Information processing method and apparatus

Abstract

Embodiments of the present specification provide an information processing method and device, wherein the information processing method comprises: determining a session field corresponding to original inquiry information, and querying a short-term memory cache according to the session field; determining target inquiry information according to a query result and the original inquiry information; analyzing the target inquiry information to determine initial inquiry structured information, and performing local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information; processing the initial inquiry structured information using a large language model associated with the target intent information to obtain an initial inquiry item sequence; performing standardization processing on initial inquiry items contained in the initial inquiry item sequence according to a field rule engine to obtain a standard inquiry item sequence; and constructing target inquiry structured information corresponding to the original inquiry information based on the standard inquiry item sequence.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of information processing technology, and in particular to information processing methods and apparatus. Background Technology

[0002] In the digital transformation of the trading industry, intermediary platforms, acting as crucial hubs connecting upstream and downstream buyers, heavily rely on instant messaging tools for daily business interactions. Inquiries submitted by buyers through instant messaging channels are generally unstructured, containing various modalities such as text messages, product screenshots, and table images, potentially mixing core data elements like specifications, material descriptions, and quantity requirements. This type of information exhibits significant industry characteristics, such as prevalent colloquial expressions, extensive use of abbreviations and aliases, mixed use of full-width and half-width characters, line breaks disrupting text flow, and frequent sentence splitting in multi-turn conversations. Existing processing mechanisms independently judge each message, failing to effectively identify messages containing only specifications, leading to their misclassification as non-purchasing information and missed business opportunities. Furthermore, the original fields of the extracted inquiry entries lack industry rule constraints, allowing noise such as abbreviations, aliases, non-standard units, and colloquial expressions to directly enter the product matching stage, causing issues like missing product name and material fields, and the inability to infer and complete specification text, resulting in redundant and inconsistent product information and low matching accuracy. Therefore, an effective solution is urgently needed to address the above problems. Summary of the Invention

[0003] In view of the above, embodiments of this specification provide an information processing method. One or more embodiments of this specification also relate to an information processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, an information processing method is provided, comprising: Determine the session field corresponding to the original inquiry information, query the short-term memory cache based on the session field, and determine the target inquiry information based on the query result and the original inquiry information; The target inquiry information is parsed to determine the initial inquiry structured information, and local intent recognition and intent reasoning are performed on the initial inquiry structured information to obtain the target intent information; The initial query structured information is processed using a large language model associated with the target intent information to obtain an initial query item sequence. The initial query items contained in the initial query item sequence are then standardized according to the field rule engine to obtain a standard query item sequence. Based on the standard query item sequence, construct the target query structured information corresponding to the original query information.

[0005] According to a second aspect of the embodiments of this specification, an information processing apparatus is provided, comprising: The determination module is configured to determine the session field corresponding to the original inquiry information, query the short-term memory cache based on the session field, and determine the target inquiry information based on the query result and the original inquiry information; The parsing module is configured to parse the target inquiry information to determine the initial inquiry structured information, and perform local intent recognition and intent reasoning on the initial inquiry structured information to obtain the target intent information; The processing module is configured to process the initial query structured information using a large language model associated with the target intent information to obtain an initial query item sequence, and to standardize the initial query items contained in the initial query item sequence according to the field rule engine to obtain a standard query item sequence. The construction module is configured to construct the target inquiry structured information corresponding to the original inquiry information based on the standard inquiry item sequence.

[0006] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the above-described information processing method.

[0007] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the information processing method described above.

[0008] According to a fifth aspect of the embodiments of this specification, a computer program product is provided, including a computer program or instructions that, when executed by a processor, implement the steps of the information processing method described above.

[0009] The information processing method provided in this embodiment determines the conversation field corresponding to the original inquiry information, queries the short-term memory cache based on the conversation field, and determines the target inquiry information based on the query result and the original inquiry information. It then parses the target inquiry information to determine initial inquiry structured information, performs local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information. Finally, it processes the initial inquiry structured information using a large language model associated with the target intent information to obtain an initial inquiry item sequence. The initial inquiry items contained in the initial inquiry item sequence are then standardized according to a field rule engine to obtain a standard inquiry item sequence. Based on the standard inquiry item sequence, the target inquiry structured information corresponding to the original inquiry information is constructed. This method achieves efficient and accurate structured information construction by introducing a short-term memory cache to process multi-turn dialogue contexts to identify effective business opportunities in sentence-splitting scenarios, and by combining local intent recognition and large language model processing. It has the advantages of effectively identifying effective business opportunities in unstructured inquiry information, reducing omissions, and improving the accuracy and efficiency of product matching. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an information processing method provided in one embodiment of this specification; Figure 2 This is a schematic diagram of an information processing method provided in one embodiment of this specification; Figure 3 This is a flowchart illustrating the processing procedure of an information processing method provided in one embodiment of this specification. Figure 4 This is a schematic diagram of the structure of an information processing device provided in one embodiment of this specification; Figure 5 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0011] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0012] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0013] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0014] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0015] This specification provides an information processing method, and also relates to an information processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0016] In practical applications, the inquiry information transmitted via instant messaging tools in the daily operation of steel trading intermediary platforms is highly unstructured, containing colloquial expressions, mixed use of specification abbreviations, full-width characters, and multimodal data input. The lack of context in multi-turn dialogues leads to inaccurate intent determination for individual messages. For example, in scenarios where a user sends specifications first and then adds inquiry terms, a specification-only message is misjudged as not indicating a purchase intent. The coupling of intent recognition and structured extraction modules results in wasted system resources; non-purchase messages still trigger calls to the large language model, increasing unnecessary computational overhead. The raw fields output by the large model lack industry rule cleaning mechanisms; abbreviations, aliases, non-standard units, and colloquial expressions in specification text are not reduced, preventing the product matching service from performing accurate searches based on standardized data. For example, in a steel B2B trading scenario, an end customer sends a message to the platform robot via instant messaging software: "What is the price of medium plate Q235B 16×2200? Is it in stock?" The message contained material, specifications, and inquiry terms, but because the system did not associate it with historical session context (no prior spec_only listening records), it judged the single message as not indicating a purchase intent. Simultaneously, the "16×2200" extracted from the large model was not converted to a standard specification representation, and the material field "Q235B" was not validated using the industry knowledge base, causing the product matching service to fail to identify a valid business opportunity. These issues caused interruptions in structured data generation, requiring manual intervention to correct the original message and affecting the continuity of the platform's automated processing flow.

[0017] In view of this, see Figure 1 , Figure 1 A flowchart of an information processing method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0018] Step S102: Determine the session field corresponding to the original inquiry information, query the short-term memory cache according to the session field, and determine the target inquiry information according to the query result and the original inquiry information.

[0019] Step S104: Parse the target inquiry information to determine the initial inquiry structured information, and perform local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information.

[0020] Step S106: Process the initial query structured information using the large language model associated with the target intent information to obtain an initial query item sequence. Then, standardize the initial query items contained in the initial query item sequence according to the field rule engine to obtain a standard query item sequence.

[0021] Step S108: Construct the target inquiry structured information corresponding to the original inquiry information based on the standard inquiry item sequence.

[0022] The information processing method provided in this embodiment can be applied to the processing of inquiry information in any trade scenario, such as steel trade, building materials trade, office supplies trading, etc. It is used to identify effective business opportunities in sentence-splitting scenarios by introducing short-term memory caching to process multi-turn dialogue context, and to achieve efficient and accurate structured information construction by combining local intent recognition and large language model processing. It has the advantages of effectively identifying effective business opportunities in unstructured inquiry information, reducing omissions, and improving the accuracy and efficiency of product matching.

[0023] Specifically, the original inquiry information refers to the initial, unstructured message submitted by the requester through an instant messaging platform. This information may include various forms of expression such as natural language text, abbreviations, and images. Session fields refer to a set of identifiers used to uniquely identify a specific dialogue or interaction session. Their function is to maintain context between multiple messages, ensuring that subsequent processing is based on the complete dialogue context. Short-term memory (STM) cache refers to a temporary storage mechanism used to save recent or context-related information for a short period. It helps in retrieving previous messages or relevant data in an ongoing dialogue. Target inquiry information refers to the comprehensive inquiry content derived from the original inquiry information, which may be supplemented by historical or contextual data in the STM cache. This information forms the basis for subsequent processing.

[0024] Correspondingly, initial query structured information refers to the preliminary structured representation of the target query information. This process transforms raw, unstructured input into a more organized format for automated processing, but may still contain non-standardized elements. Local intent recognition refers to the process of identifying the potential purpose or intent of the query using lightweight, local processing mechanisms. This process typically involves rule systems or small, dedicated models for rapid operation. Intent reasoning refers to a more advanced process for determining the precise intent of the query, typically involving more complex models or external services. It aims to classify queries into predefined categories, such as "purchase intent" or "general query." Target intent information refers to the final, confirmed query intent obtained from local intent recognition and more comprehensive intent reasoning. This information guides subsequent processing steps, particularly the selection of a large language model.

[0025] Correspondingly, a large language model refers to an artificial intelligence model trained on large amounts of text data, capable of understanding, generating, and processing human language. In this context, it is used to extract specific entities and relationships from inquiry information. The initial inquiry item sequence refers to a list of individual items or entities extracted by the large language model from the initial structured inquiry information. These items represent core components of the inquiry, such as product name, specifications, or quantity, but may still be in a raw, non-standardized format. A field rule engine refers to a system that applies predefined rules to specific data fields. It is used to transform, validate, or standardize data according to industry-specific or business-specific requirements. Standardization processing refers to the process of converting raw or inconsistent data into a uniform, predefined format. This process ensures data consistency and facilitates accurate matching and integration with other systems. A standard inquiry item sequence refers to a sequence of inquiry items after standardization processing by the field rule engine. These items conform to predefined industry standards or internal data formats. Target inquiry structured information refers to the final, fully structured representation of the original inquiry information. It is built upon the standardized inquiry item sequence and can be directly used in downstream business processes, such as matching with product catalogs or generating quotes.

[0026] Based on this, the purpose of determining the session field corresponding to the original query information, querying the short-term memory cache based on the session field, and determining the target query information based on the query results and the original query information is to establish and maintain the session context, especially in multi-turn interactions, by associating the current query with previously related messages. This helps to form a complete understanding of the user's intent. As one implementation, the system can simply use a unique identifier associated with the communication channel or sender to determine the session field. For example, a unique ID assigned to a chat window or user account can serve as a session field. Alternatively, the session field can consist of a combination of the sender's identifier and a timestamp, where the timestamp is truncated to a specific granularity (e.g., hour or day) to group messages within a short period. When the original query information is received, the system can use this determined session field to query a temporary storage area, such as a simple key-value store, to retrieve any previously stored information associated with the session. The retrieved information (if present) can be combined with the current original query information. This combination may involve simple concatenation of text or merging of data elements. The combined information is then designated as the target query information for subsequent processing. For example, if user A sends "Q235B medium-sized board" and then sends "What's the price of 16*2200? Is it in stock?", the system can use user A's ID as a session field. The system queries the cache, finds "Q235B medium-sized board", and combines it with "What's the price of 16*2200? Is it in stock?" to form the target price inquiry information.

[0027] Furthermore, in parsing the target inquiry information to determine the initial structured information of the inquiry, and performing local intent recognition and intent reasoning on the initial structured information to obtain target intent information, the aim is to transform potentially complex and unstructured target inquiry information into a preliminary structured format and determine the user's core intent. This step is crucial for guiding subsequent, more resource-intensive processing. As one implementation, the target inquiry information can be parsed using basic text processing techniques, such as word segmentation and part-of-speech tagging, to extract key phrases and entities. This process produces a preliminary structured representation, in which potential elements such as product names, quantities, or keywords are identified. For local intent recognition, a simple keyword matching algorithm can be used. A predefined list of keywords associated with different intents (e.g., "inquiry," "want to buy" represents a purchase intent; "hello," "thank you" represents a general greeting) can be used to quickly categorize inquiries. If a sufficient number of "want to buy" related keywords exist, a preliminary "want to buy" intent can be assigned. For intent reasoning, the initial structured inquiry information can be fed into a general machine learning model, such as a Support Vector Machine (SVM) or a Naive Bayes classifier, trained on datasets of various inquiry types. This model provides a more accurate prediction of inquiry intent. The results of local intent recognition and intent reasoning are then evaluated. For example, if both methods indicate a strong purchase intent, the target intent information is identified as a purchase inquiry. If discrepancies exist, conflicts can be resolved using simple majority voting or predefined priority rules to determine the final target intent information. For example, if the target inquiry information is "What is the price of medium-sized Q235B 16*2200? Is it in stock?", the parsing process might identify "medium-sized Q235B" and "16*2200" as potential product details, and "What is the price?" and "Is it in stock?" as inquiry phrases. Local intent recognition might label "What is the price?" and "Is it in stock?" as indicating purchase intent. Intent reasoning using a general classifier might also predict purchase intent. Combining these results, the target intent information is determined to be "purchase".

