Contract appendix attribution method, attribution device, attribution system and storage medium
Through multi-stage language model processing and manual annotation optimization, the system has solved the problem of efficient and accurate analysis of commercial investment contract remarks data, realized automatic identification and multi-level attribution of discounted contract situations, and improved analysis efficiency and accuracy.
Patent Information
- Application Number
- CN202510976446.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-15
- Publication Date
- 2025-12-19
AI Technical Summary
Existing technologies struggle to efficiently and accurately process the appendix data of large volumes of commercial investment contracts, identify discounted signing situations, extract evidence, and conduct standardized, hierarchical attribution analysis. This results in challenges such as difficulty in information extraction, inconsistent attribution, difficulty in uncovering deeper causes, and difficulties in knowledge accumulation and reuse.
A multi-stage, collaborative language model processing workflow is adopted, including preprocessing, discount contract identification and evidence extraction, and multi-level attribution catalog determination. The contract appendix text is analyzed through a large-scale language model to gradually generate structured three-level discount attribution results. The model is optimized by combining manually annotated reasonable attribution examples.
It enables efficient and automatic identification of discounted contract signing situations in contract appendices and accurate extraction of evidence information, improving the efficiency and quality of attribution analysis and providing enterprises with more accurate business insights.
Smart Images
Figure CN121168412A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to a contract appendix attribution method, a contract appendix attribution device, a contract appendix attribution system, and a computer-readable storage medium. Background Technology
[0002] In the development of commercial real estate, a large number of signed and archived commercial leasing contracts have been accumulated. These contracts are the core documents binding the rights and obligations of both parties. In addition to the main contract, contract appendices often record supplements, modifications, or special agreements for specific situations, which may contain important discount information, such as rent reductions, extensions of rent-free periods, and additional subsidies. Analyzing these appendices to identify discounted contracts and their reasons is crucial for companies' cost control, risk assessment, leasing strategy optimization, and business decision-making. With the expansion of companies and the surge in the number of contracts, efficiently and accurately processing this appendice data has become a major challenge. Summary of the Invention
[0003] This application provides a contract appendix attribution method, a contract appendix attribution device, a contract appendix attribution system, and a computer-readable storage medium to solve at least one of the aforementioned technical problems.
[0004] The contract appendix attribution method of the embodiments of this application includes:
[0005] The original contract appendix text is preprocessed to obtain the current contract appendix text;
[0006] The current contract appendix text is analyzed using a predefined language model to identify discounted contracts and extract discounted evidence, thereby obtaining discounted contract evidence information and generating preliminary reasons for discounted attribution.
[0007] The predetermined language model determines a multi-level discount attribution catalog based on the current contract appendix text, the preliminary discount attribution reasons, and the discount contract evidence information;
[0008] The discount attribution result of the original contract appendix text is determined based on the original contract appendix text, the discounted contract evidence information, and the multi-level discount attribution directory.
[0009] In some implementations, the preprocessing of the original contract appendix text to obtain the current contract appendix text includes:
[0010] The original contract appendix text is cleaned to filter out noise content, resulting in the current contract appendix text. The noise content includes any one or more of the following: leased resource space, leased area, renovation period project, project conditions, and deduction points.
[0011] In some implementations, the predetermined language model includes a first language model module. The step of identifying discounted contracts and extracting discounted evidence from the current contract appendix text using the predetermined language model to obtain discounted contract evidence information and generate preliminary reasons for discounted attribution includes:
[0012] The first language model module determines whether the current contract appendix text is a discounted contract.
[0013] If the current contract appendix text is a discounted contract, extract the discounted contract evidence information from the current contract appendix text;
[0014] The preliminary reasons for the discount attribution are generated based on the evidence information regarding the discounted contract.
[0015] In some implementations, the predetermined language model includes a second language model module, a third language model module, and a fourth language model module. The step of determining a multi-level discount attribution catalog based on the current contract appendix text, the preliminary discount attribution reasons, and the discount signing evidence information using the predetermined language model includes:
[0016] A predefined primary discount attribution catalog, which includes market factors, store location factors, brand factors, and negotiation factors;
[0017] Using the second language model module, based on the current contract appendix text, the preliminary discount attribution reasons, and the discount signing evidence information, the current contract appendix text is categorized into the corresponding first-level discount attribution directory in the predefined first-level discount attribution category, and the corresponding first-level discount attribution directory is used as the current first-level discount attribution directory;
[0018] Through the third language model module, based on the current contract appendix text, the current first-level discount attribution directory, the preliminary discount attribution reasons, and the discount contract evidence information, the current contract appendix text is generated or mapped to the current second-level discount attribution directory under the current first-level discount attribution directory;
[0019] The fourth language model module generates or maps the current contract postscript text to the current tertiary discount attribution directory under the current tertiary discount attribution directory based on the current contract postscript text, the current secondary discount attribution directory, the preliminary discount attribution reasons, and the discount contract evidence information, and determines the corresponding current tertiary discount attribution description.
