Semantic parsing method and device based on fine-tuning large language model, equipment and medium

Abstract

The application relates to the technical field of text analysis, and discloses a semantic analysis method and device based on fine-tuning of a large language model, equipment and a medium. The method collects bond inquiry and quotation text data, and constructs a sample set through preprocessing, labeling and data enhancement; a general large language model is supervised and fine-tuned in the bond field by using the sample set and an instruction format; a standardized template database is constructed, and the analysis basis is determined by combining the block division of the bond inquiry and quotation text to be analyzed and template matching; a bond field analysis model is used to perform analysis on the bond inquiry and quotation text to be analyzed according to public system prompts, field context and the analysis basis; the confidence evaluation and abnormal completion are combined to optimize the structured output, and then the structured output is mapped to a transaction system. The application solves the problems of low accuracy, weak generalization ability, slow response and low degree of structuring of the traditional manual and rule-driven analysis mode, can be directly connected to the transaction system, and supports intelligentization of the whole link of bond transactions.

Description

Technical Field

[0001] This application relates to the field of text parsing technology, specifically to a semantic parsing method, apparatus, device, and medium based on fine-tuning a large language model. Background Technology

[0002] In interbank bond trading, inquiries and quotes are primarily made through instant messaging tools or brokerage firms, with clients typically expressing their intentions in natural language. This type of natural language text is characterized by inconsistent formats, mixed elements, semantic ambiguity, and numerous abbreviations, encompassing multiple core elements such as bond code, abbreviation, direction, yield, settlement speed, face value, and transaction date.

[0003] Existing parsing methods mainly rely on manual input or regular expression matching rules based on fixed sentence templates, which have many drawbacks:

[0004] First, its accuracy is limited, making it difficult to handle complex semantic scenarios such as word order changes, synonym substitutions, misspellings or colloquialisms, and omitted information.

[0005] Secondly, it has weak generalization ability. As business scenarios expand, the rule base needs to be continuously expanded, which makes migration and reuse difficult.

[0006] Third, the response speed is slow, and manual intervention results in a long response time, which cannot support the needs of high-frequency scenarios.

[0007] Fourth, the degree of structuring is low, and the parsing results cannot be directly integrated into the trading system, requiring additional processing costs.

[0008] With the accelerated digital and intelligent transformation of bond trading, traditional manual or rule-driven parsing methods are no longer sufficient to meet the business requirements of high concurrency and high accuracy. There is an urgent need for an intelligent semantic understanding method for bond trading scenarios to achieve automatic parsing and standardized generation of natural language query and quotation information. Summary of the Invention

[0009] To address the aforementioned issues, embodiments of this application provide a semantic parsing method, apparatus, device, and medium based on fine-tuning a large language model. This enables automatic understanding of query and quotation semantics and outputs structured results conforming to a predefined structure, supporting automatic integration with transaction systems, significantly improving parsing accuracy and real-time performance, and reducing manual intervention and maintenance costs.

[0010] The embodiments of this application adopt the following technical solutions:

[0011] Firstly, this application provides a semantic parsing method based on fine-tuning a large language model, including:

[0012] Collect bond inquiry and quotation text data, and construct a sample set after performing preprocessing, annotation, and data augmentation on the bond inquiry and quotation text data;

[0013] Configure the instruction format, and use the sample set and instruction format to perform bond domain supervised fine-tuning on the general large language model to obtain a bond domain parsing model that generates structured output;

[0014] Extract the query and price items from each sample in the sample set, and construct a standardized template database after performing special tag replacement on the query and price items; where the query and price item is the smallest semantic unit containing independent bond information;

[0015] Receive the bond inquiry and quotation text to be parsed, extract the inquiry and quotation items to be parsed from the bond inquiry and quotation text, and select matching templates from the standardized template database based on the inquiry and quotation items to be parsed as the basis for parsing;

[0016] Configure public system prompts, form a domain context based on the query and quotation items to be parsed, and use the bond domain parsing model to parse the bond query and quotation text to be parsed according to the public system prompts, domain context and parsing criteria, and generate structured output;

[0017] The structured output is optimized based on confidence evaluation and anomaly completion, and then mapped to the trading system.

[0018] Secondly, this application also provides a semantic parsing device based on fine-tuning a large language model, comprising:

[0019] The data construction unit is used to collect bond inquiry and quotation text data, and to construct a sample set after performing preprocessing, annotation and data augmentation on the bond inquiry and quotation text data.

[0020] The model training unit is used to configure the instruction format. After performing bond domain-supervised fine-tuning on the general large language model using the sample set and instruction format, a bond domain analytical model with structured output is obtained.

[0021] The template construction unit is used to extract the query and quotation items of each sample in the sample set, and construct a standardized template database after performing special tag replacement on the query and quotation items; wherein, the query and quotation item is the smallest semantic unit containing independent bond information;

[0022] The template matching unit is used to receive the bond inquiry and quotation text to be parsed, extract the inquiry and quotation items to be parsed from the bond inquiry and quotation text, and select matching templates from the standardized template database based on the inquiry and quotation items to be parsed as the basis for parsing.

[0023] The semantic parsing unit is used to configure public system prompts, form a domain context based on the query and quotation items to be parsed, and use the bond domain parsing model to parse the bond query and quotation text to be parsed according to the public system prompts, domain context and parsing criteria, and generate structured output;

[0024] The optimization docking unit is used to optimize the structured output based on confidence evaluation and anomaly completion, and then map the optimized structured output to the transaction system.

[0025] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the semantic parsing method based on fine-tuning a large language model described above.

[0026] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the semantic parsing method based on fine-tuning a large language model described above.