[0028] Furthermore, in processing the initial query structured information using a large language model associated with target intent information to obtain a sequence of initial query entries, and then standardizing the initial query entries contained in the sequence according to the field rule engine to obtain a standard query entry sequence, the aim is to leverage advanced language understanding capabilities for detailed information extraction, and then refine the extracted information into a consistent, standardized format suitable for downstream applications. As one implementation, a specific large language model or a specific configuration of a large language model is selected based on the target intent information (e.g., a confirmed purchase intent). This large language model is then used to process the initial query structured information. The task of the large language model is to identify and extract specific entities and their attributes from the text, such as product name, material, size, and quantity. The output of this process is the initial query entry sequence, i.e., the raw, extracted data points. Subsequently, the field rule engine is invoked to process this initial query entry sequence. The field rule engine contains a set of predefined rules. These rules may include simple lookup tables for aliases, regular expressions for pattern matching, or basic arithmetic operations for unit conversion. Each initial query entry is processed by these rules. For example, if the initial query entry for "Material" is "Q235B", and the rule engine has rules that map "Q235B" to its standard form "Q235B", then it remains unchanged. If an entry is "Thickness 16", then the rule might extract "16" as a numerical value. The application of these rules transforms the initial query entries into a standardized format, resulting in a standard query entry sequence. Variations in spelling, units, or wording within this sequence have been standardized to a common standard. For example, if the large language model extracts "medium plate", "Q235B", "16", and "2200" from "medium plate Q235B 16*2200", these constitute the initial query entry sequence. The field rule engine then checks "medium plate" against the product name dictionary, checks "Q235B" against the material dictionary, and ensures that "16" and "2200" are numerical values. If "16mm" is extracted, the rule might convert it to "16". This results in standardized sequences, such as {product_name: "medium plate", material: "Q235B", thickness: "16", width: "2200"}.

[0029] Finally, constructing the target structured inquiry information corresponding to the original inquiry information based on the standard inquiry entry sequence aims to assemble the fully processed and standardized inquiry entries into a final, comprehensive structured data object for direct use by downstream business systems. According to one embodiment, once the standard inquiry entry sequence is obtained, the system starts to construct the final target structured inquiry information. This involves organizing each standardized entry (e.g., product name, material, size) into a predefined data schema. The construction process may involve mapping these standardized entries to specific fields in a structured data format (e.g., JSON or XML). For example, the standardized product name "medium plate" is mapped to the "product_name" field, "Q235B" is mapped to the "material" field, and "16" and "2200" are mapped to the "thickness" and "width" fields respectively. Additional metadata related to the original inquiry, such as the original sender or timestamp, can also be integrated into this final structured information. This ensures that the complete context of the inquiry is retained and available. Thereby, the original unstructured inquiry is converted into a complete and machine-readable representation that can be used for integration with inventory management systems, pricing engines, or customer relationship management (CRM) platforms. For example, after obtaining the standard inquiry entry sequence {product_name: "medium plate", material: "Q235B", thickness: "16", width: "2200"}, the system creates a JSON object: {"product_name": "medium plate", "material": "Q235B", "specifications": {"thickness": "16", "width": "2200"}}. This object represents the target structured inquiry information.

[0030] By way of example, suppose demander user A submits an original inquiry information to an intelligent assistant via an instant messaging platform: "What is the price of medium plate Q235B 16*2200? Do you have it in stock?". This message is a clear purchase inquiry that includes specifications, material and inquiry terms. When the system receives the original inquiry information sent by user A, it first determines the session field corresponding to the information. The session field can be composed of the identification information of the intelligent assistant and the identification information of user A. Then, the system uses the session field to query the short-term memory cache. In this example, it is assumed that there is no historical information associated with the session field in the short-term memory cache, indicating that this is a new or independent inquiry. Accordingly, the query result is empty, and the system directly takes the original inquiry information as the target inquiry information, since there is no context information to be spliced. This step ensures that even a single message can be processed completely, while providing a basis for context splicing for multi-round conversations.

[0031] Further, after the system receives the target inquiry information "What is the price of medium plate Q235B 16*2200? Do you have it in stock?", it parses the information to determine the structured information of the initial inquiry. The parsing process may involve basic natural language processing technologies such as text word segmentation and entity recognition, identifying "medium plate", "Q235B", "16", "2200" and the like in the text as potential product attributes, and identifying "what is the price" and "do you have it in stock" as inquiry keywords. Thus, a preliminary structured representation is obtained. Then, the system performs local intent recognition on the structured information of the initial inquiry. For example, through a preset keyword list, the system detects explicit purchase intent keywords such as "what is the price" and "do you have it in stock", thereby preliminarily determining that this message has a purchase intent. Meanwhile, the system performs intent reasoning on the structured information of the initial inquiry. This may involve inputting the preliminary structured information into a lightweight intent classification model, which makes a more comprehensive judgment on the inquiry intent based on its training data. By combining the quick judgment of local intent recognition and the deeper analysis of intent reasoning, the system finally obtains the target intent information, that is, it confirms that this inquiry is a "purchase" intent. This hierarchical processing mechanism enables the system to quickly filter out non-purchase information and avoid unnecessary resource consumption.

[0032] Furthermore, since the target intent information is determined as "purchase", the system will call a large language model associated with the purchase intent to process the structured information of the initial inquiry. With its semantic understanding capability, the large language model can accurately extract the product name "medium plate", material "Q235B", thickness "16", width "2200", and inquiry keywords "what is the price" and "do you have it in stock" from "What is the price of medium plate Q235B 16*2200? Do you have it in stock?". These extracted original fields constitute the initial inquiry entry sequence. Then, the system inputs the initial inquiry entry sequence into a field rule engine for standardization processing. Conversion and mapping rules for the steel industry are preset in the field rule engine. For example, for the thickness "16", the rule engine checks whether unit conversion or format unification is required; for the material "Q235B", the rule engine compares it with a standard material library to ensure that it conforms to industry-standard naming. Through the processing of the field rule engine, each item in the initial inquiry entry sequence is converted into a unified and standardized standard format, so as to obtain a standard inquiry entry sequence, for example: {product_name: "medium plate", material: "Q235B", thickness: "16", width: "2200", inquiry_keywords: ["what is the price", "do you have it in stock"]}. This standardization process solves the problems of colloquialism and abbreviation that may exist in the extraction results of the large model, and provides a high-quality data basis for subsequent product matching.

[0033] Finally, the system uses the obtained sequence of standard inquiry entries to construct structured information for the target inquiry corresponding to the original inquiry information. This involves encapsulating standardized product names, materials, specifications, and other information, along with the confirmed purchase intent, according to a preset structured data format. For example, the system will combine information such as {product_name: "medium plate", material: "Q235B", thickness: "16", width: "2200", inquiry_keywords: ["What price", "Available"]} with the unique identifier and timestamp of the original inquiry into a complete JSON object or database record. Thus, the original unstructured instant messaging message is transformed into a highly structured, standardized, and machine-processable data entity that can be directly used in subsequent business processes such as opportunity matching, inventory lookup, or quotation generation.

[0034] In summary, by introducing short-term memory caching to process multi-turn dialogue context to identify effective business opportunities in sentence-segmentation scenarios, and by combining local intent recognition with large language model processing to achieve efficient and accurate structured information construction, this approach has the advantages of effectively identifying effective business opportunities in unstructured inquiry information, reducing omissions, and improving the accuracy and efficiency of product matching.

[0035] Furthermore, in real-world instant messaging scenarios, raw inquiry information often lacks clear contextual identifiers, making it difficult for the system to accurately distinguish the inquiry intentions of different requesters, different assistants, and different time periods. This hinders the effective construction of accurate session context, leading to deviations in subsequent intent recognition and information processing. Therefore, in this embodiment, determining the session field corresponding to the raw inquiry information includes: Receive the original inquiry information submitted by the demander to the intelligent assistant through an instant messaging platform; determine the assistant identification information corresponding to the intelligent assistant, the demander identification information corresponding to the demander, and the inquiry time information corresponding to the original inquiry information; and use the assistant identification information, the demander identification information, and the inquiry time information as the session field corresponding to the original inquiry information.

[0036] Specifically, receiving the original inquiry information submitted by the requester to the intelligent assistant through the instant messaging platform is the starting point of the entire information processing flow. Its purpose is to obtain the raw, unprocessed inquiry content sent by the user via instant messaging tools. This original inquiry information is a direct expression of the user's intent and may include multiple modalities such as text, images, and voice, forming the basis for all subsequent structured processing and intent recognition. The system can deploy a message listening module that continuously monitors the instant messaging platform's API interface or message queue. Once a requester sends a message to the intelligent assistant, it immediately captures and receives the message as the original inquiry information. Alternatively, the intelligent assistant itself can integrate message receiving functionality; when the requester interacts directly with the intelligent assistant, the intelligent assistant directly processes the received user input as the original inquiry information.

[0037] Determining the assistant identifier information corresponding to the smart assistant refers to obtaining a unique identifier for the smart assistant, used to distinguish different smart assistant instances or services. In multi-smart assistant collaborative work or multi-tenant scenarios, accurately identifying which smart assistant received the inquiry information is crucial, as it helps to assign the inquiry information to the correct processing logic and business process. When a smart assistant registers within the system, it can be assigned a globally unique ID, which is extracted as the assistant identifier information when a message is received; alternatively, when the instant messaging platform forwards a message to the smart assistant, it can attach the smart assistant's unique identity credential or account information to the message metadata, which the system can directly parse as the assistant identifier information.

[0038] Determining the demander's identifier refers to obtaining a unique identifier for the user (or client) initiating the inquiry, used to distinguish different users. When processing inquiry information, knowing which user initiated the inquiry is crucial for maintaining user session state, providing personalized services, and subsequent business follow-up. Instant messaging platforms typically assign a unique OpenID or UserID to each user. When the system receives a message, it can extract this information from the sender field as the demander's identifier. Alternatively, if the demander interacts with a smart assistant through a custom client, the client can include its registered user ID or session token in the request header or message body when sending the inquiry information; the system then parses this information as the demander's identifier.

[0039] Determining the inquiry time information corresponding to the original inquiry information refers to obtaining the specific timestamp when the original inquiry information was submitted or received. Time information plays a crucial role in constructing session context, determining message timeliness, handling multi-turn conversations, and performing debouncing (e.g., merging related messages within a specific time window). Instant messaging platforms typically include a timestamp field in the message body when sending messages; the system can directly read this field as the inquiry time information. Alternatively, if the instant messaging platform does not provide a precise timestamp, the system can immediately record the current system time as the inquiry time information upon receiving the original inquiry information.

[0040] Using the assistant identifier, the demander identifier, and the inquiry time information as the session field corresponding to the original inquiry information means combining these three types of identifier information into a unique session identifier. The session field is a set of key metadata used to uniquely identify and manage an interaction session between a specific user and a specific intelligent assistant. By combining the assistant identifier, demander identifier, and inquiry time information into a single session field, the system can establish a clear contextual association for each original inquiry message, thereby achieving isolation and management of different sessions and providing an accurate index for subsequent short-term memory cache queries. These three types of identifier information can be concatenated as strings, for example, using specific delimiters (such as "_" or "|") to connect them, forming a unique string as the session field; alternatively, these three types of identifier information can be encapsulated into a data structure (such as a JSON object or a hash table), and a hash operation can be performed on the data structure to generate a hash value as the session field, or the data structure can be directly used as the query key.