[0020] In some implementations, determining the multi-level discount attribution result of the original contract appendix text based on the original contract appendix text, the discounted contract evidence information, and the multi-level discount attribution directory includes:
[0021] The original contract appendix text, the discounted contract evidence information, the current first-level discounted price attribution directory, the current second-level discounted price attribution directory, the current third-level discounted price attribution directory, and the corresponding current third-level discounted price attribution description are structurally integrated to form the third-level discounted price attribution result of the original contract appendix text.
[0022] In some implementations, the contract appendix attribution method further includes:
[0023] Obtain reasonable attribution examples from manually annotated files;
[0024] The reasonable attribution examples are dynamically added to the prompt words of the predetermined language model, wherein the prompt words are used to guide the predetermined language model to determine the multi-level discount attribution catalog.
[0025] In some implementations, the contract appendix attribution method further includes:
[0026] Traverse multiple original contract appendix texts to determine the multi-level discount attribution directory corresponding to each original contract appendix text;
[0027] Based on the multi-level discount attribution directory corresponding to each original contract appendix text, update the global discount attribution directory corresponding to multiple original contract appendix texts according to the hierarchical nesting rules.
[0028] The contract appendix attribution device according to the embodiments of this application includes:
[0029] The preprocessing module is used to preprocess the original contract appendix text to obtain the current contract appendix text;
[0030] The preliminary attribution module is used to identify discounted contracts and extract discounted evidence from the current contract appendix text using a predefined language model, obtain discounted contract evidence information, and generate preliminary attribution reasons for the discounted contracts.
[0031] The multi-level attribution module is used to determine a multi-level discount attribution catalog based on the current contract appendix text, the preliminary discount attribution reasons, and the discount contract evidence information through the predetermined language model;
[0032] The attribution determination module is used to determine the discount attribution result of the original contract appendix text based on the original contract appendix text, the discounted contract evidence information, and the multi-level discount attribution directory.
[0033] The contract appendix attribution system of this application includes one or more processors and a memory. The memory stores a computer program, which, when executed by the processor, implements the contract appendix attribution method of any of the above embodiments.
[0034] The computer-readable storage medium of the present application embodiments stores a computer program that, when executed by a processor, implements the contract appendix attribution method of any of the above embodiments.
[0035] The contract appendix attribution method, device, system, and computer-readable storage medium described in this application construct a multi-stage, sequential, and collaborative language model processing flow. This flow progressively completes the process from identifying discounted contracts and extracting evidence to determining a multi-level discount attribution catalog. This enables efficient and automatic identification of discounted contract situations in contract appendices and accurate extraction of relevant evidence information, as well as automated, standardized, and multi-level attribution analysis of the reasons for discounted contracts. This application can improve the efficiency and quality of large-scale contract appendix attribution analysis, providing enterprises with more accurate and in-depth business insights.
[0036] Additional aspects and advantages of embodiments of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of embodiments of this application. Attached Figure Description
[0037] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein:
[0038] Figure 1 This is a flowchart illustrating the contract appendix attribution method of certain embodiments of this application;
[0039] Figure 2 This is a schematic diagram of the overall working process of the contract appendix attribution method in some embodiments of this application;
[0040] Figure 3 This is a flowchart illustrating the contract appendix attribution method of certain embodiments of this application;
[0041] Figure 4 This is a schematic diagram of a predetermined language model for certain embodiments of this application;
[0042] Figure 5 This is a flowchart illustrating the contract appendix attribution method of certain embodiments of this application;
[0043] Figure 6 This is a flowchart illustrating the contract appendix attribution method of certain embodiments of this application;
[0044] Figure 7 This is a flowchart illustrating the contract appendix attribution method of certain embodiments of this application;
[0045] Figure 8 This is a schematic diagram of the contract appendix attribution device for certain embodiments of this application;
[0046] Figure 9 This is a schematic diagram of the contract appendix attribution system in some embodiments of this application;
[0047] Figure 10 This is a schematic diagram illustrating the connection state between a computer-readable storage medium and a processor according to certain embodiments of this application.
[0048] Explanation of reference numerals in the attached figures:
[0049] The system includes a contract appendix attribution device 100, a preprocessing module 10, a preliminary attribution module 20, a multi-level attribution module 30, an attribution determination module 40, a first language model module 101, a second language model module 102, a third language model module 103, a fourth language model module 104, a contract appendix attribution system 200, a processor 210, a memory 220, a computer-readable storage medium 300, a computer program 310, and a processor 320. Detailed Implementation
[0050] The embodiments of this application will be further described below with reference to the accompanying drawings. The same or similar reference numerals in the drawings denote the same or similar elements or elements having the same or similar functions throughout. Furthermore, the embodiments of this application described below with reference to the accompanying drawings are exemplary and are only used to explain the embodiments of this application, and should not be construed as limiting this application.
[0051] Research revealed the following business pain points in attribution analysis of commercial postscript texts:
[0052] (1) Difficulty in extracting information: Investment contracts are characterized by large volume, complex structure and diverse clauses, which makes it inefficient and easy to miss when manually searching for and understanding the clauses and underlying reasons related to discounts.
[0053] (2) Lack of consistency in attribution: Different people may have different understandings and attribution logics of “discount”, resulting in subjective attribution results that are difficult to standardize.
