the update cases are hot-updated to the preset case knowledge base.

[0030] If the reasoning quality inspection result indicates that the intention reasoning is incorrect, the second product intention analysis is performed using the preset large parameter model and the reasoning prompt text to obtain the second target inquiry product and determine the second product knowledge base corresponding to the second target inquiry product.

[0031] The first target query product, the query request, the second target query product, and the routing records between the second target query product and the second product knowledge base are updated to the preset training dataset to obtain the updated training dataset.

[0032] In one possible implementation, the system further includes: a model training module;

[0033] The model training module performs SFT training on the preset small parameter model based on the updated training dataset.

[0034] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any possible product intent recognition method of the first aspect of claim.

[0035] Compared to existing technologies, this application offers the following advantages: This application provides a product intent recognition method, system, and computer-readable storage medium. In this method, after responding to an inquiry request, historical inquiry cases are first matched from a preset case knowledge base and injected as ICL examples into a preset prompt word framework. Then, based on a preset small-parameter model, the first product intent analysis and knowledge base routing are completed. This leverages the lightweight nature of the small-parameter model to avoid the poor response speed of the large-parameter model, adapting to the real-time interaction needs of customer scenarios. Furthermore, the injection of dynamic ICL examples compensates for the limitations of the small-parameter model's ability to reason independently, enabling it to accurately identify product intent in conversational inquiries and thus determine the user's target inquiry product. This solves the problem of insufficient accuracy in product intent recognition by traditional small-parameter models. Simultaneously, based on determining the first target inquiry product, routing the inquiry request to the corresponding product knowledge base further reduces the pressure on subsequent retrieval and question-and-answer stages, ensuring the overall response efficiency of the question-and-answer chain and avoiding application limitations caused by the lack of real-time performance of large-parameter models. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating a product intent recognition method provided in an embodiment of this application;

[0038] Figure 2 A flowchart illustrating a reasoning quality inspection and data update method provided in an embodiment of this application;

[0039] Figure 3 This is a schematic diagram of the structure of a product intent recognition system provided in an embodiment of this application. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and accompanying drawings. It should be particularly noted that the embodiments described in this application are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0041] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0042] As described earlier, current solutions for identifying user product intent in intelligent question-answering scenarios often suffer from challenges due to the colloquial, diverse, and ambiguous nature of user descriptions of product-related questions. Traditional intent recognition models struggle to accurately identify the specific product type being inquired about. This further leads to the problem that when performing intelligent question-answering based on the knowledge base corresponding to the queried product, the model may route to the wrong product knowledge base, resulting in low accuracy in product intent recognition. While current solutions utilize large-parameter language models to address this issue, these models have long response times, failing to meet real-time customer needs and limiting their widespread application.

[0043] Based on this, embodiments of this application provide a product intent recognition method, system, and computer-readable storage medium. In this method, after responding to an inquiry request, historical inquiry cases are first matched from a preset case knowledge base and injected as ICL examples into a preset prompt word framework. Then, based on a preset small-parameter model, the first product intent analysis and knowledge base routing are completed. This approach leverages the lightweight nature of the small-parameter model to avoid the poor response speed of the large-parameter model, adapting to the real-time interaction needs of customer scenarios. Furthermore, the injection of dynamic ICL examples compensates for the limitations of the small-parameter model's ability to reason independently, enabling it to accurately identify product intent in conversational inquiries and thus determine the user's target inquiry product. This solves the problem of insufficient accuracy in product intent recognition by traditional small models. Simultaneously, based on determining the first target inquiry product, routing the inquiry request to the corresponding product knowledge base further reduces the pressure on subsequent retrieval and question-and-answer stages, ensuring the overall response efficiency of the question-and-answer chain and avoiding application limitations caused by the lack of real-time performance of large-parameter models.

[0044] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0045] See Figure 1 The figure is a flowchart illustrating a product intent recognition method provided in an embodiment of this application, specifically including the following steps:

[0046] S101: In response to the query request, determine the historical query cases that match the query request from the preset case knowledge base.