[0041] Based on this, when the system receives the original inquiry information submitted by the requester to the intelligent assistant through the instant messaging platform, the system first extracts or generates three key metadata from the interaction: the assistant's identifier information, the requester's identifier information, and the inquiry time information corresponding to the original inquiry information. These represent the message recipient, sender, and occurrence time, respectively. Subsequently, the system integrates these three types of identifier information to form a unique session field. This session field, as a composite key, can accurately identify a specific interaction session between a user and the intelligent assistant. In this way, the originally discrete and unstructured instant messaging messages are given clear identity and time attributes, transforming them into structured session units with contextual meaning. This session field, as a foundation, can provide accurate indexing for subsequent short-term memory cache queries, enabling the system to accurately locate the context of a specific session when handling multi-user, multi-round inquiries. For example, in the basic scheme, determining the session field corresponding to the original inquiry information is a prerequisite for querying the short-term memory cache. This solution combines assistant identification information, demander identification information, and inquiry time information into a session field, providing a precise and unique key-value pair for short-term memory cache queries. This enables the system to effectively manage and retrieve historical information for specific sessions, such as merging relevant inquiry information within a debouncing window, thereby ensuring the accurate construction of subsequent target inquiry information. This, in turn, improves the accuracy of initial inquiry structured information parsing, local intent recognition, and intent inference, ultimately guaranteeing the quality of target inquiry structured information construction. This refined session management mechanism effectively avoids logical errors caused by information confusion between different users or time periods, laying a solid foundation for the accuracy and robustness of the entire information processing flow.

[0042] For example, suppose in a B2B steel transaction scenario, buyer "A" sends an inquiry message to the intelligent assistant "AAA" via instant messaging software: "Do you have 10 tons of Q235B I-beams? Can you ship them today?". First, the system receives this original inquiry message submitted by buyer "A" to the intelligent assistant "AAA" via instant messaging software. Next, the system determines relevant identification information from the received message and its context. Specifically, the intelligent assistant may have been assigned a unique assistant identifier during system registration, such as "bot_steel_001". Buyer "A's" unique user ID on the communication platform, such as "user_jia_**", will be identified as the buyer's identifier. Simultaneously, the system records the precise time the message was received, such as "2024-03-15T10:30:00", as the inquiry time information. Finally, the system combines these three identified pieces of information to form a unique session field. For example, these can be concatenated into a single string using a specific delimiter: "bot_steel_001|user_jia_**|2024-03-15T10:30:00". This combined string serves as the session field corresponding to the original inquiry information, used for subsequent lookup and management in the short-term memory cache. In this way, even if "A" sends multiple messages within a short period, or other requesters also send messages to "AAA", the system can accurately identify the specific session context to which each message belongs through this unique session field.

[0043] In summary, the above processing establishes a multi-dimensional, unique session identifier for each original inquiry, thus resolving the issue of ambiguous contextual identifiers in traditional solutions. This explicit session field serves as a precise lookup key for the short-term memory cache, enabling the system to accurately identify and lock the context of a specific session when handling multi-user, multi-round inquiries, effectively avoiding logical errors caused by information confusion between different users or time periods. This not only improves the accuracy and efficiency of subsequent short-term memory cache queries but also provides a reliable foundation for the accurate construction of target inquiry information, the parsing of structured information from initial inquiries, and local intent recognition and inference. Consequently, it significantly enhances the accuracy and robustness of the entire information processing flow, especially in scenarios involving multi-round segmented inquiries and debouncing, effectively reducing misjudgments and information loss.

[0044] Furthermore, in instant messaging scenarios such as steel trading, users often have the habit of sending multiple rounds of segmented inquiries, that is, sending specification information first and then sending the inquiry intent. If the intent is determined solely by a single message, messages containing only specification information may be misjudged as non-purchasing information, resulting in the loss of contextual information. In addition, if all historical messages are spliced ​​together without filtering, they are easily interfered with by irrelevant information such as advertisements or sales, causing the system to misconceptions about purchasing intent, thereby reducing the accuracy of the constructed inquiry information. To address this, in this embodiment, the step of querying the short-term memory cache based on the session field and determining the target inquiry information based on the query result and the original inquiry information includes: The short-term memory cache is queried according to the session field, and the associated inquiry information is determined according to the query result, wherein the debouncing window corresponding to the session field is in running state; the associated inquiry information and the original inquiry information are concatenated to obtain the target inquiry information; wherein, the determination of the associated inquiry information in the short-term memory cache includes: receiving the associated inquiry information, and performing local intent recognition and intent reasoning on the associated inquiry information to obtain associated intent information; if it is determined according to the associated intent information that the associated inquiry information does not meet the conditions for constructing structured information, the associated inquiry information is stored in the short-term memory cache according to the session field corresponding to the associated inquiry information.

[0045] The session field refers to a set of identifying information used to uniquely identify an interaction between a user and the intelligent assistant. This field can consist of multiple dimensions, such as the assistant's identifier, the requester's identifier, and the request time information corresponding to the original request. Alternatively, it can be a combination of user ID, session ID, and timestamp to ensure accurate tracking and management of specific dialogue contexts in complex and ever-changing instant messaging environments. Short-term memory (SSM) caching is a storage mechanism for temporarily storing session-related information. This caching can be implemented using various technologies, such as high-speed read / write based on in-memory databases (e.g., Redis, Memcached) or persistence based on distributed file systems or local storage. Its core function is to quickly access historical request information associated with specific session fields. The debouncing window is a mechanism used to control the frequency of querying or updating the SSM cache within a specific time period. This window can be a preset time interval, such as 30 seconds or 1 minute. Within this interval, the system waits for subsequent relevant messages to arrive to determine whether information concatenation or cache status updates are necessary. This can be achieved by setting a timer or recording a timestamp in a cached entry and comparing the results. Associated query information refers to historical query information associated with the same session field and stored in a short-term memory cache before the arrival of the current original query information. This information may be incomplete or fragmented, such as messages containing only specifications and lacking a clear query intent. Original query information refers to the latest query message submitted by the user to the intelligent assistant through the instant messaging platform, awaiting processing. Target query information refers to the complete query information, after processing, used for subsequent parsing and intent recognition. This information may be the original query information itself or the result of concatenating the original query information with associated query information in the short-term memory cache. Local intent recognition refers to the preliminary intent judgment of query information using lightweight, efficient algorithms. This recognition can be based on preset keyword rules, regular expression matching, or simple statistical models (such as Naive Bayes classifiers or decision trees), aiming to quickly filter or identify intent information with obvious characteristics.

[0046] Correspondingly, intent reasoning refers to the in-depth analysis and judgment of intent in inquiry information through more complex models and algorithms. This reasoning can utilize Natural Language Processing (NLP) techniques, such as deep learning-based text classification models (e.g., FastText, BERT), or knowledge graph-based semantic matching, to identify users' more precise purchasing or non-purchasing intentions. Related intent information refers to the intent judgment results obtained after local intent recognition and intent reasoning of related inquiry information. This information is used to assess whether the related inquiry information has the conditions for constructing structured information. Not meeting the conditions for constructing structured information means that, after intent recognition and reasoning, the related inquiry information is judged to not contain sufficiently clear purchasing intent or key structured fields (e.g., product name, material, quantity), and cannot independently form a complete structured inquiry data that can be used for subsequent business opportunity processing. For example, a message containing only product specifications without inquiry terms. Storing to short-term memory cache refers to writing related inquiry information that does not meet the conditions for constructing structured information, along with its session fields and other metadata, into the short-term memory cache. Storage can be done in key-value pairs, where the key is a session field and the value is the associated query information and its status. Concatenation refers to merging the original query information with the associated query information retrieved from the short-term memory cache. This operation can be simple, such as by connecting text (e.g., adding newlines or specific delimiters), or it can be more complex, involving semantic fusion to form a more complete and context-rich text sequence.

[0047] Based on this, when the system receives a raw inquiry, it first queries the short-term memory cache based on the corresponding session field, such as a unique identifier composed of assistant identification, demander identification, and inquiry time information. This query aims to determine whether there is historical related information in the current session and whether the debouncing window corresponding to the session field is active. The debouncing window ensures that the system can wait for and collect subsequent messages under the same session within a specific time period, thus effectively handling scenarios where users split inquiries in multiple rounds. During the query process, if there is inquiry information in the short-term memory cache associated with the current session field and the debouncing window is still active, the system will extract this associated inquiry information. This associated inquiry information is not stored arbitrarily but is preprocessed: when the system previously received this associated inquiry information, it performs local intent recognition and intent reasoning to obtain associated intent information. Only when this associated intent information indicates that the associated inquiry information does not meet the conditions for directly constructing structured information (e.g., it only contains specification information without a clear purchase intent) will the system store it in the short-term memory cache according to its corresponding session field. This conditional storage mechanism avoids redundantly storing irrelevant information or already processed complete information in the cache, effectively preventing the cache from being diluted by irrelevant information such as advertisements or sales pitches. Once valid and relevant inquiry information is obtained, the system concatenates it with the current original inquiry information. This concatenation operation integrates fragmented historical messages with the current message into a coherent whole, forming the target inquiry information. In this way, even if the user sends the specifications and intent of the inquiry multiple times, the system can recombine these scattered pieces of information to form a target inquiry information containing a complete context. This target inquiry information is then sent to subsequent parsing, local intent recognition, and intent reasoning processes, providing a more comprehensive and accurate input for subsequent processing and construction of structured inquiry information using large language models. This mechanism, in conjunction with basic information processing methods, enables more accurate capture of the user's true intent when dealing with complex and ever-changing instant messaging inquiries, and improves the accuracy and completeness of structured information construction.

[0048] For example, suppose in a B2B steel transaction scenario, a customer sends an inquiry to a smart assistant via an instant messaging platform. The customer first sends a message "16×2200" at 10:28 AM, which only contains product specifications and no clear intention to inquire about pricing. Then, at 10:35 AM, the customer sends another message, "Found it? Do you have it in stock?", which contains a clear inquiry. These two messages are 7 minutes apart, constituting a typical multi-round, segmented inquiry scenario. When the first message "16×2200" is received by the smart assistant, the system processes it based on its session fields (e.g., a combination of the smart assistant identifier, customer identifier, and message timestamp). Since there is no associated information about this session in the short-term memory cache at this time, the system performs local intent recognition and intent inference. Through preset local rule checks (e.g., pure numeric rules, keyword rules) and remote text intent inference (e.g., AC automaton matching, FastText model classification), the message is determined to not contain a clear purchasing intent, for example, the intent is "other," and the confidence level is below a preset threshold. Given that the message contains a specification pattern, the system marks it as "pending_spec" and stores it in the short-term memory cache based on its session field. During storage, a debouncing window (e.g., 30 seconds) and a time-to-live (TTL) (e.g., 45 minutes) are set for this cache entry, and its state is set to "SPEC_WATCHING". At this point, the system does not immediately push it as a business opportunity, nor does it invoke the large language model and rule engine. Seven minutes later, when the second message "Found it? Is it in stock?" is received by the smart assistant, the system again queries the short-term memory cache based on its session field. This time, the system finds that the session is in the "SPEC_WATCHING" state, and the debouncing window is not yet closed (because the second message arrived within the TTL of the debouncing window after the first message was stored). The system retrieves the first message "16×2200" from the short-term memory cache and concatenates it with the current second message "Found it? Is it in stock?" in chronological order, for example, concatenating it as "16×2200\nFound it? Is it in stock?". Before concatenation, the system filters the second message to ensure it contains no irrelevant features such as advertisements, sales pitches, or casual conversation. The concatenated text, as the target inquiry information, is then fed into the subsequent text intent inference module. At this point, the AC automaton might detect purchase requests such as "Found it?" or "Is it in stock?", while the FastText model might output the intent "buy" with high confidence. Based on this, the system upgrades the intent of the conversation to "confirmed".Subsequently, this target inquiry information containing complete context ("16×2200\nDid you find it? Is it in stock?") will be used for subsequent large language model extraction and field rule engine standardization processing, so as to accurately identify product specifications and purchasing intentions, and perform operations such as product name completion, and finally construct complete target inquiry structured information.

[0049] In summary, the above processing effectively solves the problems of missing contextual information and interference from irrelevant information caused by multi-round sentence-splitting inquiries in instant messaging scenarios. By introducing short-term memory caching and a debouncing window mechanism, the system can intelligently identify and temporarily store fragmented inquiry information that does not meet the independent structuring conditions, such as messages containing only specifications. When subsequent related messages arrive within the debouncing window, the system can accurately piece together these scattered pieces of information to form a target inquiry information containing complete context, thereby avoiding inquiry loss due to insufficient intent judgment of a single message. At the same time, by performing local intent recognition and intent reasoning on related inquiry information, and only storing it in the cache when the structuring conditions are not met, the system effectively filters out the dilution and interference of irrelevant information such as advertisements and sales pitches, ensuring the quality and accuracy of the pieced information. This significantly improves the accuracy and completeness of converting the original inquiry information into target inquiry structured information, providing high-quality input for subsequent large language model processing and structured data construction, thereby improving the efficiency and robustness of the entire information processing flow.