[0027] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0028] This application collects bond inquiry and quotation text data, constructs a sample set through preprocessing, annotation, and data augmentation; uses the sample set and instruction format to perform bond-domain supervised fine-tuning on a general large language model; constructs a standardized template database, and determines the parsing basis by combining the block segmentation of the bond inquiry and quotation text to be parsed with template matching; uses a bond-domain parsing model to parse the bond inquiry and quotation text to be parsed based on public system prompts, domain context, and parsing basis; and optimizes the structured output by combining confidence evaluation and anomaly completion, and then maps it to the trading system.

[0029] This application, through fine-tuning of a general-purpose large language model for the bond domain and combining techniques such as message segmentation, template retrieval, and confidence assessment, achieves a structured expression of natural language transaction intentions, and has the following significant advantages:

[0030] (1) High accuracy: The semantic parsing accuracy is improved to over 95%, which can effectively handle complex semantic scenarios such as word order changes, synonym substitution, misspellings or colloquialisms, and omitted information, which is significantly better than existing technologies;

[0031] (2) High generalization: By maintaining a standardized template database, the bond domain analytical model can be easily migrated to other scenarios, and can still be easily adapted to new scenarios with insufficient sample data.

[0032] (3) High response speed: The semantic parsing process does not require manual intervention, which greatly shortens the response time and supports high-frequency scenarios;

[0033] (4) Highly structured: Outputs results in a standard format, which can be directly integrated into the trading system without additional format conversion, reducing system integration costs;

[0034] This application realizes intelligentization of the entire bond trading chain, from natural language understanding to standardized execution, and provides reusable core technology modules for intelligent trading systems, which has significant practical application value and market prospects. Attached Figure Description

[0035] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0036] Figure 1 A flowchart illustrating a semantic parsing method based on a fine-tuned large language model according to an embodiment of this application is shown.

[0037] Figure 2 A schematic diagram of the structure of a semantic parsing apparatus based on a fine-tuned large language model according to an embodiment of this application is shown;

[0038] Figure 3 A schematic diagram of the resulting electronic device according to an embodiment of this application is shown. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] The purpose of this application is to provide a semantic parsing method based on fine-tuning a large language model, in order to solve the problems of limited accuracy, weak generalization ability, slow response speed, and low structuredness of existing parsing methods. To achieve the above objective, Figure 1 This application illustrates a semantic parsing method based on a fine-tuned large language model according to an embodiment of the present application, from... Figure 1 As can be seen, this embodiment includes steps S110 to S160:

[0041] Step S110: Collect bond inquiry and quotation text data, and construct a sample set after preprocessing, labeling and data augmentation of the bond inquiry and quotation text data.

[0042] This embodiment begins with sample construction. Real bond inquiry and quotation text data from the interbank market are collected, and after preprocessing, annotation, and data augmentation, a sample set is constructed to train a general-purpose large language model and validate a bond domain parsing model.

[0043] In some optional implementations, step S110 involves collecting bond inquiry and quotation text data, performing preprocessing, annotation, and data augmentation on the bond inquiry and quotation text data, and then constructing a sample set. This includes: collecting bond inquiry and quotation text data; performing desensitization, noise reduction, and unit normalization preprocessing on the bond inquiry and quotation text data; performing sentence segmentation on the preprocessed bond inquiry and quotation text data to obtain key fields, and annotating each key field according to the core elements of the bond field; performing synonym replacement, field reordering, syntactic transformation, and redundancy injection on the annotated bond inquiry and quotation text data to form an annotated enhanced text data; the annotated bond inquiry and quotation text data and the annotated enhanced text data constitute the corpus text; and sorting the inquiry and quotation text data according to the inquiry and quotation text contained in the corpus text. The corpus text is classified according to three dimensions: the number of price items, the number of key fields contained in the corpus text, and the query / quote type of the corpus text. The corpus text is sampled according to the preset classification ratio to form an original sample set. For the original sample set, the number of query / quote items contained in the original sample set exceeds the preset threshold. The number of query / quote items in the original sample set is downsampled to align with the preset threshold to process the original sample set into a short text sample set. The short text sample set is divided into a short text training set and a short text validation set according to the preset ratio. The original samples in the original sample set are concatenated with the goal of the number of tokens not being less than the preset lower limit to process the original sample set into a long text sample set. The long text sample set is divided into a long text training set and a long text validation set according to the preset ratio.

[0044] Sample construction mainly includes the following steps:

[0045] I. Data Collection.

[0046] Collect real bond inquiry and quotation text data from the interbank market. For example, collect approximately 3,000 pieces of bond inquiry and quotation text data from the past three months.

[0047] II. Data Preprocessing.

[0048] The bond inquiry and quotation text data is anonymized. For example, sensitive information such as contact information in the bond inquiry and quotation text data is anonymized to prevent the leakage of sensitive information.

[0049] Denoising is performed on the bond inquiry and quotation text data. For example, meaningless characters or redundant spaces are removed to eliminate interfering content.

[0050] The unit normalization process is applied to the bond inquiry and quotation text data. For example, the monetary units "K", "thousand", "ten thousand", "hundred million", etc. are uniformly converted into the standard unit configured in the business rules. The standard unit can be "ten thousand yuan" or other units.

[0051] III. Data labeling.

[0052] The preprocessed bond inquiry and quotation text data is segmented into sentences to obtain key fields. For example, natural language processing tools are used to segment the preprocessed bond inquiry and quotation text data into sentences.

[0053] Based on the core elements of the bond industry, key fields are manually annotated. For example, based on the 18 core elements of the bond industry listed in Table 1 below, key fields are manually annotated, and then manually verified to ensure the accuracy of the annotation.

[0054] Table 1. 18 Core Elements in the Bond Industry:

[0055]

[0056] IV. Data Augmentation.