[0054] (3) The underlying reasons are difficult to uncover: Some reasons for discounts are not directly stated on the surface and may require a comprehensive judgment based on multiple terms, business background and even market conditions.
[0055] (4) Difficulty in knowledge accumulation and reuse: Empirical attribution knowledge is difficult to systematically accumulate and transmit.
[0056] In related technologies, a scheme using text clustering and Large Language Model (LLM) attribution summarization is employed. This involves first using text clustering algorithms to cluster the contract appendix content, and then generating attribution summaries using LLM. This scheme has the following drawbacks:
[0057] First, the granularity of text clustering is difficult to determine: the number of text clusters and the coarseness of the clustering granularity are difficult to control, thus affecting the clustering effect. Understandably, an appropriate clustering granularity is the key to accurate clustering. If the granularity is too large, it may cause content with different topics to be grouped into one category, while if the granularity is too small, it may cause similar content to be scattered, thus affecting the subsequent analysis of contract appendices.
[0058] Second, there are uncertainties and instabilities in the depth and level of attribution and the evidence for attribution: When a general LLM is applied directly to a complex contract note discount attribution task without targeted optimization and structured guidance, there may be uncertainties and instabilities in the output format, depth and level of attribution, and the precise correspondence of evidence.
[0059] In view of this, the embodiments of this application provide a contract appendix attribution method, a contract appendix attribution device, a contract appendix attribution system, and a computer-readable storage medium, aiming to solve the technical problem that when processing a large number of commercial investment contract appendices (hereinafter referred to as "contract appendices"), it is difficult to efficiently and accurately automatically identify discounted contract signing situations, extract corresponding evidence, and conduct standardized, hierarchical, and in-depth attribution analysis.
[0060] Please see Figure 1 and Figure 2 The contract appendix attribution method of the embodiments of this application includes:
[0061] 010: Preprocess the original contract appendix text to obtain the current contract appendix text;
[0062] 020: By using a predefined language model, the current contract appendix text is analyzed for discounted contract recognition and discounted contract evidence extraction. This yields discounted contract evidence information and generates preliminary reasons for discounted contract attribution.
[0063] 030: Determine a multi-level discount attribution catalog based on the current contract appendix text, preliminary discount attribution reasons, and discount signing evidence information using a pre-defined language model;
[0064] 040: Determine the discount attribution result of the original contract appendix text based on the original contract appendix text, discounted contract evidence information, and multi-level discount attribution directory.
[0065] The contract appendix attribution method described in this application constructs a multi-stage, sequential, and collaborative language model processing flow. It progressively completes the process from identifying discounted contracts and extracting evidence to determining a multi-level discount attribution directory. This enables efficient and automatic identification of discounted contract situations in contract appendices and accurate extraction of relevant evidence information, as well as automated, standardized, and multi-level attribution analysis of the reasons for discounted contracts. This application can improve the efficiency and quality of large-scale contract appendix attribution analysis, providing enterprises with more accurate and in-depth business insights.
[0066] Specifically, the predefined language model can be a Large Language Model (LLM). An LLM is a deep learning-based artificial intelligence model that learns from massive amounts of text data to understand and generate human language. Its core function is to capture the complex structure, semantics, and contextual relationships of language, thus demonstrating powerful capabilities in various natural language processing tasks. LLMs possess strong language understanding and generation capabilities, efficient knowledge transfer and generalization characteristics, and are accurate and efficient in handling tasks, supporting complex reasoning and creative output.
[0067] The contract appendix attribution method of this application first preprocesses the original contract appendix text to obtain the current contract appendix text. This process can be implemented by a contract appendix text preprocessing module. Since the original contract appendix text is often complex and contains a large amount of content unrelated to the discount contract attribution analysis, this irrelevant content not only increases the complexity of subsequent analysis but may also interfere with the model's judgment of core information. Therefore, it is necessary to preprocess the original contract appendix text to filter out irrelevant content from the investment contract appendix content to obtain the current contract appendix text.
[0068] Then, a predefined language model is used to identify discounted contracts and extract evidence of discounted transactions from the current contract's appendix text, obtaining evidence of discounted transactions and generating preliminary reasons for attributing the discounted transactions. Specifically, the predefined language model can determine whether the current contract's appendix text constitutes a discounted contract by identifying explicit or implicit discounted factors. Explicit discounted factors may include clear rent reduction clauses, rent-free period extensions, etc.; implicit discounted factors may be hidden in complex clause wording, requiring the model to deeply understand the semantics for identification. For discounted contracts, the predefined language model extracts evidence of discounted transactions and generates preliminary reasons for attributing the discounted transactions, providing direction for subsequent in-depth attribution analysis.
[0069] Next, a multi-level discount attribution catalog is determined based on the current contract appendix text, preliminary discount attribution reasons, and discount contract evidence information using a pre-defined language model. The multi-level category includes two or more levels. For example, the multi-level category can be two, three, four, or five levels, etc., without limitation. Preferably, the multi-level category is three levels, and the pre-defined language model performs a progressively detailed attribution analysis based on the current contract appendix text, preliminary discount attribution reasons, and discount contract evidence information to determine the three-level discount attribution catalog.