[0050] Existing information processing methods, when parsing target inquiry information and performing intent identification and inference, often struggle to balance accuracy and real-time processing when relying solely on a single intent identification method. Furthermore, they lack refined, layered processing of the intent identification process, resulting in insufficient robustness in intent judgment when faced with complex and ever-changing inquiry texts. This makes it difficult to effectively distinguish intent information from different sources, thus affecting the quality of subsequent structured processing. Therefore, in this embodiment, the process of parsing the target inquiry information to determine initial structured inquiry information, and performing local intent identification and inference on the initial structured inquiry information to obtain target intent information, includes: The target inquiry information is parsed to obtain the inquiry text. The inquiry text is normalized to obtain initial inquiry structured information. The initial inquiry structured information is subjected to local intent recognition according to preset local heuristic rules to obtain preliminary intent information. The preliminary intent information is inferred by calling the non-local intent reasoning interface to obtain inferred intent information. The target intent information is determined based on the preliminary intent information and the inferred intent information.

[0051] Specifically, parsing the target inquiry information aims to extract the plain text of the inquiry from the raw information, which may contain various formats and content. This can be achieved through Natural Language Processing (NLP) techniques, such as using text extraction tools to identify and extract text content from the message body, or using regular expression matching to remove non-text elements such as HTML tags and special symbols. Normalizing the inquiry text aims to eliminate noise and inconsistencies, converting it into a unified and standardized format for subsequent intent recognition and reasoning. For example, this can involve converting full-width characters to half-width characters, traditional Chinese characters to simplified Chinese characters, synonym replacement (e.g., unifying "threaded steel" and "thread"), and unit standardization (e.g., unifying "ton" and "t"). Furthermore, stop words and punctuation can be removed, or lemma can be restored. The initial structured inquiry information is a structured representation of the normalized inquiry text, transforming unstructured text into a data format that is easily understood and processed by machines. For example, normalized text can be encapsulated as a JSON object containing metadata such as text content, source, and timestamp, or stored as key-value pairs, such as {"content": "normalized text", "source": "IM"}.

[0052] Correspondingly, local heuristic rules are a set of predefined judgment rules based on domain knowledge and experience, used to quickly and efficiently identify common inquiry intents. These rules can include keyword matching rules (e.g., detecting words like "wanting to buy," "do you have stock," "what price," etc.), regular expression rules (e.g., identifying non-wanting information such as phone numbers and license plate numbers), or rules based on text length or specific sentence structures. Local intent recognition refers to using these heuristic rules to quickly match and judge the initial structured information of the inquiry in the local system environment to preliminarily determine its intent. For example, a rule engine can traverse the rule set, and once a match is successful, the intent judgment result is immediately given. Preliminary intent information is the output of local intent recognition, which contains the preliminarily determined intent type and its confidence level. For example, if the keyword "wanting to buy" is matched, the preliminary intent information might be {"intent": "wanting to buy", "confidence": 1.0, "source": "local_rule"}.

[0053] Correspondingly, a non-local intent reasoning interface refers to an external intent recognition service interface, typically based on more powerful computing resources and complex models. This interface can be a remote API service, such as a natural language understanding service based on a cloud platform, or a microservice deployed on a dedicated server. Intent reasoning refers to using this non-local interface to perform deep semantic analysis and intent judgment on the initial structured information of the inquiry, leveraging advanced machine learning or deep learning models (such as large language models based on the Transformer architecture, recurrent neural networks, etc.). This reasoning can capture complex semantic patterns and contextual information that are difficult to cover by local heuristic rules. The inferred intent information is the output of non-local intent reasoning, providing the intent type, confidence level, and possible entity information determined by the complex model. For example, the inferred intent information might be {"intent": "inquiry", "confidence": 0.95, "entities": {"product":"rebar"}}.

[0054] Accordingly, determining the target intent information refers to comprehensively considering the prior intent information obtained from local intent recognition and the inferred intent information obtained from non-local intent reasoning, and ultimately determining the most accurate and reliable inquiry intent through a certain fusion strategy. Fusion strategies may include, but are not limited to: priority arbitration (e.g., prioritizing the adoption of local rules when the confidence level is 1.0), weighted voting (weighted averaging based on the confidence levels of different sources), or decision models based on meta-learning. For example, if both local rules and non-local models determine it as "wanting to buy," the confidence level of that intent can be strengthened; if the two conflict, a selection is made according to a preset arbitration logic. The target intent information is the finally determined inquiry intent used to guide subsequent business processes; it represents the system's final judgment on the user's true intent.

[0055] Based on this, after receiving the original inquiry information and identifying the target inquiry information, the system first parses the target inquiry information to obtain the inquiry text, and then normalizes this text to obtain the initial structured inquiry information. This preprocessing step aims to eliminate noise and format inconsistencies in the original text, providing a standardized and clean input foundation for subsequent intent recognition. On this basis, the system performs two layers of intent recognition in parallel or sequentially. The first layer is local intent recognition, which quickly judges the initial structured inquiry information according to preset local heuristic rules to obtain preliminary intent information. This approach utilizes domain knowledge and empirical rules, enabling rapid identification of most common and explicit inquiry intents with extremely low computational cost, thus effectively filtering invalid information or quickly confirming high-confidence intents while ensuring real-time performance. The second layer is intent reasoning, which uses a non-local intent reasoning interface to perform deep semantic analysis of the initial structured inquiry information using a more powerful model and computing power to obtain reasoned intent information. This non-local reasoning can handle complex semantics and deep intents that are difficult for local rules to cover, compensating for the shortcomings of heuristic rules in terms of flexibility and generalization. Ultimately, the system makes a comprehensive decision based on both prior intent information and inferred intent information to determine the final target intent information. This fusion mechanism fully leverages the efficiency of local rules and the intelligence of non-local models, significantly enhancing the robustness of intent recognition through the complementarity and verification of multi-source information. Through this hierarchical and fusion-based intent recognition mechanism, this application can more accurately and efficiently understand user intent, providing high-quality input for subsequent processing of the initial query structured information using a large language model associated with the target intent information, thereby obtaining the initial query item sequence. This improves the efficiency and accuracy of the entire information processing method and reduces unnecessary computational consumption and erroneous extraction.

[0056] For example, suppose a customer submits a raw inquiry to the intelligent assistant via an instant messaging platform: "Do you have rebar? What's the price? Urgent!". The system first determines the conversation field corresponding to the raw inquiry and identifies the target inquiry based on the query results and the raw inquiry. Next, the system parses the target inquiry, extracting the inquiry text "Do you have rebar? What's the price? Urgent!" and normalizing it. For example, punctuation and modifiers are removed, converting the text to "Do you have rebar? What's the price? Urgent!", thus obtaining the initial structured inquiry information, represented as {"text": "Do you have rebar? What's the price? Urgent!"}. Subsequently, the system performs local intent recognition on this initial structured inquiry information according to preset local heuristic rules. For example, the local rule base contains purchasing keywords such as "what price" and "is it in stock?". In this example, because the text contains the keyword "what price", the local intent recognizer quickly determines the intent to be "want to buy" and generates preliminary intent information, such as {"intent": "want to buy", "confidence": 1.0, "source":"local_rule"}. Simultaneously, or if the local intent recognition fails to provide a high-confidence result, the system calls the non-local intent inference interface, sending the initial price inquiry structured information to this interface for intent inference. This non-local interface may utilize a pre-trained deep learning model (such as a Transformer-based text classification model) to perform semantic analysis on "Do you have rebar? What price? Urgently needed". After model inference, it may output inferred intent information, such as {"intent": "price inquiry", "confidence": 0.97, "source": "remote_model"}. Finally, the system determines the final target intent information based on the preliminary intent information and the inferred intent information. In this example, both the local rule and the non-local model point to either "wanting to buy" or "inquiring about prices" (the two are semantically similar). The system can employ an arbitration strategy; for example, if the confidence level of the local rule is 1.0, the result of the local rule is adopted first, or the result with the higher confidence level is taken. Ultimately, the target intent information is determined to be "wanting to buy," with a confidence level of 0.97 or 1.0. This explicit target intent information will guide the subsequent structured extraction by the large language model, ensuring the accuracy and efficiency of subsequent processing.

[0057] In summary, by integrating local and non-local intent information, this application can more accurately determine the user's true intent, significantly enhance the robustness of intent judgment, effectively distinguish intent information from different sources, provide a reliable decision-making basis for subsequent structured information construction, and thus improve the efficiency and accuracy of the entire information processing process.

[0058] Single reasoning methods often struggle to balance accuracy and robustness in intent recognition, easily leading to intent judgment bias and consequently affecting the accuracy of subsequent price inquiry processing. Therefore, in this embodiment, the invocation of a non-local intent reasoning interface to perform intent reasoning on the initial structured price inquiry information to obtain reasoned intent information includes: The non-local intent reasoning interface is invoked to perform intent reasoning on the initial query structured information according to a preset matching algorithm to obtain first inferred intent information; the non-local intent reasoning interface is invoked to input the initial query structured information into the intent classification model for intent reasoning to obtain second inferred intent information; and the inferred intent information is determined based on the first inferred intent information and the second inferred intent information.

[0059] Specifically, the non-local intent reasoning interface is used to communicate with external intent reasoning services or modules to obtain intent judgments from the initial query structured information. Its implementation methods may include, but are not limited to: interacting with independent intent reasoning microservices via Remote Procedure Call (RPC); calling cloud service-based API interfaces via HTTP / HTTPS; or exchanging data with asynchronously processed intent reasoning modules via message queues. The pre-defined matching algorithm aims to quickly and deterministically identify the intent of the initial query structured information using predefined rules, patterns, or knowledge bases. Its implementation methods may include, but are not limited to: precise matching algorithms based on keywords or phrases, such as AC automata or BM algorithms; pattern matching based on regular expressions; or rule engines built based on domain expert experience. The intent classification model is a classifier built using machine learning or deep learning techniques, capable of learning and identifying the potential intent categories from the initial query structured information. The implementation methods can include, but are not limited to: models based on text embeddings (such as Word2Vec, FastText) and traditional machine learning classifiers (such as Support Vector Machines (SVM), Logistic Regression); models based on deep neural networks (such as Recurrent Neural Networks (RNN), Long Short-Term Memory Networks (LSTM), Transformers); or classifiers fine-tuned from pre-trained large language models (such as BERT, GPT series). The first and second inference intent information represent the intent judgment results obtained through a preset matching algorithm and intent classification model, respectively. They typically include the intent category (e.g., "seeking purchase," "consultation," "advertisement," etc.) and possible confidence or matching degree. Determining the inference intent information aims to comprehensively utilize the first and second inference intent information to derive the final, more reliable inference intent information. The implementation methods can include, but are not limited to: when the two intent information are consistent, directly adopting the intent; when there is a conflict, making a decision based on preset arbitration rules (e.g., prioritizing rule matching results, prioritizing the model with higher confidence, or weighted fusion combining manually labeled data); or making a comprehensive judgment through voting mechanisms, expert systems, etc.

[0060] Based on this, when intent reasoning is required on the initial quoting structured information, the system executes two different reasoning strategies in parallel or sequentially through a non-local intent reasoning interface. First, the non-local intent reasoning interface is invoked, using a pre-defined matching algorithm to perform a fast and deterministic analysis of the initial quoting structured information, thereby obtaining the first inferred intent information. This matching algorithm is typically based on predefined rules or patterns and can efficiently identify explicit intent expressions. Simultaneously, the system invokes the non-local intent reasoning interface again, inputting the same initial quoting structured information into a trained intent classification model. This intent classification model utilizes its generalization ability to learn complex features from the data to determine the intent, thereby obtaining the second inferred intent information. In this way, the rule-based matching algorithm provides high-precision and interpretable intent judgment, while the model-based classifier provides robust recognition capabilities for complex, ambiguous, or novel expressions. Finally, the system makes a comprehensive judgment based on the inferred intent information from these two different sources, for example, through comparison, arbitration, or fusion strategies, to determine the final inferred intent information. This dual-verification mechanism effectively compensates for the limitations of single reasoning methods, significantly reduces the bias in intent judgment, and provides more accurate and reliable input for subsequent determination of target intent information and processing of large language models.