[0057] The labeled bond inquiry and quotation text data is expanded to form a corpus text through methods such as synonym substitution, field reordering, syntactic transformation, and redundancy injection. For example, after performing data augmentation on approximately 3,000 labeled bond inquiry and quotation text data, a total of approximately 80,000 corpus text data is obtained, including both labeled bond inquiry and quotation text data and labeled augmented text data.

[0058] V. Stratified sampling.

[0059] The corpus texts are classified according to three dimensions: the number of query and quotation items contained in the corpus text (text length), the number of key fields contained in the corpus text (number of elements), and the query and quotation types in the corpus text.

[0060] In practice, classification based on the number of query and price items contained in the corpus text includes: using 5 query and price items (preset threshold) as the standard for classification. If the number of query and price items in a corpus text exceeds 5, it is a long corpus text; if the number of query and price items in a corpus text does not exceed 5, it is a short corpus text.

[0061] It should be noted that the query / quote item is the smallest semantic unit containing independent bond information. The method for determining the query / quote item in the corpus text is as follows: For a corpus text, the corpus text is divided into blocks, distinguishing between independent lines (containing bond abbreviations and / or bond codes) and shared lines (containing neither bond abbreviations nor bond codes, which are context information), and combining an independent line with its preceding shared line into a query / quote item.

[0062] In practice, classification based on the number of key fields contained in the corpus text includes: using 10 and 15 key fields as the standard for classification, if a corpus text has fewer than 10 key fields, it is a low-feature corpus text; if a corpus text has no less than 10 and no more than 15 key fields, it is a medium-feature corpus text; and if a corpus text has more than 15 key fields, it is a multi-feature corpus text.

[0063] In practice, classification based on the query / quotation type of the corpus text includes: using query / quotation type as unilateral Bid, unilateral Ofr, bilateral, or unknown as the standard for classification.

[0064] For example, a corpus text is ultimately classified as "long text + multi-element corpus text + bilateral".

[0065] The corpus text is stratified and sampled according to a preset classification ratio to form the original sample set. For example, short texts account for 50%, long texts for 50%, few-factor texts for 30%, medium-factor texts for 40%, multi-factor texts for 30%, one-sided (Bid) for 25%, one-sided (Ofr) for 25%, two-sided (Blank) for 25%, and unknown for 25%. After the original sample set is formed, it is further processed into short text sample sets and long text sample sets.

[0066] VI. Construction of short text samples.

[0067] For original samples where the number of inquiry / quotation items exceeds a preset threshold, downsampling is performed to align the number of inquiry / quotation items to the preset threshold. For example, for original samples containing more than 5 inquiry / quotation items, downsampling is performed until only 5 inquiry / quotation items are included.

[0068] The short text sample set is divided into a short text training set and a short text validation set according to a preset ratio (such as 8:2 or 7:3).

[0069] Short text sample sets are used to prevent bias in the analytical model of the bond domain.

[0070] VII. Construction of Long Text Samples

[0071] Perform concatenation on the original samples in the original sample set with the goal of having no less than a preset lower limit of tokens. For example, perform self-repeated concatenation or cross-concatenation on the original samples to form long text samples. Long text samples have no less than 10,000 (preset lower limit) tokens.

[0072] The long text sample set is divided into a long text training set and a long text validation set according to a preset ratio (such as 8:2 or 7:3).

[0073] The long text sample set is used to ensure the parsing ability of the bond domain parsing model for long texts.

[0074] Step S120: Configure the instruction format, and use the sample set and instruction format to perform bond domain supervised fine-tuning on the general large language model to obtain a bond domain analytical model with structured output.

[0075] This embodiment then performs model fine-tuning. A configuration instruction format is provided to explicitly guide the output of the bond domain analytical model, ensuring it produces structured output results that meet the requirements.

[0076] In some optional implementations, step S120, configuring the instruction format, and using the sample set and instruction format to perform supervised fine-tuning of the general-purpose large language model in the bond domain to obtain a bond domain parsing model with structured output, includes: configuring the instruction format containing task description, core element definition, and output template; using a general-purpose large language model with small parameters as a base, inputting a short text training set, a long text training set, and the instruction format to perform supervised fine-tuning of the general-purpose large language model, and inputting a short text validation set, a long text validation set, and the instruction format to perform performance verification of the supervised fine-tuned general-purpose large language model to obtain a bond domain parsing model with structured output in JSON Schema; wherein, the objective function of supervised fine-tuning includes parsing accuracy constraints and structure consistency constraints.

[0077] To ensure that the general language model clearly defines the task objectives, element standards, and output specifications for bond inquiry and quotation semantic parsing, an instruction format is configured that includes three parts: task description, core element definition, and output template.

[0078] The task description clearly defines the core tasks to be completed, avoiding ambiguity. For example, "Perform semantic parsing of natural language inquiry and quotation texts in the interbank bond market, extract 18 core elements in the bond field, generate structured data that meets the access requirements of the trading system, ensure that no elements are omitted or incorrect, and the output format must strictly follow JSON Schema."

[0079] The definition of core elements provides a detailed explanation of the value range and identification characteristics of core elements in the bond field. For example, please refer to the explanatory examples in Table 1 above.

[0080] The output template clearly defines the standards for structured output, specifying the key names and data types of each core element to ensure consistency in structured output.