[0047] In intelligent question-answering scenarios that respond to user inquiries, the user's inquiry is often based on a specific product. In practice, users frequently use non-standard product names, such as sales names, aliases, or abbreviations. For example, in a question-answering scenario regarding insurance products, a user might abbreviate "Blue Medical Insurance Long-Term Medical Insurance Good Doctor Good Medicine Edition" as simply "Good Doctor Good Medicine." Traditional static prompts are insufficient to cover such complex scenarios. Therefore, to ensure accurate identification of the target product, this step uses a pre-set case knowledge base to determine historical inquiry cases, which are then used as small-parameter models to provide inference prompt text.

[0048] In this application, the pre-defined case knowledge base is a collection of historical query data that has been validated in practice. The stored content includes key information such as historical user query text, corresponding product intent recognition results, and correct knowledge base routing records, and it has dynamic update capabilities. Upon receiving a user's query request, a matching process for historical query cases is triggered immediately. This eliminates the need for complex reasoning based on large parameter models; relevant cases can be quickly located using lightweight retrieval logic. Specifically, in determining historical query cases from the pre-defined case knowledge base based on the user's query request, multiple logics, such as semantic similarity algorithms and product intent relevance analysis, can be combined to perform a deep comparison between the query request and historical cases. For example, when a user queries "the coverage of long-term medical care," historical cases containing core semantics such as "long-term medical care" and "Blue Medical Insurance Long-Term Medical Insurance" are prioritized for retrieval, rather than being limited to queries with identical wording. Simultaneously, the matching rules pre-set a relevance threshold, selecting only historical query cases with a relevance higher than the threshold as candidates. This ensures that the injected ICL examples have strong reference value and avoids irrelevant cases interfering with the reasoning and judgment of the pre-defined small parameter model.

[0049] S102: Inject the historical query cases as ICL examples into the preset prompt word framework to obtain the reasoning prompt text.

[0050] In this step, injecting historical query cases as ICL examples into the preset prompt word framework serves to capture the correlation between user statements and product intent for the preset small-parameter model. As introduced in the previous section on historical query cases, each case includes a complete triplet: the user's original query, the product intent identification result, and the corresponding knowledge base routing target. The preset prompt word framework provides the corresponding reasoning guidance structure, including model task instructions such as "Please identify the insurance product intent in the user's query and match it to the corresponding knowledge base," a case display area, and the current user's query input field. In practice, after receiving the user's current query, the small model first uses the knowledge base retrieval mechanism to match historical cases with the highest semantic similarity and the closest correlation to product intent from the case library. Typically, 1-3 core cases are selected to balance reference value and reasoning efficiency, avoiding excessive cases that could distract the model. These cases are then injected into the case area of ​​the preset prompt word framework in a unified format, forming a complete reasoning prompt text together with the guidance and the current user's query statement, providing the small model with a clear reasoning data foundation.

[0051] In this step, the inference hint text helps the pre-defined small-parameter model quickly understand the true product intent behind ambiguous expressions. When a user submits a simplified query like "long-term medical care," if the injected inference hint text contains a record such as "User asks: Long-term medical care; Identification result: Blue Health Insurance Long-term Medical Insurance (20 years); Routing to: Blue Health Insurance 20-year Long-term Medical Insurance Knowledge Base," the small model can directly reference past correct inference logic without starting from scratch, improving recognition accuracy and ensuring real-time response speed. Furthermore, these examples are all from real-world business scenarios, covering various expressions such as sales names, standard names, and aliases for different products, enabling the inference hint text to accurately meet the business needs of multiple products and effectively reduce routing errors caused by product name confusion.

[0052] S103: Based on the preset small parameter model, perform first product intent analysis through the inference prompt text to obtain the first target inquiry product, and route the inquiry request to the first product knowledge base corresponding to the first target inquiry product.