[0061] For example, when receiving normalized initial structured inquiry information (e.g., an inquiry text "What price is available for rebar?"), it can be simultaneously input into two inference modules called by the non-local intent inference interface. One module can use the AC automaton as a preset matching algorithm to perform longest phrase matching on the inquiry text. Assuming the AC automaton hits the two preset purchase phrases "what price" and "available", it outputs the first inference intent information as "purchase", along with a list of matched phrases. The other module can input the inquiry text into an intent classification model trained on FastText. After inferring the text, the model outputs the second inference intent information as "purchase", with a confidence score of 0.97. Since both the first and second inference intent information indicate "purchase", and there is no conflict between them, the system can arbitrate and output the final inference intent information as "purchase", adopting the confidence score of 0.97 given by the intent classification model, and marking its source as "model + rule". This combination ensures the accuracy and reliability of intent judgment.

[0062] In summary, by combining pre-defined matching algorithms and intent classification models in parallel or sequentially for dual reasoning, we can fully leverage the determinism and efficiency of matching algorithms in identifying explicit intents, as well as the generalization ability and robustness of intent classification models in handling ambiguous, colloquial, or non-standard expressions. This multi-dimensional cross-validation mechanism significantly improves the accuracy and reliability of intent identification from the initial structured query information, effectively avoiding misjudgments. This provides high-quality, high-confidence intent input for subsequent target intent information determination, large language model processing, and final query item standardization, thereby ensuring the accuracy of the entire information processing flow.

[0063] The raw fields directly output by large language models often lack industry standardization, containing colloquial expressions, abbreviations, aliases, and non-standard units, leading to low product matching hit rates and redundancy / inconsistency between raw and standard fields. To address this, in this embodiment, the standardization process performed by the field rule engine on the initial query item sequence to obtain a standard query item sequence includes: According to the field transformation rules in the field rule engine, the initial inquiry entries contained in the initial inquiry entry sequence are transformed to obtain the intermediate inquiry entry sequence; according to the field mapping rules in the field rule engine, the intermediate inquiry entries contained in the intermediate inquiry entry sequence are standard mapped to obtain the standard inquiry entry sequence.

[0064] A field rule engine is a configurable system or module used to define, store, and execute a set of rules for data fields. These rules can include operations such as data format validation, data transformation, data mapping, and data cleaning. The field rule engine can be a rule parser based on XML or JSON configuration files, executing data processing logic by reading predefined rule sets; alternatively, it can be a rule engine framework implemented in a programming language, allowing developers to define complex business rules by writing code or configuring a domain-specific language (DSL).

[0065] Said standardization processing includes converting initial inquiry entries included in an initial inquiry entry sequence according to field conversion rules in a field rule engine, to obtain an intermediate inquiry entry sequence. Said field conversion rules are a specific rule type in the field rule engine, which are used for converting data fields in non-standard formats or expressions into unified, standardized formats. For example, said field conversion rules may be defined as regular expression matching and replacement rules, which are used for converting "16mm" into "16", or unifying multiple abbreviations of "Q235B" into a standard name; alternatively, said field conversion rules may be defined as a lookup table or dictionary mapping, which convert colloquial expressions (such as "what price") into standard keywords (such as "price"), or convert non-standard units (such as "ton") into standard units (such as "kilogram").

[0066] On this basis, said standardization processing further includes performing standard mapping on intermediate inquiry entries included in an intermediate inquiry entry sequence according to field mapping rules in said field rule engine, to obtain a standard inquiry entry sequence. Said field mapping rules are another specific rule type in the field rule engine, which are used for associating or aligning converted data fields with predefined standard data items. For example, said field mapping rules may be defined as matching rules based on a knowledge base or standard database, which map the converted product name "medium plate" to a standard product name ID inside the system; alternatively, said field mapping rules may be defined as fuzzy matching or semantic matching rules, which map attribute values in intermediate inquiry entries to the closest standard attribute values in combination with industry dictionaries and ontology.

[0067] Based on this, by introducing a field rule engine, after processing the initial inquiry structured information using the large language model associated with the target intention information and obtaining the initial inquiry entry sequence, further standardization processing is performed on the initial inquiry entry sequence. The field rule engine first converts each initial inquiry entry in the initial inquiry entry sequence according to preset field conversion rules. This conversion aims to eliminate noise such as colloquial expressions, abbreviated aliases, and non-standard units that may exist in the direct output of the large language model, and convert non-standard original text into an intermediate inquiry entry sequence with uniform format and consistent semantics. This stage of processing ensures that the data basis for subsequent operations is clean and standardized. Subsequently, the field rule engine aligns the intermediate inquiry entries in the intermediate inquiry entry sequence with the industry standard database according to preset field mapping rules, thereby mapping the converted data into a standard inquiry entry sequence conforming to industry specifications. This staged conversion and mapping processing not only effectively solves the problem of noise interference of large model-generated content in specific industry applications, but also eliminates redundancy and inconsistency by accurately aligning original fields with standard fields. In this way, the solution of the present application ensures the accuracy, consistency and matchability of key commodity entry data in the entire processing flow from original inquiry information to target inquiry structured information, thereby significantly improving the accuracy of subsequent commodity matching and the robustness of system processing.

[0068] For example, when the system processes the initial inquiry structured information using the large language model associated with the target intention information and obtains the initial inquiry entry sequence, for example: [{product_name: "medium plate", material: "Q235B", thickness: "16", width: "2200", inquiry_keywords: "what is the price, do you have stock"}]. At this time, this entry only contains the fields originally extracted by the large model, and needs to enter the field rule engine for standardization processing. The field rule engine can be specifically implemented as an Industry SKU Field Rule Engine (IndustrySkuFieldRuleEngine).

[0069] First, in the field conversion stage, the field rule engine converts the initial inquiry entries in the initial inquiry entry sequence according to the field conversion rules. For example, for the specification fields thickness="16" and width="2200", the field conversion rule of "trim decimal trailing zeros" (trimDecimalZero) can be executed. If the original value is "16.00", it will be converted to "16"; in this example, "16" and "2200" themselves do not contain decimal trailing zeros, so they remain unchanged. After processing at this stage, the intermediate inquiry entry sequence is obtained.

[0070] Subsequently, in the field mapping stage, the field rule engine performs standard mapping on intermediate inquiry entries in the intermediate inquiry entry sequence according to field mapping rules. For example, for the product name productName="medium plate", it can be sent to the batch resolution interface (resolveBatch) of the knowledge base and mapped to the standard name canonicalName="medium plate", which may be achieved through exact matching (LINKED). Similarly, for the material material="Q235B", it can also be sent to said batch resolution interface and mapped to the standard name canonicalName="Q235B". In this process, product name completion model inference can also be performed. For example, the specification text "16*2200", the original product name "medium plate", the original material "Q235B" and a candidate product name list are sent to a large language model for inference, and the model outputs the selected product name "medium plate" and the corrected specification "16*2200". Finally, after the output cleaning stage, only specifications, product line codes and standard fields are retained, and internal auxiliary fields are removed, so as to obtain a standard inquiry entry sequence, for example: {product_name: "medium plate", material: "Q235B", thickness: "16", width: "2200", specification: "16*2200", inquiry_keywords: "what is the price, do you have stock"}.

[0071] In summary, the above-mentioned processing effectively solves the noise problems such as colloquial expressions, abbreviated aliases, and non-standard units existing in the initial inquiry entry sequence directly output by the large language model, as well as the resulting problems of low subsequent product matching hit rate and redundant inconsistency between original fields and standard fields. Specifically, through the field conversion rules in the field rule engine, the initial inquiry entries can be effectively cleaned and normalized, ensuring data consistency; then through the field mapping rules, the converted data is accurately aligned with industry standards, thereby obtaining a high-quality standard inquiry entry sequence. This enables significant improvement in the accuracy and availability of structured information when subsequently constructing the target inquiry structured information corresponding to the original inquiry information based on the standard inquiry entry sequence, greatly enhances the robustness and efficiency of the entire information processing method in specific industry scenarios, and provides a reliable data foundation for downstream product matching and services.

[0072] Due to the lack of a mechanism for quality verification and link tracing of standardized processed entries, the constructed structured information may have logical errors or data inconsistency, and it is difficult to trace the information source, which cannot meet the high requirements for data accuracy and traceability in steel trade transactions. To address this problem, in this embodiment, said constructing the target inquiry structured information corresponding to the original inquiry information based on said standard inquiry entry sequence comprises: The standard inquiry entries contained in the standard inquiry entry sequence are detected according to the preset post-check rules; the target inquiry intent information, structured inquiry entries, and link tracing identifier are determined based on the detection results; and the target inquiry structured information corresponding to the original inquiry information is constructed based on the target inquiry intent information, the structured inquiry entries, and the link tracing identifier.

[0073] Specifically, pre-defined post-check rules refer to a series of predefined conditions or algorithms used to further verify the data quality, logical consistency, and business compliance of inquiry information after initial standardization. These rules can be configured based on domain expert experience. For example, for steel products, reasonable ranges for specifications such as thickness, width, and length can be set, or logical conflicts between product name and material can be checked. Furthermore, machine learning models can be trained to learn patterns in historical valid data, thereby identifying abnormal or illogical data combinations. A standard inquiry item sequence refers to a set of inquiry information that conforms to predefined formats and specifications after standardization by the field rule engine. Each item in this sequence has had colloquial expressions, abbreviations, aliases, non-standard units, and other noise from the original inquiry information normalized. For example, it can be a list of key-value pairs containing fields such as product name, material, specifications (e.g., thickness, width, length), quantity, and unit, or a structured data fragment conforming to specific industry data exchange standards, facilitating subsequent structured information construction and system processing. Detection refers to the process of systematically verifying a sequence of standard inquiry entries by applying pre-defined post-checking rules. This process can examine each standard inquiry entry in the sequence one by one, for example, by traversing each entry and applying a series of pre-defined rules, including but not limited to data type checks, numerical range checks, field integrity checks, and logical checks on the relationships between different fields. Detection can also be done in batches, inputting the entire sequence into a rule engine or data quality verification module, outputting a detection report all at once, or directly marking abnormal entries that do not conform to the rules. The detection result is feedback information generated after performing the detection operation, indicating whether the sequence of standard inquiry entries has passed all pre-defined post-checking rules and detailing any potential problems. The detection result can be a simple Boolean value (e.g., pass or fail) accompanied by a detailed error report listing the specific entries that do not conform to the rules and their reasons. Alternatively, the detection result can be a corrected or filtered sequence of standard inquiry entries, where entries that do not conform to the rules have been removed, corrected, or explicitly marked. Target inquiry intent information refers to the final confirmed and verified purpose or type of inquiry; it is a precise expression of the core intent of the original inquiry information. For example, it can be a label within a predefined intent category, such as "Seeking Purchase," "Inquiry," "Consultation," or "Not Seeking Purchase," reflecting the true business intent of the demander. Target inquiry intent information can also be a structured object containing intent type, confidence score, and related explanations, providing clear guidance for subsequent business processes. A structured inquiry item refers to an inquiry information unit that, after testing and confirmation, conforms to business logic and data specifications.These entries are the core components of the final structured information for the target inquiry. They have been cleaned, standardized, and verified to be error-free, and can be directly used in downstream business processes such as product matching and order processing. Structured inquiry entries can be valid entries selected from a standard inquiry entry sequence, or they can be formed by correcting the original entries based on the detection results. The link tracing identifier is a unique credential used to uniquely identify and trace specific inquiry information throughout its entire process, from reception, parsing, intent identification, standardization processing to final structured construction. This identifier can be a globally unique string, such as a UUID (Universally Unique Identifier), or a composite code combining a timestamp, session ID, and random number. The link tracing identifier enables end-to-end traceability of the inquiry information processing flow, facilitating problem investigation, data auditing, and process optimization. Constructing the target inquiry structured information corresponding to the original inquiry information means integrating the verified target inquiry intent information, structured inquiry entries, and link tracing identifier to form a complete, standardized inquiry data object that can be directly used by downstream systems. This process can encapsulate the above information into a JSON object or XML document, containing explicit intent fields, a structured list of entries, and a unique tracking ID. Alternatively, this data can be written to a specific table in a database, with each field corresponding to a data item, thus completing the transformation from unstructured raw inquiry information to structured business opportunity data.