[0081] For example, the following is an example output template:

[0082] "Bond Abbreviation": "String type"

[0083] "Bond Code": "String type"

[0084] "Quote direction": "String type", "Value range: bid / ofr / double / unknown"

[0085] "Buy Quantity": "Number Type", "Unit: 10,000 Yuan"

[0086] "Sales Volume": "Number Type", "Unit: Ten Thousand Yuan"

[0087] "Buy Price": "Number type", "Keep 4 decimal places"

[0088] "Selling Price": "Number type", "Keep 4 decimal places"

[0089] "Buy Price Type": "String type", "Value range: NET / Dirty / Yield / Spread / Intentional"

[0090] "Selling price type": "String type", "Value range: NET / Dirty / Yield / Spread / Intentional"

[0091] "Request a quote to buy": "Boolean type", "Value range: true / false"

[0092] "Request a quote to sell": "Boolean type", "Value range: true / false"

[0093] "Custom preferences": "String type"

[0094] "Mid-price": "Numerical type"

[0095] "Execution Price": "Numerical Type"

[0096] "Due Date": "Date Type", "Format: YYYY-MM-DD"

[0097] "Coupon Rate": "Number Type", "Unit: %", "Retain 4 decimal places"

[0098] "Buy Reference Price": "Number Type"

[0099] "Sell Reference Price": "Number Type".

[0100] A general-purpose large language model with small parameters (approximately 1.5B in size) is selected as the base, such as Qwen2.5-1.5B-Instruct. Choosing a general-purpose large language model with small parameters can significantly reduce computational requirements while maintaining parsing accuracy.

[0101] Short text training sets, long text training sets, and instruction formats are input into a general large language model to perform supervised fine-tuning (SFT) in the bond domain. The objective function of the supervised fine-tuning includes parsing accuracy constraints and structure consistency constraints.

[0102] The analytical accuracy constraint is calculated using the analytical accuracy loss function, which calculates the matching error between the predicted core elements and the labeled core elements.

[0103] The structured consistency constraint is calculated using the structured consistency loss function, which calculates the matching error between the output format and the output template.

[0104] During supervised fine-tuning, an early stopping strategy can be incorporated to prevent overfitting. Short text validation sets, long text validation sets, and instruction formats are input into the supervised fine-tuned general-purpose large language model for performance verification. When the performance verification meets the specified criteria, a bond domain analytical model is obtained. For example, accuracy ≥ 95%, recall ≥ 95%, and F1 ≥ 95%.

[0105] Table 2 below shows a performance comparison of the bond domain parsing model with other general large language models in the bond inquiry and pricing domain.

[0106] Table 2. Performance comparison of each model:

[0107]

[0108] Step S130: Extract the query and quotation items for each sample in the sample set, perform special label replacement on the query and quotation items, and construct a standardized template database.

[0109] This embodiment then constructs a standardized template database. By maintaining this standardized template database, the bond domain analytical model can be easily migrated to other scenarios (such as interest rate bonds, credit bonds, derivatives), improving its adaptability to new scenarios with insufficient sample data.

[0110] In some optional implementations, step S130, extracting the inquiry and quotation items for each sample in the sample set, and constructing a standardized template database after performing special mark replacement on the inquiry and quotation items, includes: dividing each sample in the sample set into blocks, distinguishing the independent rows and shared rows of each sample in the sample set, and combining an independent row with the preceding shared row of that independent row into an inquiry and quotation item; wherein, the independent row contains the bond abbreviation and / or bond code, and the shared row contains neither the bond abbreviation nor the bond code; replacing the English characters, Chinese characters, and numbers in the inquiry and quotation items with exclusive special marks to form the original template; and performing deduplication on each original template to form a standardized template database.

[0111] The method for determining the query and quotation items for each sample is the same as that for the query and quotation items in the above corpus text: For a sample, the sample is divided into blocks, and independent lines (containing bond abbreviations and / or bond codes) and shared lines (containing neither bond abbreviations nor bond codes, which are context information) are distinguished. An independent line and its preceding shared line are combined into a query and quotation item.

[0112] The process of applying special markings involves replacing English characters, Chinese characters, numbers, etc., in the quotation items with the corresponding special markings. For example, replacing the bond code in "specific number.market suffix" with "[bond code]", and replacing the median price of the value with "[amount]", etc.

[0113] The original template is formed by replacing the special tags in each query and quotation item. Since there may be duplicates in the original template, a standardized template database is constructed after deduplication of the original template.

[0114] Step S140: Receive the bond inquiry and quotation text to be parsed, extract the inquiry and quotation items to be parsed from the bond inquiry and quotation text, and select matching templates from the standardized template database based on the inquiry and quotation items to be parsed as the basis for parsing.

[0115] This embodiment continues with template retrieval. Based on a standardized template database, the closest matching template is found for the bond inquiry and quotation text to be parsed, thereby improving the bond domain parsing model's ability to understand the bond inquiry and quotation text.

[0116] In some optional implementations, step S140, receiving the bond inquiry and quotation text to be parsed, extracting the inquiry and quotation items to be parsed from the bond inquiry and quotation text to be parsed, and selecting matching templates from the standardized template database based on the inquiry and quotation items to be parsed, includes: receiving the bond inquiry and quotation text to be parsed; dividing the bond inquiry and quotation text to be parsed into blocks, distinguishing between the independent lines to be parsed and the shared lines to be parsed, combining an independent line to be parsed with its preceding shared line to be parsed into an inquiry and quotation item to be parsed; for an inquiry and quotation item to be parsed, using the longest common subsequence algorithm to calculate the similarity between the inquiry and quotation item to be parsed and each template in the standardized template database, selecting candidate templates with similarity higher than the similarity threshold and ranking in the first preset position, using the maximum marginal relevance algorithm to re-rank the candidate templates, and selecting the template ranking in the second preset position as the matching template.

[0117] The method for determining the query and quotation items in the bond query and quotation text to be parsed is the same as the method for determining the query and quotation items in the above-mentioned corpus text: For a bond query and quotation text to be parsed, the text is divided into blocks, and independent lines (containing bond abbreviations and / or bond codes) and shared lines (containing neither bond abbreviations nor bond codes, which are context information) are distinguished. An independent line and its preceding shared line are combined into a query and quotation item to be parsed.