[0070] Finally, the attribution result of the original contract's appendix text is determined based on the original contract's appendix text, evidence of discounted signing, and the multi-level attribution directory. This process can be implemented by the attribution combination and evidence integration module. The attribution result of the original contract's appendix text can be comprehensively determined based on the original contract's appendix text, evidence of discounted signing, and the multi-level attribution directory.
[0071] Please see Figure 2 In some implementations, the original contract appendix text is preprocessed to obtain the current contract appendix text (i.e., 010), including:
[0072] Data cleaning is performed on the original contract appendix text to filter out noisy content, resulting in the current contract appendix text. Noisy content includes any one or more of the following: leased resource space, leased area, renovation period project, project conditions, and deduction points.
[0073] Specifically, the contract appendix text preprocessing module can employ a large language model (such as Deepseek-r1). In its implementation, prompts can be pre-designed to guide the large language model in cleaning the input raw contract appendix text, filtering out noisy content and irrelevant information such as leased resource locations, leased area, renovation period engineering, engineering conditions, and deductions. This enhances data quality, improves information focus, and increases the efficiency of subsequent analysis.
[0074] It's understandable that while the aforementioned details regarding leased space, leased area, renovation work, construction conditions, and commission rates are important for the basic terms and execution details of the contract, they are not strongly related to the core attribution logic of discounted contracts. For example, leased space mainly concerns the specific location and layout of the leased premises; leased area concerns the size of the leased space; renovation work and construction conditions focus on the renovation and facility requirements of the leased premises; and commission rates are usually related to the rent calculation method. When analyzing why a contract is discounted, this information often doesn't directly reveal the reason for the discount but instead increases the complexity and interference of subsequent analysis. Therefore, filtering them out as noise will make subsequent analysis more accurate and efficient.
[0075] Please see Figure 3 and Figure 4 In some implementations, the predetermined language model includes a first language model module 101. The predetermined language model is used to identify discounted contracts and extract discounted evidence from the current contract appendix text, obtaining discounted contract evidence information and generating preliminary discounted attribution reasons (i.e., O20), including:
[0076] 021: Determine whether the current contract appendix text is a discounted contract through the first language model module 101;
[0077] 022: If the current contract appendix is a discounted contract, extract the discounted contract evidence information from the current contract appendix.
[0078] 023: Generate preliminary reasons for the discount attribution based on the evidence information of the discounted contract.
[0079] Specifically, the first language model module 101 can be a first LLM module. The first language model module 101 is used to implement discounted contract recognition, discounted evidence extraction, and preliminary discounted attribution reason generation. In some examples, the workflow of the first language model module 101 is as follows:
[0080] (1) Input: The preprocessed text of the current contract remarks;
[0081] (2) Handling calls to large language models (such as deepseek-r1). Guided by carefully crafted prompts, deepseek-r1 performs the following tasks:
[0082] Task 1 (Discounted Contract Recognition): Based on the content of the current contract's appendix text, determine whether the current contract's appendix text contains any discounted contract terms (such as rent reduction, rent-free period extension, etc.). The output can be a Boolean value (Yes / No).
[0083] Task 2 (Evidence Extraction of Discounted Contract): If the current contract appendix text is determined to be a discounted contract, then extract specific text fragments from the current contract appendix text that directly support this determination as evidence of discounted contract.
[0084] Task 3 (Generating Preliminary Reasons for Price Discount): Based on the evidence of the price discount contract, summarize and generate the core reasons that led to the price discount contract, i.e., the preliminary reasons for price discount.
[0085] Example Prompt: "Please analyze the following remarks in the investment agreement: '{Remarks Text}'. 1. Determine whether this remarks indicate a discounted contract (e.g., rent reduction, extended rent-free period, additional subsidies, relaxation of specific conditions, etc., or long vacancy period of the shop, brand entering the area for the first time, brand entering the project for the first time, unified investment promotion by the group, cross-project joint investment promotion, strong brand momentum, market downturn, irregular shop location, impact of the epidemic, etc.). Answer 'Yes' or 'No'. 2. If the answer is 'Yes', please extract the key sentences or phrases in the remarks that directly prove the discounted contract as evidence and list them. Also, please briefly summarize the core reason for this discounted contract, describing it in one sentence or several keywords. Output requirements: Return in strict JSON format:"
[0086] Discounted contract, output format: {{"is_discount":"","discount_evidence":"xx","discount_reason":"xx"}}
[0087] For non-discounted contracts, the output format is {{"is_discount":"No"}}
[0088] (3) Output: Whether the contract was signed at a discount, evidence of the discounted contract, and preliminary reasons for the discount.