[0053] The core of the process of conducting first-product intent analysis based on a pre-set small-parameter model is to accurately capture the target product indicated in the user's inquiry by leveraging reasoning prompts incorporating historical cases, while ensuring real-time response. In this embodiment, the pre-set small-parameter model is a model with a parameter scale of no more than 10B (a 7B model can be used in actual scenarios). During the analysis, the pre-set small-parameter model first parses the guiding instructions and case demonstrations in the reasoning prompts to establish a mapping logic between colloquial expressions and the standard name of the target product. Then, it focuses on the user's current inquiry request, extracting key product-related information sentence by sentence. Whether it is an abbreviation, alias, sales name, or vague description, the model can rely on the experience accumulated in the cases to eliminate semantic ambiguity and accurately lock onto the core product. For example, when a user asks for information about the coverage of long-term medical insurance, the model will refer to the identification logic of similar cases in the prompt text to clarify that the long-term medical insurance in the inquiry is not a generalized concept, but a specific product "Blue Medical Insurance Long-Term Medical Insurance (20 years)". It will then identify it as the first target product for inquiry. The entire analysis process closely revolves around the needs of the business scenario, which not only avoids the recognition limitations of small models that rely solely on prompt words, but also maintains a high processing speed.

[0054] Once the primary target product for inquiry is identified, the knowledge base routing process is automatically initiated. Based on the pre-defined mapping relationship between products and knowledge bases, the user's inquiry request is routed to the primary product knowledge base corresponding to the primary target product. In one possible implementation, the knowledge base can be pre-segmented by product dimension, with each product having its own independent product knowledge base. This routing method fundamentally avoids the search pressure caused by the mixing of knowledge from multiple products and reduces answer errors caused by interference between different product information. The routing process requires no manual intervention and is entirely automatically matched by the standard identifier of the primary target product. For example, once the target product is identified as "Blue Medical Insurance Long-Term Medical Insurance Good Doctor Good Medicine Edition," the user is directly routed to the knowledge base corresponding to that product, ensuring that subsequent search steps can quickly locate relevant product information.

[0055] In the product intent recognition method of this embodiment, after determining the first target query product and completing the routing process, to ensure the correctness of the reasoning logic of the preset small parameter model, it is necessary to perform quality checks on the reasoning results of the preset small parameter model through the preset large parameter model. Specifically, this involves verifying whether the first target query product determined by the preset small parameter model belongs to the actual target product indicated in the query request. If the reasoning result of the preset small parameter model is determined to be correct, all relevant data from this reasoning by the small parameter model can be used as an update case and updated to the preset case knowledge base to improve the response speed of the preset small parameter model. Conversely, if the reasoning result is determined to be incorrect, the preset large parameter model can be used to determine the correct recognition result, and both the correct and incorrect recognition results can be included in the training dataset of the preset small parameter model. This facilitates subsequent fine-tuning and training of the small parameter model, thereby optimizing the model's performance in recognizing product intent.

[0056] It is understandable that, in order to ensure the real-time nature of the question-and-answer response, the process of identifying product intent through a preset small parameter model in this application embodiment is often carried out in a real-time online process. However, since the analysis process of the preset large parameter model is relatively long, the process of quality inspection of the inference results through the preset large parameter model is executed in the backend offline process.

[0057] Next, with reference to the accompanying drawings of specific process implementation examples, we will introduce the execution process of performing inference quality checks through a preset large parameter model, and then updating the preset case knowledge base or training dataset.

[0058] See Figure 2 The figure is a flowchart illustrating a reasoning quality inspection and data update method provided in an embodiment of this application, specifically including the following steps:

[0059] S201: Based on the preset large parameter model, perform reasoning result quality inspection on the first target query product output by the preset small parameter model to obtain the reasoning quality inspection result.

[0060] In this embodiment, a pre-defined large-parameter model is used for inference quality control of the first target query product output by the pre-defined small-parameter model. In one possible implementation, the pre-defined large-parameter model is typically a high-performance model at the 72-bit level. Leveraging its powerful semantic understanding and reasoning capabilities, it verifies the target query product identified by the pre-defined small-parameter model. The quality control process focuses on determining whether the product intent output by the small model accurately matches the user's actual needs and whether it precisely corresponds to the correct product category. This step is performed offline, which does not affect the online real-time response efficiency and fully utilizes the high accuracy advantage of the large model. It compensates for potential biases in the small model when handling colloquial, diverse, and ambiguous user queries, providing a reliable basis for subsequent system optimization.

[0061] S202: Update the preset case knowledge base or preset training dataset based on the inference quality inspection results and the first target query product; the preset training dataset is the training data set of the preset small parameter model.