[0118] For a query to be parsed, the Longest Common Subsequence (LCS) algorithm is used to calculate the similarity between the query to be parsed and each template in the standardized template database. Candidate templates with similarity higher than the similarity threshold (e.g., LCS-sim≥0.6) and ranking (e.g., Top-k) belonging to the first preset position (e.g., k=20) are selected.

[0119] The candidate templates are re-ranked using the maximum marginal relevance algorithm (MMR, lambda=0.7), and the matching templates whose ranking (e.g., Top-k') belongs to the second-to-last preset position (e.g., k'=5) are selected as the basis for analysis.

[0120] Step S150: Configure public system prompts, form a domain context based on the query and quotation items to be parsed, and use the bond domain parsing model to parse the bond query and quotation text to be parsed according to the public system prompts, domain context and parsing criteria to generate structured output.

[0121] This embodiment then proceeds to perform model inference. The bond inquiry and quotation text to be parsed is used as input to the bond domain parsing model. Based on shared public system hints and domain context knowledge, the bond domain parsing model parses the bond inquiry and quotation text to be parsed, generating structured output.

[0122] In some optional implementations, step S150 involves configuring common system prompts, forming a domain context based on the query and quotation items to be parsed, and using a bond domain parsing model to parse the bond query and quotation text to be parsed according to the common system prompts, domain context, and parsing criteria to generate structured output. This includes: determining the batch size of the bond query and quotation text to be parsed based on the request queue length; configuring common system prompts for each batch that include the task objective, core requirements, and output format; concatenating each query and quotation item to be parsed in each batch into a knowledge context for each batch, and concatenating the knowledge context of each batch with general bond domain knowledge into a domain context for each batch; and using the bond domain parsing model, the common system prompts and domain context of each batch, and parsing the bond query and quotation text to be parsed in each batch according to the parsing criteria to generate JSON Schema structured output.

[0123] Based on the request queue length, the batch size of the bond inquiry and quotation texts to be parsed is dynamically adjusted to achieve parallel parsing and improve overall throughput. For example, under low load (request queue length < 20), the batch size is automatically reduced to process 5-10 bond inquiry and quotation texts to be parsed each time to reduce the waiting time of individual requests; under medium load (20 ≤ request queue length ≤ 50), a medium batch size is maintained to process 10-20 bond inquiry and quotation texts to be parsed each time to balance throughput and latency; under high load (request queue length > 50), the batch size is automatically increased to process 20-30 bond inquiry and quotation texts to be parsed each time to maximize computing power utilization and avoid congestion.

[0124] Configure common system prompts, which represent fixed instructions shared by all batches of bond inquiry and quotation texts to be parsed. These prompts can include the task objective, core requirements, and output format. The common system prompts can be the same as or different from the instruction format.

[0125] Each batch of query quotes to be parsed is concatenated into a knowledge context, which provides contextual information within each batch. This knowledge context is then combined with general knowledge in the bond industry to form a domain context.

[0126] The bond domain parsing model shares common system hints and domain context from each batch, makes full use of prefixes and key-value caches, reduces redundant calculations, thereby improving inference speed and controlling the average response latency to the second level.

[0127] The bond domain parsing model combines public system prompts and domain context for each batch, and performs parallel parsing of the bond inquiry and quotation texts to be parsed for each batch according to the parsing criteria, ultimately generating JSON Schema structured output.

[0128] Step S160: Optimize the structured output based on confidence evaluation and anomaly completion, and map the optimized structured output to the trading system.

[0129] This embodiment concludes with optimization and transaction mapping. The structured output generated by the bond domain analytical model is further optimized, and the optimized structured output is directly connected to the trading system to achieve end-to-end intelligent management of bond trading.

[0130] In some optional implementations, step S160, optimizing the structured output based on confidence evaluation and anomaly completion, and mapping the optimized structured output to the transaction system, includes: performing validation on the JSON Schema structured output; wherein, the validation includes: core element integrity validation, data type compliance validation, and value range rationality validation; if the validation fails, a confidence threshold is determined based on the minimum logarithmic probability of the JSON Schema structured output sequence, and output content below the confidence threshold is removed to form a preliminary optimization result; anomaly completion is performed on the preliminary optimization result based on preset business rules or manual review to form an optimized JSON Schema structured output; the optimized JSON Schema structured output is output to the transaction system through an interface, and the optimized JSON Schema structured output is mapped to the transaction system template according to the mapping rules.

[0131] Perform schema validation on the JSON schema structured output generated by the bond domain parsing model. The validation includes the following three aspects:

[0132] First, verify the integrity of core elements. This means verifying whether the JSON Schema structured output covers the 18 core elements.

[0133] Second, data type compliance verification. This involves verifying whether the data type conforms to the definition; for example, whether the bond abbreviation is a string type.

[0134] Third, verify the reasonableness of the value range. That is, verify whether the value range is within a reasonable range. For example, the quotation direction belongs to one of [bid (buy), ofr (sell), double (two-way), unknown (unknown)].

[0135] If the validation is successful, no further optimization is needed, and the JSON Schema structured output can be directly mapped to the transaction system.

[0136] If validation fails, optimization steps are required. Calculate the confidence level of each output token based on the minimum logarithmic probability of the JSON Schema structured output sequence (lower probability equals lower confidence). Set a confidence threshold (e.g., 0.8), and exclude output tokens with confidence levels below this threshold from output, thus generating preliminary optimization results.