[0074] Based on this, by introducing post-checking and link tracing mechanisms, the system performs quality control and traceability management on the standard inquiry item sequence obtained by the above method. Upon receiving the standardized inquiry item sequence, the system first checks each standard inquiry item in the sequence according to preset post-checking rules. These rules aim to identify and filter out logical conflicts, data anomalies, or items that do not conform to business specifications that may have been left over from the initial standardization process, thereby ensuring the accuracy and reliability of the subsequently constructed structured information. For example, the rules can check the rationality of steel specifications, the matching of product name and material, and the correctness of quantity units. Based on the detection results, the system further determines the final target inquiry intent information, the verified structured inquiry items, and a unique link tracing identifier. The target inquiry intent information is the final confirmation of the core intent of the original inquiry information, the structured inquiry items are the valid product details that have undergone rigorous screening and verification, and the link tracing identifier provides end-to-end traceability for the entire processing flow. In this way, this application transforms unstructured original inquiry information into business opportunity data with a clear intent, a well-defined structure, and traceable origin. Ultimately, based on the determined target inquiry intent information, structured inquiry entries, and link tracing identifiers, the system constructs the target inquiry structured information corresponding to the original inquiry information. This construction process not only ensures the integrity and standardization of the output data but also provides crucial evidence for subsequent business auditing, data correction, model iteration, and seamless integration with downstream systems through link tracing identifiers. This solution, combined with the standard inquiry entry sequence obtained through the large language model and field rule engine in the aforementioned methods, forms a closed loop from the initial inquiry information reception to the final structured output, effectively improving the robustness and data quality of the entire information processing flow.

[0075] For example, suppose the system receives a raw inquiry message. After initial parsing, intent recognition, large language model processing, and field rule engine standardization, a standard inquiry item sequence is generated. This sequence may contain multiple standard inquiry items, such as: Item A: {Product Name: "Rebar", Material: "HRB400E", Specification: "Φ16*12000", Quantity: "10 tons"}; Item B: {Product Name: "I-beam", Material: "Q235B", Specification: "200*100*7*10", Quantity: "5000 kg"}; Item C: {Product Name: "Channel Steel", Material: "Q235B", Specification: "16#", Quantity: "2000 meters"}; Item D: {Product Name: "Scrap Steel", Material: "Unlimited", Specification: "Unlimited", Quantity: "5 tons"}. At this point, the system will inspect this standard inquiry item sequence according to preset post-checking rules. For example, preset rules may include: 1. Specification completeness check: For conventional steel products such as rebar and I-beams, the specification field must include all key dimensions (such as diameter, length, cross-sectional dimensions, etc.). 2. Name and material matching check: Check whether the name and material conform to industry standards; for example, "rebar" usually corresponds to grades such as "HRB400E". 3. Quantity unit reasonableness check: Check whether the quantity unit matches the name; for example, "rebar" is usually measured in "tons," while "meters" may be unreasonable. 4. Intention Figure 1 Consistency Check: If the sequence contains non-core business categories such as "scrap steel," it may trigger intent adjustment or marking. During the detection process, the system found that: Items A and B passed all checks and were determined to be valid structured inquiry items. Item C's "Quantity: 2000 meters" may not conform to the conventional trading unit for channel steel (usually tons), or its specification "16#" may need further refinement. According to the rules, this item may be marked for manual review or directly filtered. Item D, "scrap steel," may trigger non-purchasing intent or special processing procedures because it does not belong to conventional steel purchasing. Based on the detection results, the system can determine: Target inquiry intent information: for example, determined as "purchasing (partially valid)" or "purchasing (conventional steel) + non-purchasing (scrap steel)." Structured Inquiry Items: Only items A and B are retained as the final structured inquiry items. Link Trace Identifier: A unique identifier, such as "trace_20240315_001," is generated to trace the entire processing process of this inquiry information. Finally, based on the determined target query intent information, structured query entries (entry A and entry B), and tracing identifiers, the system constructs the final target query structured information. This information can be a JSON object containing the intent, a list of valid entries, and a tracing ID, for example: json { "intent": "Request to purchase", "items": [ {"product_name": "Rebar", "material": "HRB400E", "spec": "Φ16*12000", "quantity": "10", "unit": "tons"}, {"product_name": "I-beam", "material": "Q235B", "spec": "200*100*7*10", "quantity": "5000", "unit": "kilograms"} ], "trace_id": "trace_20240315_001" } In this way, even if the original inquiry information is complex and varied, it can still output high-quality, traceable structured data after being processed by this solution.

[0076] In summary, the combination of the standard query item sequence obtained through the large language model and field rule engine forms a more complete and robust information processing closed loop. This ensures that the transformation from unstructured raw query information to high-quality structured business opportunity data is controllable, reliable, and efficient, thus meeting the high requirements for data accuracy, consistency, and traceability in steel trading scenarios and providing a solid data foundation for downstream business operations such as commodity matching and order management.

[0077] Furthermore, when processing mixed price inquiry information containing both image and text, existing technologies often lack effective utilization of the visual features of images and struggle to distinguish between promotional information and genuine purchasing intent within images. This results in limited accuracy in intent recognition under multimodal input, making it impossible to accurately extract valid price inquiries from mixed image and text information. In this embodiment, when the target price inquiry information includes both image and text information, the process of parsing the target price inquiry information to determine initial structured price inquiry information, and performing local intent recognition and intent reasoning on the initial structured price inquiry information to obtain target intent information, includes: The process involves parsing the target inquiry information to determine initial image-text inquiry structured information, which includes initial image inquiry structured information and initial text inquiry structured information. If, according to a preset local heuristic rule, the initial image inquiry structured information is determined to be non-promotional information, an image intent reasoning interface is invoked to perform image intent reasoning on the initial image inquiry structured information to obtain image intent information. Similarly, a non-local intent reasoning interface is invoked to perform intent reasoning on the initial text inquiry structured information to obtain text intent information. Target intent information is then determined based on the image intent information and the text intent information. The step of processing the initial inquiry structured information using a large language model associated with the target intent information to obtain an initial inquiry item sequence includes: identifying the initial image inquiry structured information to obtain identification structured information; and processing the identification structured information and the initial text inquiry structured information using a large language model associated with the target intent information to obtain an initial inquiry item sequence.

[0078] Specifically, "parse the target inquiry information to determine the initial image and text inquiry structured information" refers to performing preliminary structured processing on the received mixed inquiry information containing images and text, decomposing it into a format that can be further analyzed by subsequent modules. This can be done using a multimodal parser, which can process image and text data simultaneously and convert them into a unified intermediate representation; alternatively, it can be done by separately calling the image parsing module and the text parsing module, and then integrating their respective parsing results.

[0079] The "initial image and text inquiry structured information" refers to a structured data set that, after preliminary parsing, represents the original image and text inquiry information. This information may include image metadata, the original text content, and possible relationships between the two. The "initial image inquiry structured information" refers to structured data extracted from image inquiry information, such as the image URL, image feature vector, or text content extracted from the image using Optical Character Recognition (OCR) technology. The "initial text inquiry structured information" refers to structured data extracted from text inquiry information, such as a text sequence preprocessed through word segmentation, part-of-speech tagging, and named entity recognition, or the text embedding vector. The "predefined local heuristic rules" refer to a series of predefined judgment criteria based on experience or domain-specific knowledge, used for quick and lightweight preliminary screening of image content. These rules may include checks on features such as image size, file type, specific watermarks, QR codes, and advertising keywords to identify images that are not indicative of a purchase request. The term "non-promotional information" refers to content that is not advertising, marketing, spam, or other non-purchasing-related information, indicating that the image may contain genuine inquiry intent. The phrase "calling the image intent reasoning interface to perform image intent reasoning on the initial image inquiry structured information to obtain image intent information" refers to using a dedicated image analysis module or service to perform deep analysis on the structured image data to determine the user intent expressed by the image. This interface can classify images based on deep learning models (such as convolutional neural networks, Vision Transformer, etc.) and output whether the image represents a purchase request, inquiry, or other type of intent.

[0080] The "image intent information" refers to the user intent inferred from the image content, such as "purchase intent," "inquiry intent," or "non-purchase intent." "Calling a non-local intent reasoning interface to perform intent reasoning on the initial text inquiry structured information to obtain text intent information" refers to using a more complex intent recognition service, possibly deployed on a remote server, to perform semantic analysis on the text inquiry information to infer the user intent expressed by the text. This interface can be based on a large language model (LLM), natural language processing (NLP) model, or a complex rule engine to perform deep understanding of the text and identify intents such as inquiries, consultations, and complaints. The "text intent information" refers to the user intent inferred from the text content, such as "purchase intent," "inquiry intent," or "non-purchase intent." "Determining the target intent information based on the image intent information and the text intent information" refers to fusing the intent information obtained separately from the image and text to derive the final, comprehensive user intent. This can be achieved through weighted averaging, decision trees, logical judgments (e.g., if either the image or text displays a purchase intention, the final intention is to purchase), or more complex intention fusion models. The phrase "identifying the structured information of the initial image inquiry to obtain the identified structured information" refers to extracting and structuring the specific content contained in the image. This typically involves Optical Character Recognition (OCR) technology, converting text in the image into editable text; it can also include object detection in the image, table recognition, etc., converting visual information into structured data fields. The phrase "identifying structured information" refers to the specific data identified and extracted from the image, such as text recognized by OCR, row and column table data in a table, or product names, specifications, etc., detected in the image. The phrase "processing the identified structured information and the initial text query structured information using a large language model associated with the target intent information to obtain an initial query item sequence" refers to taking the structured data identified from the image and the structured data parsed from the text as input, providing them to a large language model that adjusts or prompts based on the target intent information. This model then performs deep semantic understanding and information extraction to generate a preliminary query item list. The large language model can adjust its extraction strategy based on the target intent information; for example, if the target intent is "to purchase," the model will focus more on extracting key information such as product name, quantity, and specifications. The "initial query item sequence" refers to the preliminary list of unstandardized product or service items extracted by the large language model from the multimodal input. Each item may contain fields such as product name, material, specifications, and quantity.

[0081] Based on this, a multimodal processing mechanism was introduced to achieve deep analysis and intent coordination of mixed text and image inquiry information. First, when the target inquiry information includes both image and text inquiries, the system parses it into initial structured text and image inquiry information. This information is further subdivided into initial structured image inquiry information and initial structured text inquiry information. This classification ensures that data of different modalities can be analyzed in a targeted manner. To improve processing efficiency and reduce resource waste, the system performs preliminary screening of the initial structured image inquiry information according to preset local heuristic rules to determine whether it is non-promotional information, thereby quickly filtering out invalid or non-purchasing image content. After determining that the image is valid information, the system calls a dedicated image intent reasoning interface in parallel or sequentially to perform image intent reasoning on the initial structured image inquiry information to obtain image intent information; simultaneously, it calls a non-local intent reasoning interface to perform intent reasoning on the initial structured text inquiry information to obtain text intent information. Subsequently, the system fuses image intent information and text intent information to determine the final target intent information, thus accurately capturing the user's true purchasing intent. In the subsequent inquiry item sequence generation stage, the system identifies the structured information of the initial image inquiry, for example, by extracting text from the image using OCR technology to obtain the identified structured information. Finally, using a large language model associated with the target intent information, the system processes both the identified structured information (from the image) and the initial text inquiry structured information (from the text) as input to obtain the initial inquiry item sequence. This strategy of deeply integrating visual recognition results with text semantic information can fully utilize the complementary information in multimodal data, enabling the large language model to more comprehensively and accurately understand user needs and generate a more complete and accurate initial inquiry item sequence. This solution optimizes the basic information processing methods for multimodal input scenarios. Through a refined intent recognition and information extraction process, it effectively solves the problem of information loss or misjudgment in complex inquiry scenarios under single-modal processing, significantly improving the accuracy and efficiency of inquiry information structuring.

[0082] For example, suppose a customer sends a screenshot of a chat via an instant messaging platform. The screenshot contains a handwritten specifications table: "Q355B 20×2000 5 sheets, Q235B 12×1800 10 sheets" along with the inquiry text "What is the price for the above specifications, and what is the delivery time?" This message is an image type, containing both image and text inquiry information.

[0083] First, after receiving the target inquiry information, the system will parse it to determine the initial image and text inquiry structured information. This includes extracting the image URL and representing it as the initial image inquiry structured information, while also representing the text content in the image (via OCR or other methods) and the independent text inquiry information "What is the price for the above specifications, and what is the delivery time" as the initial text inquiry structured information.