[0089] Please see Figure 4 and Figure 5 In some implementations, the predetermined language model includes a second language model module 102, a third language model module 103, and a fourth language model module 104. The predetermined language model determines a multi-level discount attribution catalog (i.e., 030) based on the current contract appendix text, preliminary discount attribution reasons, and discount signing evidence information, including:
[0090] 031: Predefined primary discount attribution list, which includes market factors, store location factors, brand factors, and negotiation factors;
[0091] 032: Using the second language model module 102, based on the current contract appendix text, preliminary reasons for price reduction attribution, and evidence of price reduction signing, the current contract appendix text is classified into the corresponding first-level price reduction attribution directory in the predefined first-level price reduction attribution category, and the corresponding first-level price reduction attribution directory is used as the current first-level price reduction attribution directory;
[0092] 033: Through the third language model module 103, based on the current contract appendix text, the current first-level discount attribution directory, the preliminary discount attribution reasons, and discount contract evidence information, the current contract appendix text is generated or mapped to the current second-level discount attribution directory under the current first-level discount attribution directory;
[0093] 034: Using the fourth language model module 104, based on the current contract appendix text, the current secondary discount attribution directory, the preliminary discount attribution reasons, and discount contract evidence information, the current contract appendix text is generated or mapped to the current tertiary discount attribution directory under the current secondary discount attribution directory, and the corresponding current tertiary discount attribution description is determined.
[0094] Specifically, the second language model module 102 can be a second LLM module. The second language model module 102 is used to generate and map the first-level discount attribution directory. First, a first-level discount attribution directory is predefined and generated. Then, a large language model (such as deepseek-r1) is used to combine the current contract appendix text, the preliminary discount attribution reasons, and the discount signing evidence information to classify the current contract appendix text into the predefined first-level discount attribution directory.
[0095] In some examples, the workflow of the second language model module 102 is as follows:
[0096] (1) Obtain the four categories of attribution for discounted contracts summarized and refined by human experts based on their experience: market factors, store location factors, brand factors, and negotiation factors, as a predefined first-level attribution directory for discounted contracts.
[0097] (2) Input: Current contract appendix text, preliminary reasons for price reduction attribution, attribution evidence information (i.e., evidence information of price reduction contract signing, the same below), and four categories of primary price reduction attribution factors extracted by experts.
[0098] (3) Processing logic: With the help of the reasoning and analysis capabilities of large language models (such as deepseek-r1), the current contract appendix text is assigned to the most appropriate first-level discount attribution factor and used as the current first-level discount attribution directory through prompt optimization.
[0099] Prompt Example: The "known" preliminary attribution reason is: '{Preliminary Attribution Reason}', and the relevant comments and evidence are: '{Comment Text}', '{Evidence Information}', '{Predefined First-Level Attribution Directory}', '{Dynamic few--short Example}'. Please match this text content to the predefined first-level attribution directory, and return the first-level attribution directory content in JSON format as the first-level attribution directory of the current text comment. Output Format:
[0100] {{"is_discount":"Yes","discount_evidence":"xx","discount_reason":"xx","reason_dir":[{{"rea son_dir1"}}]}}
[0101] (4) Output: The current first-level discount attribution directory corresponding to the current contract postscript text.
[0102] The third language model module 103 can be the third LLM module. The third language model module 103 is used to implement the generation and mapping of the second-level discount attribution directory. In some examples, the workflow of the third language model module 103 is as follows:
[0103] (1) Input: The current contract postscript text, the current first-level discount attribution directory, the preliminary discount attribution reason, and the attribution evidence information.
[0104] (2) Processing logic: Call a large language model (such as deepseek-r1). Based on the current first-level discount attribution directory, further refine it through prompts (Prompts) to generate or map to the current second-level discount attribution directory (if the current second-level discount attribution directory does not exist under the current first-level discount attribution directory, then generate the current second-level discount attribution directory here; if the current second-level discount attribution directory already exists under the current first-level discount attribution directory, then map to the current second-level discount attribution directory here). Among them, the second-level discount attribution directory is a specific expansion of the first-level discount attribution directory. The total number of second-level discount attribution directories under each first-level discount attribution directory does not exceed 15, and the name of each second-level discount attribution directory does not exceed 20 characters.
[0105] Prompt example: "Given that the first-level attribution is: '{reason_dir1}', the preliminary attribution reason is: '{preliminary attribution reason}', and '{dynamic few--short example}'. Please generate a second-level attribution directory under this first-level attribution, requiring that the total number of second-level attribution directories ≤ 15, closely related to the text, and not duplicate the existing second-level attribution directories. Output in strict json format, output format:
[0106] {{"is_discount":"yes","discount_evidence":"xx","discount_reason":"xx", "reason_dir":[{{"reason_dir1":[{{"reason_dir2"}}]}}]}}"
[0107] (3) Output: The current second-level discount attribution directory corresponding to the current contract postscript text.
[0108] The fourth language model module 104 can be the fourth LLM module. The fourth language model module 104 is used to implement the generation and mapping of the third-level discount attribution directory. In some examples, the workflow of the fourth language model module 104 is as follows:
[0109] (1) Input: Current contract postscript text, current secondary discount attribution directory, preliminary discount attribution reason, attribution evidence information.
[0110] (2) Processing logic: Call a large language model (such as deepseek - r1). Based on the current secondary discount attribution directory, generate or map to the most granular current tertiary discount attribution directory through prompts (if the current tertiary discount attribution directory does not exist under the current secondary discount attribution directory, it is to generate the current tertiary discount attribution directory here; if the current tertiary discount attribution directory already exists under the current secondary discount attribution directory, it is to map to the current tertiary discount attribution directory) and the corresponding specific tertiary discount attribution description. Among them, the tertiary discount attribution directory is a specific expansion of the secondary discount attribution directory. The total number of tertiary discount attribution directories under each secondary discount attribution directory does not exceed 30, and the name of each tertiary discount attribution directory does not exceed 20 characters. In addition, the current tertiary discount attribution description should be as specific as possible and can directly form a strong association with the evidence information in the current contract postscript text.