[0137] The system completes the missing information in the initial optimization results based on preset business rules. For example, when "+1" appears alone, the transaction date is set to the current day and the settlement speed is set to T+1.

[0138] In some cases, the preset business rules cannot cover all missing content completion situations. In such cases, the missing content will be pushed to manual review, and the operator will fill in the missing content.

[0139] After exception completion, the final output is an optimized JSON Schema structured output.

[0140] The optimized JSON Schema structured output is sent to the trading system via an API interface. According to the mapping rules, the optimized JSON Schema structured output corresponds one-to-one with the trading system template. This allows the trading system to automatically generate fixed prices, bilateral or consolidated credit bond quotes, etc., and call the internal pricing engine to return quotes in real time and trigger intended trades. This achieves a closed loop from natural language input to structured parsing and trade execution.

[0141] Figure 2 This application illustrates a semantic parsing apparatus based on a fine-tuned large language model according to an embodiment of the present application, from... Figure 2 It can be seen that the semantic parsing device 200 based on fine-tuning a large language model includes:

[0142] Data construction unit 210 is used to collect bond inquiry and quotation text data, and to construct a sample set after performing preprocessing, annotation and data augmentation on the bond inquiry and quotation text data.

[0143] Model training unit 220 is used to configure the instruction format. After performing bond domain supervised fine-tuning on the general large language model using the sample set and instruction format, a bond domain analytical model with structured output is obtained.

[0144] Template construction unit 230 is used to extract the inquiry and quotation items of each sample in the sample set, and construct a standardized template database after performing special mark replacement on the inquiry and quotation items; wherein, the inquiry and quotation item is the smallest semantic unit containing independent bond information;

[0145] The template matching unit 240 is used to receive the bond inquiry and quotation text to be parsed, extract the inquiry and quotation items to be parsed from the bond inquiry and quotation text, and select matching templates from the standardized template database based on the inquiry and quotation items to be parsed as the basis for parsing.

[0146] Semantic parsing unit 250 is used to configure public system prompts, form a domain context based on the query and quotation items to be parsed, and use the bond domain parsing model to perform parsing on the bond query and quotation text to be parsed according to the public system prompts, domain context and parsing criteria to generate structured output;

[0147] The optimization docking unit 260 is used to optimize the structured output based on confidence evaluation and anomaly completion, and then map the optimized structured output to the transaction system.

[0148] In some optional embodiments, in the above apparatus, the data construction unit 210 includes: a data preprocessing module, used to collect bond inquiry and quotation text data and perform desensitization, noise reduction, and unit normalization preprocessing on the bond inquiry and quotation text data; a data annotation module, used to perform sentence segmentation on the preprocessed bond inquiry and quotation text data to obtain each key field, and to annotate each key field according to the core elements of the bond field; a data enhancement module, used to perform synonym replacement, field reordering, syntactic transformation, and redundancy injection on the annotated bond inquiry and quotation text data to form annotated enhanced text data, the annotated bond inquiry and quotation text data and the annotated enhanced text data constituting the corpus text; and a stratified sampling module, used to sample according to the number of inquiry and quotation items contained in the corpus text. The system categorizes corpus texts based on three dimensions: quantity, number of key fields contained in the corpus texts, and query / quote types. It then samples the corpus texts according to a preset classification ratio to form an original sample set. The sample construction module is used to downsample the original samples in the original sample set where the number of query / quote items exceeds a preset threshold, aligning the number of query / quote items to the preset threshold to process the original sample set into a short text sample set. The short text sample set is then divided into a short text training set and a short text validation set according to a preset ratio. Finally, the system performs concatenation on the original samples in the original sample set, aiming to ensure the number of tokens is not less than a preset lower limit, to process the original sample set into a long text sample set. The long text sample set is then divided into a long text training set and a long text validation set according to a preset ratio.

[0149] In some optional implementations, in the above apparatus, the model training unit 220 is used to: configure an instruction format including task description, core element definition, and output template; use a general-purpose large language model with small parameters as a base, input short text training sets, long text training sets, and instruction format to perform supervised fine-tuning on the general-purpose large language model, input short text validation sets, long text validation sets, and instruction format to perform performance verification on the supervised fine-tuned general-purpose large language model, and obtain a bond domain parsing model that generates JSON Schema structured output; wherein, the objective function of supervised fine-tuning includes parsing accuracy constraints and structure consistency constraints.

[0150] In some optional embodiments, in the above apparatus, the template construction unit 230 is used to: divide each sample in the sample set into blocks, distinguish the independent rows and shared rows of each sample in the sample set, combine an independent row with the preceding shared row of the independent row into a query quotation item; replace the English characters, Chinese characters and numbers in the query quotation item with exclusive special marks to form an original template; and perform deduplication on each original template to form a standardized template database.

[0151] In some optional embodiments, in the above apparatus, the template matching unit 240 is used to: receive the bond inquiry and quotation text to be parsed; divide the bond inquiry and quotation text to be parsed into blocks, distinguish the independent lines to be parsed and the shared lines to be parsed, and combine an independent line to be parsed with the preceding shared line to be parsed into a quotation item to be parsed; for a quotation item to be parsed, use the longest common subsequence algorithm to calculate the similarity between the quotation item to be parsed and each template in the standardized template database, select candidate templates with similarity higher than the similarity threshold and ranking in the first preset position, use the maximum marginal relevance algorithm to re-rank the candidate templates, and select the template ranking in the second preset position as the matching template.

[0152] In some optional implementations, in the above apparatus, the semantic parsing unit 250 is used to: determine the batch size of the bond inquiry and quotation text to be parsed based on the request queue length; configure common system prompts for each batch, including task purpose, core requirements, and output format; concatenate each inquiry and quotation item to be parsed in each batch into the knowledge context of each batch, and concatenate the knowledge context of each batch with general knowledge in the bond domain into the domain context of each batch; and use the bond domain parsing model, the common system prompts of each batch, and the domain context to perform parallel parsing on the bond inquiry and quotation text to be parsed in each batch according to the parsing criteria, generating JSON Schema structured output.