[0084] Next, the system will judge the initial image query structured information according to preset local heuristic rules. For example, local rules will check whether the image URL is in the blacklist of advertising images, or whether the image contains obvious advertising features. If it is determined to be non-promotional information, the image intent inference interface will be called. This interface may use a pre-trained EfficientNet-B0 model to preprocess the image (e.g., scale it to 224×224 pixels and perform ImageNet normalization) and then perform inference. Assuming the model outputs a confidence score of 0.93 for the "Request to Purchase" category, which far exceeds the preset threshold, the image intent information is determined to be "Request to Purchase". At the same time, the system will call the non-local intent inference interface to perform intent inference on the initial text query structured information "What is the price of the above specifications, and what is the delivery time", to obtain the text intent information, which may also be judged as "Request to Purchase".

[0085] Subsequently, the system determines the final target intent information as "want to buy" based on the image intent information and the text intent information (both of which are "want to buy"). After obtaining the target intent information, the system will recognize the structured information of the initial image inquiry. For example, it will use OCR technology to recognize the handwritten specification table content in the image as text: "Q355B 20×2000 5 sheets, Q235B 12×1800 10 sheets", thus obtaining the recognized structured information.

[0086] Finally, the system utilizes a large language model associated with the "purchase" target intent information, processing both the recognized structured information (OCR text) and the initial text inquiry structured information ("What is the price for the above specifications, and how long is the delivery period") as input. The large language model extracts and generates an initial inquiry item sequence based on this information, for example: [{product_name: "low alloy", material: "Q355B", thickness: "20", width: "2000", quantity:"5 sheets", inquiry_keywords: "What is the price, and how long is the delivery period"}, {product_name: "medium plate", material: "Q235B", thickness: "12", width: "1800", quantity: "10 sheets", inquiry_keywords: "What is the price, and how long is the delivery period"}].

[0087] In summary, the above processing effectively solves the problems of existing technologies lacking effective utilization of image visual features, difficulty in distinguishing promotional information from genuine purchasing intentions, and limited accuracy in intent recognition under multimodal input when processing mixed inquiry information containing both image and text information. This solution, by introducing multimodal parsing, lightweight image intent pre-screening, and a parallel reasoning and fusion mechanism for image and text intents, can more accurately determine the user's true intent, avoiding invalid processing of non-purchasing information and thus saving computational resources. Furthermore, by inputting image recognition results and text information together into a large language model for information extraction, the complementary nature of multimodal data is fully utilized, enabling the system to extract effective inquiry items more comprehensively and accurately from complex mixed image and text information, significantly improving the accuracy and efficiency of structuring inquiry information. In real-world business scenarios, besides the original inquiry information, there are also scenarios involving the modification or editing of existing objects, i.e., object editing requests. Existing standardized processing flows primarily target inquiry entries and lack a mechanism for specifically scanning, mapping, and filtering object description information. This results in the inability to effectively identify and clean noise in object descriptions when processing object editing requests, making it difficult to directly convert the edited object information into standardized business data, thus affecting the accuracy of subsequent object matching and management. Therefore, in this embodiment, the method further includes: The system receives an object editing request, wherein the object editing request carries object description information associated with the target object; scans the object description information according to the field rule engine; maps the object description information according to the scan results to obtain standard object description information; and filters the standard object description information according to the field rule engine to obtain target object description information.

[0088] Specifically, an object editing request refers to an instruction initiated by a user or system to modify or update information related to existing business objects (such as products, materials, customer profiles, etc.). This request is typically received through a user interface (such as a web form or mobile application interface) or an application programming interface (API). The object description information included in the request is a textual representation of the specific attributes or characteristics of the object to be edited, such as the product's name, specifications, material, quantity, etc. This information may exist in the form of unstructured text, semi-structured text, or structured data.

[0089] Correspondingly, the field rule engine is a pre-configured set of rules used to parse, transform, and standardize input data. Scanning object description information involves using the rules in this engine to perform pattern matching, keyword recognition, or syntactic analysis on the original description text to identify the various fields and their values, and to detect potential noise or non-standard expressions. For example, regular expressions can be used to identify specifications or model numbers, or a keyword list can be used to identify product names. Mapping the scan results involves converting the identified raw field values ​​into a unified and standardized format according to preset conversion rules. This can be achieved by consulting a predefined mapping table, executing a specific conversion function, or calling an external knowledge base; for example, mapping "low alloy" to "low alloy plate," and "Q345B" to "Q355B." Through this process, the original object description information, which may contain colloquialisms, abbreviations, or aliases, is transformed into semantically clear and uniformly formatted standard object description information.

[0090] Accordingly, after obtaining the standard object description information, the field rule engine further applies filtering rules to perform quality checks and screening on this standardized information. The purpose of filtering is to eliminate invalid, redundant, or non-compliant data, ensuring that the final output target object description information is of high quality and can be directly used for subsequent business processing. For example, field values ​​that do not meet specific length or format requirements can be filtered out, or unnecessary auxiliary information can be removed based on business logic. Through this filtering step, it is ensured that the final target object description information not only conforms to business standards but also possesses high usability and accuracy, providing a reliable data foundation for subsequent object matching, updating, and management.

[0091] Based on this, by introducing a standardized processing flow for object editing requests, the problem of non-standardized descriptive information in object change scenarios is effectively solved. When the system receives an editing request containing object description information related to the target object, the object description information is first sent to the field rule engine. The field rule engine performs a detailed scan of these raw object descriptions based on its internal preset rule set. The scanning process aims to identify key fields, potential non-standard expressions, and possible noise in the description information. For example, for text descriptions, the engine can identify elements such as product name, material, and specifications. After completing the scan and identifying relevant information, the field rule engine transforms these raw inquiry information according to preset mapping rules. The purpose of this mapping process is to unify various colloquial, abbreviated, or alias expressions into a standardized format, thereby obtaining standard object description information. For example, the user-inputted "low alloy" is mapped to the system's internal standard product name "low alloy plate." Finally, the field rule engine further filters these standard object descriptions. The filtering process aims to remove information that does not conform to business specifications, is invalid, or redundant, ensuring that the final output target object description information is of high quality and can be directly used for subsequent business operations. Through this series of steps, this solution transforms raw, potentially chaotic object description information into a clearly structured, semantically explicit format that conforms to business standards. This provides an accurate and reliable data foundation for subsequent object matching, updating, and management. This process complements the mechanism described above for standardizing initial inquiry entries, together constructing a unified and efficient data standardization system. This ensures that both inquiry information and object editing information receive consistent and high-quality processing, significantly improving data consistency and usability.

[0092] For example, suppose an administrator on a steel trading platform needs to update a product record. The administrator enters or modifies product information in the product management backend, such as entering the product name as "low alloy" (colloquial abbreviation), the material as "Q345B" (old standard name), and the specifications as "20*2000 5 sheets". When the administrator submits this product change operation, the system receives an object edit request, which carries this original object description information. The system sends this original object description information to the field rule engine for processing. The field rule engine first scans the product name field "low alloy", identifies it as product name information, and maps it to the standard product name "low alloy plate" according to preset field conversion rules (such as the "product name abbreviation completion" rule). Next, the specification field "20*2000 5 sheets" is scanned to identify the specification information. The "trimDecimalZero" rule (removing trailing zeros) and the "stripChars" rule (removing non-numeric characters such as "sheets") are applied to map it to the standard specification "20*2000 5". For the material field "Q345B", the field rule engine maps it to the standard material "Q355B" based on the alias-standard name library. After obtaining standard object description information such as "low alloy plate", "Q355B", and "20*2000 5", the field rule engine further filters this information. For example, it checks whether this standard information conforms to preset format requirements or business logic to ensure there is no invalid or redundant data. Finally, the target object description information obtained by the system has been fully standardized and cleaned, for example: Product Name: "Low Alloy Plate", Material: "Q355B", Specification: "20*2000 5". This standardized information can be redundantly stored along with the original fields so that subsequent product matching services can simultaneously support exact matching by standard name and fuzzy matching by original value, thereby effectively solving the matching deviation problem caused by inconsistency between the original fields and the standard fields.

[0093] In summary, the above processing effectively solves the technical problem of lacking a targeted scanning, mapping, and filtering mechanism for object description information in object editing scenarios. By receiving object editing requests and using a field rule engine to scan, map, and filter object description information, the original object description information, which may contain colloquialisms, abbreviations, or aliases, can be transformed into unified, standardized, and high-quality standard object description information. This significantly improves the standardization and data consistency of object information, avoiding the decrease in accuracy of subsequent business processes (such as product matching and inventory management) caused by non-standard information. Furthermore, by reusing the existing field rule engine, this solution not only standardizes object editing information but also unifies it with the standardized process of inquiry information processing, reducing the complexity of system development and maintenance costs, and ensuring the continuity and efficiency of data processing across the entire platform.

[0094] The information processing method described in this embodiment focuses on constructing an intelligent processing system for steel transaction inquiries based on multimodal intent recognition and templated arrangement. This system addresses the technical challenges of existing technologies, such as low efficiency in manually extracting customer needs from unstructured chat logs, significant semantic understanding biases, inconsistencies in multi-source data, and insufficient intent gating leading to ineffective large-scale model calls and matching accuracy. For details, see [link to documentation]. Figure 2 As shown, client instant messaging messages enter through the identification gateway. After synchronous identification, the message first reads and writes the short-term memory buffer through the inquiry context orchestration layer and outputs the L2 business state. Only when confirmed is the message routed to the industry processor according to the major industry. After completing heterogeneous input parsing and pre-intent reasoning, the purchase intent triggers large model extraction, field standardization and post-cleaning, and finally outputs a structured result (intent + product item) to facilitate the completion of corresponding transaction operations within the platform and improve the transaction experience for the demand side.

[0095] The following is in conjunction with the appendix Figure 3 Taking the application of the information processing method provided in this specification in a building materials trading scenario as an example, the information processing method will be further explained. Figure 3 A flowchart illustrating the processing steps of an information processing method provided in one embodiment of this specification is shown, specifically including the following steps.

[0096] Step S302: Receive the original inquiry information submitted by the demander to the intelligent assistant through the instant messaging platform.

[0097] Step S304: Determine the assistant identification information corresponding to the smart assistant, the demander identification information corresponding to the demander, and the inquiry time information corresponding to the original inquiry information.

[0098] Step S306: The assistant identification information, the demander identification information, and the inquiry time information are used as the session fields corresponding to the original inquiry information.

[0099] Step S308: Query the short-term memory cache based on the session field, determine the associated inquiry information based on the query result, wherein the debounce window corresponding to the session field is in running state, and the associated inquiry information and the original inquiry information are concatenated to obtain the target inquiry information.

[0100] Step S310: Parse the target inquiry information to obtain the inquiry text, and obtain the initial inquiry structured information by normalizing the inquiry text.

[0101] Step S312: Perform local intent recognition on the initial price inquiry structured information according to the preset local heuristic rules to obtain the prior intent information.

[0102] Step S314: Call the non-local intent reasoning interface to perform intent reasoning on the initial query structured information according to the preset matching algorithm to obtain the first inferred intent information. Call the non-local intent reasoning interface to input the initial query structured information into the intent classification model to perform intent reasoning to obtain the second inferred intent information.

[0103] Step S316: Determine the reasoning intent information based on the first reasoning intent information and the second reasoning intent information, and determine the target intent information based on the prior intent information and the reasoning intent information.

[0104] Step S318: Process the initial query structured information using the large language model associated with the target intent information to obtain the initial query item sequence. Then, according to the field transformation rules in the field rule engine, transform the initial query items contained in the initial query item sequence to obtain the intermediate query item sequence.

[0105] Step S320: According to the field mapping rules in the field rule engine, perform standard mapping on the intermediate query entries contained in the intermediate query entry sequence to obtain the standard query entry sequence.

[0106] Step S322: Detect the standard inquiry entries contained in the standard inquiry entry sequence according to the preset post-check rules, and determine the target inquiry intent information, structured inquiry entries and link tracing identifier based on the detection results.

[0107] Step S324: Based on the target inquiry intent information, structured inquiry entries, and link tracing identifier, construct the target inquiry structured information corresponding to the original inquiry information.