[0111] Prompt example: "Given that the primary attribution is: '{reason_dir1}', the secondary attribution is: '{reason_dir2}', and the preliminary attribution reason is: '{preliminary attribution reason}'. Please give the most specific tertiary attribution directory and the corresponding specific description (not exceeding 20 characters) of the tertiary attribution directory according to the postscript content '{postscript text}' and the evidence '{evidence information}', ensuring that this description can be directly supported by the evidence. Require that the total number of tertiary directories under each secondary directory does not exceed 30, and return in strict json format. Output example: {{"is_discount": "yes", "reason_dir": [{{"market factor": {{"macroeconomic downturn": {{"women's clothing industry slump": [{"affected by the aftermath of the epidemic and the overall market environment, the Wheatfield Rose in Times天街 is the only remaining store within the ** system"]}}}}}}, {{"store location factor": {{"traffic": {{"end of the corridor": [{"the store is at the end of the corridor, with poor visibility and difficult access for customers", "the store's sales are mainly based on repeat customers of old customers, with a serious decline in sales and huge operating pressure"]}}}}}}, {{"negotiation factor': {{"short - term transition arrangement": {{"adjust new location for seamless connection": [{"this short - term renewal is for 2 months, to cooperate with the seamless operation of the Wheatfield Rose's adjustment to the new location and at the same time争取时间 for negotiating the next new brand"]}}}}}}, {{"brand factor": {{"operating pressure": {{"high rent - to - sales ratio": [{"the current monthly average performance of the ** store is only about 13, and the rent - to - sales ratio is **%. For this renewal, multiple rounds of communication have been carried out with the other party, and finally this renewal condition has been reached"]}}}}}}]}}
[0112] (3) If it is not a discounted contract, only return no.
[0113] (4) Output: The current three-level discount attribution directory corresponding to the current contract remarks text.
[0114] This application innovatively designs a processing pipeline composed of multiple dedicated LLM modules connected in series. Each LLM module focuses on a specific subtask (discount identification and evidence extraction, preliminary attribution, primary attribution, secondary attribution, and tertiary attribution), guided by carefully designed prompt engineering, progressing step by step to collaboratively complete the complex analysis task from the original contract appendix text to a structured three-level discount attribution catalog. This architecture decomposes complex problems, ensuring the focus and accuracy of each step of the analysis, which is not achieved by directly applying or simply combining general-purpose LLMs.
[0115] Furthermore, this application emphasizes that at every level of attribution (especially at the most granular level of tertiary discount attribution), the results must be directly supported and anchored by explicit textual evidence extracted from contract appendices. It not only provides attribution conclusions but, more importantly, offers textual evidence of "why this attribution is made," significantly enhancing the interpretability, traceability, and credibility of the attribution results.
[0116] In some implementations, the multi-level discount attribution result (i.e., 040) of the original contract appendix text is determined based on the original contract appendix text, discounted contract evidence information, and a multi-level discount attribution directory, including:
[0117] The original contract appendix text, discounted contract evidence information, current first-level discounted attribution directory, current second-level discounted attribution directory, current third-level discounted attribution directory, and corresponding current third-level discounted attribution descriptions are structurally integrated to form the third-level discounted attribution results of the original contract appendix text.
[0118] Specifically, the attribution integration and evidence integration module can integrate the outputs of the first language model module 101, the second language model module 102, the third language model module 103, and the fourth language model module 104 to form a complete analysis record of each contract appendix text, including its specific path in the three-level discount attribution catalog and the direct evidence supporting that path.
[0119] In some examples, the workflow of the attribution integration and evidence integration module is as follows:
[0120] (1) Input: Original contract remarks text, attribution evidence information, current first-level discount attribution directory, current second-level discount attribution directory, current third-level discount attribution directory and current third-level discount attribution description.
[0121] (2) Processing logic: All the above information is structured and integrated to form a complete analysis record of each contract appendix text, including its specific path in the three-level discount attribution directory and direct evidence supporting that path.
[0122] (3) Output: Complete three-level discount attribution results containing clear evidentiary information corresponding to a single contract appendix text.
[0123] Please see Figure 2 and Figure 6 In some implementations, the contract appendix attribution method further includes:
[0124] 050: Obtain reasonable attribution examples from manually annotated data;
[0125] 060: Dynamically add reasonable attribution examples to the cue words of the predefined language model, whereby the cue words are used to guide the predefined language model to determine a multi-level discount attribution catalog.
[0126] Specifically, 010 to 040 above describe the attribution catalog processing for a single contract appendix text. However, in this application's implementation, the attribution catalog processing for multiple contract appendix texts can combine annotations from front-end human experts to obtain reasonable few-shot examples, which are dynamically added to the corresponding Prompt reserved variables. For example, the variable "dynamic few-short examples" reserved in the Prompts of the aforementioned second language model module 102 and third language model module 103.