[0153] In some optional implementations, in the above apparatus, the optimization docking unit 260 is used to: perform validation on the JSON Schema structured output; wherein, the validation includes: core element integrity validation, data type compliance validation, and value range rationality validation; if the validation fails, a confidence threshold is determined based on the minimum logarithmic probability of the JSON Schema structured output sequence, and output content below the confidence threshold is removed to form a preliminary optimization result; based on preset business rules or manual review, anomaly completion is performed on the preliminary optimization result to form an optimized JSON Schema structured output; the optimized JSON Schema structured output is output to the transaction system through the interface, and the optimized JSON Schema structured output is mapped to the transaction system template according to the mapping rules.

[0154] It should be noted that the semantic parsing device 200 based on the fine-tuned large language model can implement the aforementioned semantic parsing method based on the fine-tuned large language model, which will not be elaborated further.

[0155] Figure 3 This invention illustrates a schematic diagram of the structure of an electronic device according to an embodiment of the present application. Figure 3As shown, the electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external devices via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a semantic parsing method based on fine-tuning a large language model.

[0156] In one embodiment, the electronic device provided in this application includes a memory and a processor. The memory stores a database and a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the aforementioned semantic parsing method based on fine-tuning a large language model.

[0157] The above is as stated in this application. Figure 2 The method executed by the semantic parsing device based on fine-tuning a large language model disclosed in the illustrated embodiments can be applied to a processor or implemented by a processor. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The steps of the method disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0158] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the aforementioned semantic parsing method based on fine-tuning a large language model.

[0159] It should be noted that the functions or steps that the above-mentioned electronic devices or computer-readable storage media can achieve can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0160] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0161] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0162] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A semantic parsing method based on fine-tuning a large language model, characterized in that, include: Collect bond inquiry and quotation text data, and construct a sample set after performing preprocessing, annotation, and data augmentation on the bond inquiry and quotation text data; Configure the instruction format, and use the sample set and instruction format to perform bond domain supervised fine-tuning on the general large language model to obtain a bond domain parsing model that generates structured output; Extract the query and price items from each sample in the sample set, and construct a standardized template database after performing special tag replacement on the query and price items; where the query and price item is the smallest semantic unit containing independent bond information; Receive the bond inquiry and quotation text to be parsed, extract the inquiry and quotation items to be parsed from the bond inquiry and quotation text, and select matching templates from the standardized template database based on the inquiry and quotation items to be parsed as the basis for parsing; Configure public system prompts, form a domain context based on the query and quotation items to be parsed, and use the bond domain parsing model to parse the bond query and quotation text to be parsed according to the public system prompts, domain context and parsing criteria, and generate structured output; The structured output is optimized based on confidence evaluation and anomaly completion, and then mapped to the trading system. The process of collecting bond inquiry and quotation text data, performing preprocessing, annotation, and data augmentation on the bond inquiry and quotation text data, and then constructing a sample set includes: Collect bond inquiry and quotation text data, and perform preprocessing such as desensitization, noise reduction, and unit normalization on the bond inquiry and quotation text data; After performing sentence segmentation on the preprocessed bond inquiry and quotation text data, each key field is obtained, and each key field is annotated according to the core elements of the bond field. After performing synonym replacement, field reordering, syntactic transformation and redundancy injection on the annotated bond inquiry and quotation text data, the annotated enhanced text data is formed. The annotated bond inquiry and quotation text data and the annotated enhanced text data constitute the corpus text. The corpus texts are classified according to three dimensions: the number of query and quotation items contained in the corpus texts, the number of key fields contained in the corpus texts, and the query and quotation types in the corpus texts. The corpus texts are then sampled according to the preset classification ratio to form the original sample set. For original samples whose number of query and quotation items exceeds a preset threshold, downsampling is performed to align the number of query and quotation items to the preset threshold to process the original sample set into a short text sample set. The short text sample set is then divided into a short text training set and a short text validation set according to a preset ratio. The original samples in the original sample set are concatenated with the goal of having no less than a preset lower limit of token quantity to process the original sample set into a long text sample set. The long text sample set is then divided into a long text training set and a long text verification set according to a preset ratio. The configuration instruction format, which utilizes a sample set and instruction format to perform bond domain-supervised fine-tuning on a general large language model, yields a bond domain parsing model that generates structured output, including: The configuration includes the instruction format, which includes task description, core element definition, and output template. Using a general-purpose large language model with small parameters as a base, supervised fine-tuning is performed on the general-purpose large language model by inputting short text training sets, long text training sets, and instruction formats. Performance verification is performed on the supervised fine-tuned general-purpose large language model by inputting short text validation sets, long text validation sets, and instruction formats, resulting in a bond domain parsing model that generates JSON Schema structured output. The objective function of supervised fine-tuning includes parsing accuracy constraints and structure consistency constraints.

2. The method according to claim 1, characterized in that, The query and quotation items for each sample in the extracted sample set are used to construct a standardized template database after performing special tag replacement on the query and quotation items, including: Each sample in the sample set is divided into blocks, and independent rows and shared rows of each sample in the sample set are distinguished. An independent row and its preceding shared row are combined into a query / quote item. The independent row contains the bond abbreviation and / or bond code, while the shared row contains neither the bond abbreviation nor the bond code. Replace the English characters, Chinese characters, and numbers in the inquiry and quotation items with unique special marks to form the original template; After deduplication of each original template, a standardized template database is formed.