[0108] In summary, by introducing short-term memory caching to process multi-turn dialogue context to identify effective business opportunities in sentence-segmentation scenarios, and by combining local intent recognition with large language model processing to achieve efficient and accurate structured information construction, this approach has the advantages of effectively identifying effective business opportunities in unstructured inquiry information, reducing omissions, and improving the accuracy and efficiency of product matching.

[0109] Corresponding to the above method embodiments, this specification also provides embodiments of an information processing apparatus. Figure 4 A schematic diagram of the structure of an information processing apparatus according to one embodiment of this specification is shown. Figure 4 As shown, the device includes: The determination module 402 is configured to determine the session field corresponding to the original inquiry information, query the short-term memory cache according to the session field, and determine the target inquiry information according to the query result and the original inquiry information; The parsing module 404 is configured to parse the target inquiry information to determine the initial inquiry structured information, and perform local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information; Processing module 406 is configured to process the initial query structured information using a large language model associated with the target intent information to obtain an initial query item sequence, and to standardize the initial query items contained in the initial query item sequence according to the field rule engine to obtain a standard query item sequence. The construction module 408 is configured to construct the target inquiry structured information corresponding to the original inquiry information based on the standard inquiry item sequence.

[0110] In an optional embodiment, determining the session field corresponding to the original inquiry information includes: Receive original inquiry information submitted by the demander to the intelligent assistant through an instant messaging platform; Determine the assistant identifier information corresponding to the intelligent assistant, the demander identifier information corresponding to the demander, and the inquiry time information corresponding to the original inquiry information; The assistant identification information, the demander identification information, and the inquiry time information are used as the session fields corresponding to the original inquiry information.

[0111] In an optional embodiment, the step of querying the short-term memory cache based on the session field and determining the target inquiry information based on the query result and the original inquiry information includes: The short-term memory cache is queried based on the session field, and the associated price inquiry information is determined based on the query results, wherein the debounce window corresponding to the session field is in running state; The associated inquiry information and the original inquiry information are concatenated to obtain the target inquiry information; The determination of the associated price inquiry information in the short-term memory cache includes: Receive associated price inquiry information, and perform local intent recognition and intent reasoning on the associated price inquiry information to obtain associated intent information; If, based on the associated intent information, it is determined that the associated inquiry information does not meet the conditions for constructing structured information, the associated inquiry information is stored in the short-term memory cache according to the session field corresponding to the associated inquiry information.

[0112] In an optional embodiment, the step of parsing the target inquiry information to determine initial inquiry structured information, and performing local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information, includes: The target inquiry information is parsed to obtain the inquiry text, and the initial inquiry structured information is obtained by normalizing the inquiry text. According to the preset local heuristic rules, the initial query structured information is subjected to local intent recognition to obtain the prior intent information; Call the non-local intent reasoning interface to perform intent reasoning on the initial query structured information to obtain the reasoned intent information; The target intent information is determined based on the prior intent information and the inferred intent information.

[0113] In an optional embodiment, the step of invoking the non-local intent reasoning interface to perform intent reasoning on the initial query structured information to obtain inferred intent information includes: The non-local intent reasoning interface is invoked to perform intent reasoning on the initial query structured information according to a preset matching algorithm, thereby obtaining the first inferred intent information; The initial query structured information is input into the intent classification model by calling the non-local intent reasoning interface to perform intent reasoning and obtain the second inferred intent information; The reasoning intent information is determined based on the first reasoning intent information and the second reasoning intent information.

[0114] In an optional embodiment, the standardization process of the initial query entries contained in the initial query entry sequence according to the field rule engine to obtain a standard query entry sequence includes: According to the field transformation rules in the field rule engine, the initial query entries contained in the initial query entry sequence are transformed to obtain the intermediate query entry sequence; According to the field mapping rules in the field rule engine, the intermediate query entries contained in the intermediate query entry sequence are standardly mapped to obtain the standard query entry sequence.

[0115] In an optional embodiment, constructing the target query structured information corresponding to the original query information based on the standard query item sequence includes: The standard inquiry entries contained in the standard inquiry entry sequence are checked according to the preset post-check rules; Based on the detection results, determine the target inquiry intent information, structured inquiry items, and link tracing identifiers; Based on the target inquiry intent information, the structured inquiry entries, and the link tracing identifier, the target inquiry structured information corresponding to the original inquiry information is constructed.

[0116] In an optional embodiment, when the target inquiry information includes image inquiry information and text inquiry information, the step of parsing the target inquiry information to determine initial inquiry structured information, and performing local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information includes: The target inquiry information is parsed to determine the initial image and text inquiry structured information, wherein the initial image and text inquiry structured information includes initial image inquiry structured information and initial text inquiry structured information; If the initial image pricing structured information is determined to be non-promotional information according to the preset local heuristic rules, the image intent reasoning interface is called to perform image intent reasoning on the initial image pricing structured information to obtain image intent information, and the non-local intent reasoning interface is called to perform intent reasoning on the initial text pricing structured information to obtain text intent information. Determine the target intent information based on the image intent information and the text intent information; The step of processing the initial query structured information using a large language model associated with the target intent information to obtain an initial query item sequence includes: The structured information of the initial image inquiry is identified to obtain the identified structured information. The identified structured information and the initial text inquiry structured information are then processed using a large language model associated with the target intent information to obtain an initial inquiry item sequence.

[0117] In an optional embodiment, the method further includes: Receive an object editing request, wherein the object editing request carries object description information associated with the target object; The object description information is scanned according to the field rule engine, and the object description information is mapped according to the scan results to obtain standard object description information; The standard object description information is filtered according to the field rule engine to obtain the target object description information.

[0118] The above is an illustrative scheme of an information processing device according to this embodiment. It should be noted that the technical solution of this information processing device and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the information processing device, please refer to the description of the technical solution of the information processing method described above.

[0119] Figure 5 A structural block diagram of a computing device 500 according to one embodiment of this specification is shown. The components of the computing device 500 include, but are not limited to, a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and a database 550 is used to store data.

[0120] The computing device 500 also includes an access device 540, which enables the computing device 500 to communicate via one or more networks 560. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 540 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0121] In one embodiment of this specification, the above-described components of the computing device 500 and Figure 5 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 5 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0122] The computing device 500 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 500 can also be a mobile or stationary server.

[0123] The processor 520 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-described information processing method.

[0124] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the information processing method described above.

[0125] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described information processing method.

[0126] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the information processing method described above.

[0127] An embodiment of this specification also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described information processing method.

[0128] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the information processing method described above.

[0129] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0130] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0131] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0132] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0133] The preferred embodiments disclosed above are merely illustrative of this specification. Optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described in this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification.

Claims

1. An information processing method, characterized in that, include: Determine the session field corresponding to the original inquiry information, query the short-term memory cache based on the session field, and determine the target inquiry information based on the query result and the original inquiry information; The target inquiry information is parsed to determine the initial inquiry structured information, and local intent recognition and intent reasoning are performed on the initial inquiry structured information to obtain the target intent information; The initial query structured information is processed using a large language model associated with the target intent information to obtain an initial query item sequence. The initial query items contained in the initial query item sequence are then standardized according to the field rule engine to obtain a standard query item sequence. Based on the standard query item sequence, construct the target query structured information corresponding to the original query information.

2. The information processing method according to claim 1, characterized in that, The step of determining the session fields corresponding to the original inquiry information includes: Receive original inquiry information submitted by the demander to the intelligent assistant through an instant messaging platform; Determine the assistant identifier information corresponding to the intelligent assistant, the demander identifier information corresponding to the demander, and the inquiry time information corresponding to the original inquiry information; The assistant identification information, the demander identification information, and the inquiry time information are used as the session fields corresponding to the original inquiry information.

3. The information processing method according to claim 1, characterized in that, The step of querying the short-term memory cache based on the session field and determining the target inquiry information based on the query result and the original inquiry information includes: The short-term memory cache is queried based on the session field, and the associated price inquiry information is determined based on the query results, wherein the debounce window corresponding to the session field is in running state; The associated inquiry information and the original inquiry information are concatenated to obtain the target inquiry information; The determination of the associated price inquiry information in the short-term memory cache includes: Receive associated price inquiry information, and perform local intent recognition and intent reasoning on the associated price inquiry information to obtain associated intent information; If, based on the associated intent information, it is determined that the associated inquiry information does not meet the conditions for constructing structured information, the associated inquiry information is stored in the short-term memory cache according to the session field corresponding to the associated inquiry information.

4. The information processing method according to claim 1, characterized in that, The process of parsing the target inquiry information to determine initial inquiry structured information, and performing local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information includes: The target inquiry information is parsed to obtain the inquiry text, and the initial inquiry structured information is obtained by normalizing the inquiry text. According to the preset local heuristic rules, the initial query structured information is subjected to local intent recognition to obtain the prior intent information; Call the non-local intent reasoning interface to perform intent reasoning on the initial query structured information to obtain the reasoned intent information; The target intent information is determined based on the prior intent information and the inferred intent information.

5. The information processing method according to claim 4, characterized in that, The step of calling the non-local intent reasoning interface to perform intent reasoning on the initial query structured information to obtain inferred intent information includes: The non-local intent reasoning interface is invoked to perform intent reasoning on the initial query structured information according to a preset matching algorithm, thereby obtaining the first inferred intent information; The initial query structured information is input into the intent classification model by calling the non-local intent reasoning interface to perform intent reasoning and obtain the second inferred intent information; The reasoning intent information is determined based on the first reasoning intent information and the second reasoning intent information.

6. The information processing method according to claim 1, characterized in that, The standardization process, performed by the field rule engine on the initial query entries contained in the initial query entry sequence to obtain a standard query entry sequence, includes: According to the field transformation rules in the field rule engine, the initial query entries contained in the initial query entry sequence are transformed to obtain the intermediate query entry sequence; According to the field mapping rules in the field rule engine, the intermediate query entries contained in the intermediate query entry sequence are standardly mapped to obtain the standard query entry sequence.

7. The information processing method according to claim 1, characterized in that, The construction of the target inquiry structured information corresponding to the original inquiry information based on the standard inquiry item sequence includes: The standard inquiry entries contained in the standard inquiry entry sequence are checked according to the preset post-check rules; Based on the detection results, determine the target inquiry intent information, structured inquiry items, and link tracing identifiers; Based on the target inquiry intent information, the structured inquiry entries, and the link tracing identifier, the target inquiry structured information corresponding to the original inquiry information is constructed.

8. The information processing method according to claim 1, characterized in that, When the target inquiry information includes image inquiry information and text inquiry information, the step of parsing the target inquiry information to determine initial inquiry structured information, and performing local intent recognition and intent reasoning on the initial inquiry structured information to obtain target intent information includes: The target inquiry information is parsed to determine the initial image and text inquiry structured information, wherein the initial image and text inquiry structured information includes initial image inquiry structured information and initial text inquiry structured information; If the initial image pricing structured information is determined to be non-promotional information according to the preset local heuristic rules, the image intent reasoning interface is called to perform image intent reasoning on the initial image pricing structured information to obtain image intent information, and the non-local intent reasoning interface is called to perform intent reasoning on the initial text pricing structured information to obtain text intent information. Determine the target intent information based on the image intent information and the text intent information; The step of processing the initial query structured information using a large language model associated with the target intent information to obtain an initial query item sequence includes: The structured information of the initial image inquiry is identified to obtain the identified structured information. The identified structured information and the initial text inquiry structured information are then processed using a large language model associated with the target intent information to obtain an initial inquiry item sequence.

9. The information processing method according to claim 6, characterized in that, The method further includes: Receive an object editing request, wherein the object editing request carries object description information associated with the target object; The object description information is scanned according to the field rule engine, and the object description information is mapped according to the scan results to obtain standard object description information; The standard object description information is filtered according to the field rule engine to obtain the target object description information.

10. An information processing device, characterized in that, include: The determination module is configured to determine the session field corresponding to the original inquiry information, query the short-term memory cache based on the session field, and determine the target inquiry information based on the query result and the original inquiry information; The parsing module is configured to parse the target inquiry information to determine the initial inquiry structured information, and perform local intent recognition and intent reasoning on the initial inquiry structured information to obtain the target intent information; The processing module is configured to process the initial query structured information using a large language model associated with the target intent information to obtain an initial query item sequence, and to standardize the initial query items contained in the initial query item sequence according to the field rule engine to obtain a standard query item sequence. The construction module is configured to construct the target inquiry structured information corresponding to the original inquiry information based on the standard inquiry item sequence.

11. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 9.

Images