[0127] Specifically, the Grado and Flask frameworks can be used to build front-end and back-end services that dynamically acquire contract attribution examples annotated by human experts and conform to business logic. This enables real-time front-end display of contract attribution tasks and the collection and immediate feedback of appropriate annotation tasks to back-end tasks. Annotated examples are dynamically appended to the reserved variables in the attribution task's Prompt, allowing the model to better understand how to generate a reasonable three-level discount attribution catalog that meets business requirements.
[0128] It is understandable that in the process of processing the attribution catalog of multiple contract appendices, as the number of reasonable attribution examples manually labeled increases, the three-level discount attribution catalog generated by the model will become more accurate. Subsequently, the model can learn and generate more attribution examples autonomously without having to rely on manual labeling.
[0129] Please see Figure 2 and Figure 7 In some implementations, the contract appendix attribution method further includes:
[0130] 070: Traverse multiple original contract appendix texts to determine the multi-level discount attribution directory corresponding to each original contract appendix text;
[0131] 080: Based on the multi-level discount attribution directory corresponding to each original contract appendix text, update the global discount attribution directory corresponding to multiple original contract appendix texts according to the hierarchical nesting rules.
[0132] Specifically, in the embodiments of this application, the attribution catalog processing flow for multiple contract appendix texts involves traversing multiple original contract appendix texts to generate a three-level attribution catalog and supporting evidence for the overall discounted contract.
[0133] In some examples, the backend logic for processing all the original contract postscript text can be constructed as follows:
[0134] (1) Traverse multiple original contract appendix texts.
[0135] (2) Add reasonable attribution examples marked by front-end human experts to the processing flow of single contract remarks text, generate the current three-level discount attribution directory, and update the current three-level discount attribution directory to the global discount attribution directory.
[0136] (3) Merging logic of the global discount attribution directory: The hierarchical management of the global discount attribution directory is implemented using a hierarchical nested dictionary structure (e.g., a tree-like nested dictionary structure). For example, based on the three-level discount attribution directory corresponding to each original contract appendix text, the global discount attribution directory is updated according to the hierarchical nesting rules and saved as a JSON file.
[0137] (4) Obtain the complete three-level discount attribution catalog corresponding to all original contract appendix texts, i.e., the global discount attribution catalog, to support business experience-based decision-making.
[0138] In summary, the contract appendix attribution method of the present application has at least the following innovative features:
[0139] First, a multi-stage, sequential LLM collaborative processing method for contract appendices: This application innovatively designs a processing pipeline composed of multiple dedicated LLM modules connected in series. Each LLM module focuses on a specific subtask (discount identification and evidence extraction, preliminary attribution, primary attribution, secondary attribution, and tertiary attribution), and through carefully designed prompts and engineering guidance, it delves deeper step by step, collaboratively completing the complex analysis task from the original contract appendice text to a structured three-level discount attribution catalog. This architecture decomposes complex problems, ensuring the focus and accuracy of each step of the analysis, which is not available in the direct application or simple combination of current general-purpose LLMs.
[0140] Second, the evidence-driven hierarchical attribution generation mechanism: This application emphasizes that at each level of attribution (especially at the most granular third-level discount attribution), the results must be directly supported and anchored by explicit textual evidence extracted from contract appendices. It not only provides attribution conclusions but, more importantly, offers textual evidence of "why this attribution is made," significantly enhancing the interpretability, traceability, and credibility of the attribution results.
[0141] Third, dynamically add reasonable few-shot examples: collect reasonable attribution examples annotated by front-end experts in real time to improve the accuracy of LLM attribution catalog generation.
[0142] Please see Figure 4 and Figure 8 The contract appendix attribution device 100 of this application includes a preprocessing module 10, a preliminary attribution module 20, a multi-level attribution module 30, and an attribution determination module 40. The preprocessing module 10 preprocesses the original contract appendix text to obtain the current contract appendix text. The preliminary attribution module 20 uses a predetermined language model to identify discounted contracts and extract discounted evidence from the current contract appendix text, obtaining discounted contract evidence information and generating preliminary discounted attribution reasons. The multi-level attribution module 30 uses a predetermined language model to determine a multi-level discounted attribution directory based on the current contract appendix text, the preliminary discounted attribution reasons, and the discounted contract evidence information. The attribution determination module 40 determines the discounted attribution result of the original contract appendix text based on the original contract appendix text, the discounted contract evidence information, and the multi-level discounted attribution directory.
[0143] In some implementations, the preprocessing module 10 is specifically used to perform data cleaning on the original contract appendix text to filter out noise content in the original contract appendix text and obtain the current contract appendix text. The noise content includes any one or more of the following: leased resource space, leased area, decoration period project, project conditions, and deduction points.
[0144] In some implementations, the predetermined language model includes a first language model module 101. The preliminary attribution module 20 is specifically used to: determine whether the current contract appendix text is a discounted contract through the first language model module 101; if the current contract appendix text is a discounted contract, extract discounted contract evidence information from the current contract appendix text; and summarize and generate preliminary discounted attribution reasons based on the discounted contract evidence information.