3. The method according to claim 2, characterized in that, The process of receiving the bond inquiry and quotation text to be parsed, extracting the inquiry and quotation items to be parsed from the text, and selecting matching templates from the standardized template database based on the inquiry and quotation items to be parsed includes: Receive the bond inquiry and quotation text to be parsed; The bond inquiry and quotation text to be parsed is divided into blocks, and the independent lines to be parsed and the shared lines to be parsed are distinguished. An independent line to be parsed and its preceding shared line to be parsed are combined into a single inquiry and quotation item to be parsed. For a query to be parsed, the longest common subsequence algorithm is used to calculate the similarity between the query to be parsed and each template in the standardized template database. Candidate templates with similarity higher than the similarity threshold and ranking in the first preset position are selected. The maximum marginal relevance algorithm is used to re-rank the candidate templates, and the template ranking in the second preset position is selected as the matching template.

4. The method according to claim 3, characterized in that, The configuration public system prompts, based on the domain context formed by the query and quotation items to be parsed, utilize the bond domain parsing model to parse the bond query and quotation text according to the public system prompts, domain context, and parsing criteria, generating structured output, including: The batch size of the bond inquiry and quotation text to be parsed is determined based on the length of the request queue; Configure common system prompts for each batch, including task objectives, core requirements, and output format; Each query and quotation item to be parsed in each batch is concatenated into the knowledge context of each batch, and the knowledge context of each batch is concatenated with general knowledge in the bond field to form the domain context of each batch. Using the bond domain parsing model, public system prompts and domain context for each batch, parallel parsing is performed on the bond inquiry and quotation text to be parsed for each batch according to the parsing criteria, generating JSON Schema structured output.

5. The method according to claim 4, characterized in that, The optimization of the structured output based on confidence evaluation and anomaly completion maps the optimized structured output to the trading system, including: Perform validation on the JSON Schema structured output; the validation includes: core element integrity validation, data type compliance validation, and value range reasonableness validation. If the verification fails, the confidence threshold is determined based on the minimum log probability of the JSON Schema structured output sequence. Output content below the confidence threshold is then removed to form a preliminary optimization result. Based on preset business rules or manual review, the preliminary optimization results are filled with anomalies to form an optimized JSONSchema structured output; The optimized JSON Schema structured output is sent to the trading system through an interface, and the optimized JSON Schema structured output is matched with the trading system template according to the mapping rules.

6. A semantic parsing device based on fine-tuning a large language model, characterized in that, include: The data construction unit is used to collect bond inquiry and quotation text data, and to construct a sample set after performing preprocessing, annotation and data augmentation on the bond inquiry and quotation text data. The model training unit is used to configure the instruction format. After performing bond domain-supervised fine-tuning on the general large language model using the sample set and instruction format, a bond domain analytical model with structured output is obtained. The template construction unit is used to extract the query and quotation items of each sample in the sample set, and construct a standardized template database after performing special tag replacement on the query and quotation items; wherein, the query and quotation item is the smallest semantic unit containing independent bond information; The template matching unit is used to receive the bond inquiry and quotation text to be parsed, extract the inquiry and quotation items to be parsed from the bond inquiry and quotation text, and select matching templates from the standardized template database based on the inquiry and quotation items to be parsed as the basis for parsing. The semantic parsing unit is used to configure public system prompts, form a domain context based on the query and quotation items to be parsed, and use the bond domain parsing model to parse the bond query and quotation text to be parsed according to the public system prompts, domain context and parsing criteria, and generate structured output; The optimization docking unit is used to optimize the structured output based on confidence evaluation and anomaly completion, and then map the optimized structured output to the transaction system; The data construction unit includes: The data preprocessing module is used to collect bond inquiry and quotation text data and perform desensitization, noise reduction, and unit normalization preprocessing on the bond inquiry and quotation text data. The data annotation module is used to segment the preprocessed bond inquiry and quotation text data into sentences to obtain key fields, and to annotate each key field according to the core elements of the bond field. The data augmentation module is used to perform synonym replacement, field reordering, syntactic transformation and redundancy injection on the labeled bond inquiry and quotation text data to form labeled augmented text data. The labeled bond inquiry and quotation text data and the labeled augmented text data constitute the corpus text. The stratified sampling module is used to classify the corpus text according to three dimensions: the number of query and quotation items contained in the corpus text, the number of key fields contained in the corpus text, and the query and quotation type of the corpus text. It then performs sampling on the corpus text according to the preset classification ratio to form the original sample set. The sample construction module is used to perform downsampling on the original samples in the original sample set where the number of query and quotation items exceeds a preset threshold, aligning the number of query and quotation items to the preset threshold to process the original sample set into a short text sample set. The short text sample set is then divided into a short text training set and a short text verification set according to a preset ratio. Additionally, the module performs concatenation on the original samples in the original sample set with the goal of ensuring that the number of tokens is not less than a preset lower limit to process the original sample set into a long text sample set. The long text sample set is then divided into a long text training set and a long text verification set according to a preset ratio. The model training unit is used to configure the instruction format, which includes task descriptions, core element definitions, and output templates. A general-purpose large language model with small parameters is used as the base. Short text training sets, long text training sets, and the instruction format are input to perform supervised fine-tuning of the general-purpose large language model. Short text validation sets, long text validation sets, and the instruction format are input to perform performance verification of the supervised fine-tuned general-purpose large language model, resulting in a bond domain parsing model that generates JSON Schema structured output. The objective function of the supervised fine-tuning includes parsing accuracy constraints and structure consistency constraints.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the semantic parsing method based on fine-tuning a large language model as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the semantic parsing method based on fine-tuning a large language model as described in any one of claims 1 to 5.

Images