[0062] Based on the aforementioned reasoning quality control results and the primary target query product, the dynamic update process of the preset case knowledge base or preset training dataset is automatically initiated. If the quality control results show that the primary target query product output by the small-parameter model is accurate, this valid case will be integrated into the preset case knowledge base through hot updates, becoming the ICL input resource for subsequent online reasoning by the small-parameter model. This allows the small-parameter model to learn from more high-quality historical cases when processing similar user queries, continuously improving the accuracy of real-time recognition. If the quality control finds that the recognition results of the small model have deviations, these erroneous examples will be promptly included in the preset training dataset. The training dataset is specifically designed as targeted training material for the preset small-parameter model. In this way, the preset case knowledge base can always maintain real-time updates of cases, and the training dataset can accurately accumulate erroneous examples, laying the foundation for iterative optimization of the small model. Ultimately, this builds a closed loop of continuous learning, driving the entire system to continuously improve its intent recognition and routing capabilities in multi-product question-answering scenarios.

[0063] S203: Determine whether the reasoning quality control result is that the reasoning is correct;

[0064] S204: If the reasoning quality inspection result is that the intention reasoning is correct, the query request, the first target query product, and the routing record for the query request are determined as update cases, and the update cases are hot-updated to the preset case knowledge base;

[0065] S205: If the reasoning quality inspection result is that the intention reasoning is incorrect, perform second product intention analysis through the preset large parameter model and the reasoning prompt text to obtain the second target inquiry product, and determine the second product knowledge base corresponding to the second target inquiry product;

[0066] S206: Update the first target query product, the query request, the second target query product, and the routing records between the second target query product and the second product knowledge base to the preset training dataset to obtain the updated training dataset.

[0067] Furthermore, when the judgment result indicates that the intent reasoning is correct, it means that the preset small-parameter model, based on historical similar cases and the ICL mechanism, accurately identified the product intent in the user's query request and successfully completed the routing to the corresponding product knowledge base. At this point, the complete and valid interaction information is integrated into an updated case, specifically including the user's original query request, the first target query product output by the small model, and the routing record between the target product and the corresponding knowledge base. This updated case is then injected into the preset case knowledge base through hot updating. This hot updating mechanism ensures that the case library can absorb high-quality interaction data in real time, allowing the subsequent small-parameter model to have richer and more realistic historical cases as ICL input when processing similar queries, further enhancing the adaptive effect of dynamic ICL and helping the small model better cope with conversational and diverse product queries.

[0068] If the reasoning quality control result indicates that the intent reasoning is incorrect, a correction and data accumulation process needs to be initiated. The strong reasoning capabilities of the pre-set large-parameter model are used to compensate for the recognition bias of the small model. The pre-set large-parameter model is invoked, and combined with the reasoning prompt text, the user's query request is re-analyzed for the second product intent to determine the user's true product needs, ultimately obtaining the accurate second target query product. Simultaneously, the corresponding second product knowledge base is identified, completing the correct route matching. Based on this, key information from this interaction is integrated, including the first target query product initially output by the small-parameter model (incorrect result), the user's original query request, the second target query product corrected by the large model (correct result), and the correct route record between the second target query product and the second product knowledge base. This information is then updated to the pre-set training dataset, forming a complete set of error examples, resulting in the updated training dataset. These high-quality error examples, containing both incorrect and correct comparisons, will provide material for subsequent targeted training of the pre-set small-parameter model, thereby continuously improving the intent recognition accuracy of the pre-set small-parameter model.

[0069] Based on this, after obtaining the updated training dataset, the pre-defined small-parameter model can be trained using SFT. The core of this approach lies in using the control samples in the training dataset to specifically address the shortcomings of the small-parameter model in product intent recognition. The updated training dataset gathers a large number of high-quality samples containing complete interaction chains. Each sample covers the user's original query request, the initially incorrect first target query product output by the small model, the accurate second target query product after correction by the large-parameter model, and the correct routing records between the second target query product and the corresponding second product knowledge base. These samples focus on complex scenarios such as colloquial product aliases, abbreviations, and sales names that the small-parameter model previously failed to accurately identify, providing solid support for targeted training.

[0070] During training, the SFT fine-tuning technique allows the small model to learn the causes of errors and the correct recognition logic from clearly labeled control samples, gradually enhancing its sensitivity to diverse product descriptions and optimizing the intent reasoning algorithm. This continuously improves the accuracy of product intent recognition and knowledge base routing while maintaining responsiveness. This training method eliminates the need to reconstruct the model architecture; it focuses on targeted optimization based solely on real-world error examples accumulated in actual business scenarios. This reduces the development cost of model iteration and allows the small model to quickly adapt to new product scenarios and user query habits. Ultimately, it improves the intent recognition accuracy of the pre-defined small-parameter model while maintaining its lightweight and high-response speed advantages, further perfecting the system's continuous learning loop.

[0071] This application provides a product intent recognition method. In this method, after responding to an inquiry request, historical inquiry cases are first matched from a preset case knowledge base and injected as ICL examples into a preset prompt word framework. Then, based on a preset small-parameter model, the first product intent analysis and knowledge base routing are completed. This method leverages the lightweight nature of the small-parameter model to avoid the poor response speed of the large-parameter model, adapting to the real-time interaction needs of customer scenarios. Furthermore, the injection of dynamic ICL examples compensates for the limitations of the small-parameter model's ability to reason independently, enabling it to accurately identify product intent in conversational inquiries and thus determine the user's target inquiry product. This solves the problem of insufficient accuracy in product intent recognition by traditional small models. Simultaneously, based on determining the first target inquiry product, routing the inquiry request to the corresponding product knowledge base further reduces the pressure on subsequent question-and-answer retrieval stages, ensuring the overall response efficiency of the question-and-answer chain and avoiding application limitations caused by the lack of real-time performance of large-parameter models.

[0072] The following describes a product intent recognition system provided in an embodiment of this application. The product intent recognition system described below can be referred to in correspondence with the product intent recognition method described above.

[0073] See Figure 3 The figure is a schematic diagram of the structure of a product intent recognition system provided in an embodiment of this application, which specifically includes the following modules:

[0074] The case determination module 100 is used to respond to an inquiry request and determine historical inquiry cases that match the inquiry request from a preset case knowledge base.

[0075] The prompt construction module 200 is used to inject the historical query cases as ICL examples into the preset prompt word framework to obtain inference prompt text;

[0076] The intent analysis module 300 is used to perform first product intent analysis based on a preset small parameter model and the inference prompt text to obtain a first target inquiry product, and to route the inquiry request to the first product knowledge base corresponding to the first target inquiry product.

[0077] In one possible implementation, the system further includes: an inference quality control module; the inference quality control module is specifically used for:

[0078] Based on the preset large parameter model, the inference result quality inspection is performed on the first target query product output by the preset small parameter model to obtain the inference quality inspection result.

[0079] Based on the reasoning quality inspection results and the first target query product, the preset case knowledge base or preset training dataset is updated; the preset training dataset is the training data set of the preset small parameter model.

[0080] The parameter scale of the preset large parameter model is larger than that of the preset small parameter model.

[0081] In one possible implementation, the inference quality control module includes: an update unit, which is specifically used for:

[0082] If the reasoning quality inspection result indicates that the intention reasoning is correct, the query request, the first target query product, and the routing record for the query request are identified as update cases, and the update cases are hot-updated to the preset case knowledge base.

[0083] If the reasoning quality inspection result indicates that the intention reasoning is incorrect, the second product intention analysis is performed using the preset large parameter model and the reasoning prompt text to obtain the second target inquiry product and determine the second product knowledge base corresponding to the second target inquiry product.

[0084] The first target query product, the query request, the second target query product, and the routing records between the second target query product and the second product knowledge base are updated to the preset training dataset to obtain the updated training dataset.

[0085] In one possible implementation, the system further includes: a model training module;

[0086] The model training module performs SFT training on the preset small parameter model based on the updated training dataset.

[0087] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a computer-readable storage medium storing computer instructions for causing the computer to execute the product intent recognition method as described in any of the above embodiments.

[0088] The computer-readable media in this application embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0089] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the product intent recognition method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0090] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the system, method, and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The system, method, and medium embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated 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 modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0091] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A product intent recognition method characterized by, The method comprises the following steps: in response to an inquiry request, determining a historical inquiry case matching the inquiry request from a preset case knowledge base according to the inquiry request; injecting the historical inquiry case into a preset prompt word framework as an ICL example to obtain reasoning prompt text; based on a preset small parameter model, performing first product intent analysis through the reasoning prompt text to obtain a first target inquiry product, and routing the inquiry request to a first product knowledge base corresponding to the first target inquiry product.

2. The method of claim 1, wherein, After obtaining the first target inquiry product, the method further comprises: based on a preset large parameter model, performing reasoning result quality inspection on the first target inquiry product output by the preset small parameter model to obtain a reasoning quality inspection result; updating the preset case knowledge base or a preset training data set according to the reasoning quality inspection result and the first target inquiry product; the preset training data set is a training data set of the preset small parameter model; wherein the parameter size of the preset large parameter model is larger than that of the preset small parameter model.

3. The method of claim 2, wherein, The updating of the preset case knowledge base or the preset training data set according to the reasoning quality inspection result and the first target inquiry product comprises: in the case that the reasoning quality inspection result is correct intent reasoning, determining the inquiry request, the first target inquiry product and the routing record for the inquiry request as an update case, and hot updating the update case to the preset case knowledge base; in the case that the reasoning quality inspection result is incorrect intent reasoning, performing second product intent analysis through the preset large parameter model and the reasoning prompt text to obtain a second target inquiry product, and determining a second product knowledge base corresponding to the second target inquiry product; updating the first target inquiry product, the inquiry request, the second target inquiry product and the routing record between the second target inquiry product and the second product knowledge base to the preset training data set to obtain an updated training data set.

4. The method of claim 2, wherein, After updating the first target inquiry product, the inquiry request, the second target inquiry product and the routing record between the second target inquiry product and the second product knowledge base to the preset training data set to obtain an updated training data set, the method further comprises: based on the updated training data set, performing SFT training on the preset small parameter model.

5. The method of claim 1, wherein, The preset small parameter model is a model with a parameter size not greater than 10B.

6. A product intent recognition system characterized by, The method comprises the following steps: a case determination module configured to determine a historical inquiry case matching an inquiry request from a preset case knowledge base according to the inquiry request in response to the inquiry request; a prompt construction module configured to inject the historical inquiry case into a preset prompt word framework as an ICL example to obtain reasoning prompt text; an intent analysis module configured to perform first product intent analysis through the reasoning prompt text based on a preset small parameter model to obtain a first target inquiry product, and route the inquiry request to a first product knowledge base corresponding to the first target inquiry product.

7. The system of claim 6, wherein, The system further comprises an inference quality inspection module, and the inference quality inspection module is specifically configured to: perform inference result quality inspection on the first target inquiry product output by the preset small parameter model based on a preset large parameter model, to obtain an inference quality inspection result; update the preset case knowledge base or the preset training data set according to the inference quality inspection result and the first target inquiry product; and the preset training data set is a training data set of the preset small parameter model. The parameter scale of the preset large parameter model is greater than that of the preset small parameter model.

8. The system of claim 7, wherein, The inference quality inspection module comprises an updating unit, and the updating unit is specifically configured to: in a case where the inference quality inspection result is correct intent inference, determine the inquiry request, the first target inquiry product, and a routing record for the inquiry request as an update case, and hot update the update case to the preset case knowledge base; in a case where the inference quality inspection result is incorrect intent inference, perform second product intent analysis through the preset large parameter model and the inference prompt text to obtain a second target inquiry product, and determine a second product knowledge base corresponding to the second target inquiry product; update the first target inquiry product, the inquiry request, the second target inquiry product, and a routing record between the second target inquiry product and the second product knowledge base to the preset training data set to obtain an updated training data set.

9. The system of claim 7, wherein, The system further comprises a model training module. The model training module performs SFT training on the preset small parameter model based on the updated training data set.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by a processor to implement the product intent recognition method in any one of claims 1-5. The program is executed by a processor to implement the product intent recognition method in any one of claims 1-5.

Images