[0145] In some implementations, the predetermined language model includes a second language model module 102, a third language model module 103, and a fourth language model module 104. The multi-level attribution module 30 is specifically used for: predefining a first-level discount attribution directory, which includes market factors, store location factors, brand factors, and negotiation factors; using the second language model module 102, based on the current contract appendix text, preliminary discount attribution reasons, and discount signing evidence information, classifying the current contract appendix text into the corresponding first-level discount attribution directory within the predefined first-level discount attribution directory, and using the corresponding first-level discount attribution directory as the current first-level discount attribution directory; using the third language model module 103, based on the current... The contract appendix text, the current first-level discount attribution directory, the preliminary discount attribution reasons, and the discount signing evidence information are used to generate or map the current contract appendix text to the current second-level discount attribution directory under the current first-level discount attribution directory. Through the fourth language model module 104, based on the current contract appendix text, the current second-level discount attribution directory, the preliminary discount attribution reasons, and the discount signing evidence information, the current contract appendix text is generated or mapped to the current third-level discount attribution directory under the current second-level discount attribution directory, and the corresponding current third-level discount attribution description is determined.
[0146] In some implementations, the attribution determination module 40 is specifically used to: structurally integrate the original contract appendix text, discounted contract evidence information, the current first-level discounted attribution directory, the current second-level discounted attribution directory, the current third-level discounted attribution directory, and the corresponding current third-level discounted attribution description to form the third-level discounted attribution result of the original contract appendix text.
[0147] In some embodiments, the contract attribution device 100 further includes an acquisition module and an addition module. The acquisition module is used to acquire manually annotated reasonable attribution examples. The addition module is used to dynamically add reasonable attribution examples to prompts in a predetermined language model, wherein the prompts are used to guide the predetermined language model to determine a multi-level discount attribution catalog.
[0148] In some embodiments, the contract appendix attribution device 100 further includes a traversal module and an update module. The traversal module is used to traverse multiple original contract appendix texts to determine the multi-level discount attribution directory corresponding to each original contract appendix text. The update module is used to update the global discount attribution directory corresponding to multiple original contract appendix texts according to the hierarchical nesting rules based on the multi-level discount attribution directory corresponding to each original contract appendix text.
[0149] It should be noted that the explanation of the contract appendix attribution method in the foregoing embodiments also applies to the contract appendix attribution device 100 in the embodiments of this application, and will not be elaborated here.
[0150] Please see Figure 9The contract appendix attribution system 200 of this application includes one or more processors 210 and a memory 220, the memory 220 storing a computer program. When the computer program is executed by the processor 210, it implements the contract appendix attribution method of any of the above embodiments.
[0151] For example, when a computer program is executed by processor 210, the following contract appendix attribution method is implemented:
[0152] 010: Preprocess the original contract appendix text to obtain the current contract appendix text;
[0153] 020: By using a predefined language model, the current contract appendix text is analyzed for discounted contract recognition and discounted contract evidence extraction. This yields discounted contract evidence information and generates preliminary reasons for discounted contract attribution.
[0154] 030: Determine a multi-level discount attribution catalog based on the current contract appendix text, preliminary discount attribution reasons, and discount signing evidence information using a pre-defined language model;
[0155] 040: Determine the discount attribution result of the original contract appendix text based on the original contract appendix text, discounted contract evidence information, and multi-level discount attribution directory.
[0156] For example, when a computer program is executed by processor 210, the following contract appendix attribution method is implemented:
[0157] 011: Perform data cleaning on the original contract appendix text to filter out noisy content and obtain the current contract appendix text. The noisy content includes any one or more of the following: leased resource space, leased area, decoration period project, project conditions, and deduction points.
[0158] It should be noted that the explanation of the contract appendix attribution method in the foregoing embodiments also applies to the contract appendix attribution system 200 in the embodiments of this application, and will not be elaborated here.
[0159] Please see Figure 10 The computer-readable storage medium 300 of this application embodiment stores a computer program 310 thereon. When the program is executed by the processor 320, it implements the contract appendix attribution method of any of the above embodiments.
[0160] For example, when the program is executed by processor 320, the following contract appendix attribution method is implemented:
[0161] 010: Preprocess the original contract appendix text to obtain the current contract appendix text;
[0162] 020: By using a predefined language model, the current contract appendix text is analyzed for discounted contract recognition and discounted contract evidence extraction. This yields discounted contract evidence information and generates preliminary reasons for discounted contract attribution.
[0163] 030: Determine a multi-level discount attribution catalog based on the current contract appendix text, preliminary discount attribution reasons, and discount signing evidence information using a pre-defined language model;
[0164] 040: Determine the discount attribution result of the original contract appendix text based on the original contract appendix text, discounted contract evidence information, and multi-level discount attribution directory.
[0165] For example, when the program is executed by processor 320, the following contract appendix attribution method is implemented:
[0166] 011: Perform data cleaning on the original contract appendix text to filter out noisy content and obtain the current contract appendix text. The noisy content includes any one or more of the following: leased resource space, leased area, decoration period project, project conditions, and deduction points.
[0167] It should be noted that the explanation of the contract appendix attribution method in the foregoing embodiments also applies to the computer-readable storage medium 300 of the embodiments of this application, and will not be elaborated here.
[0168] In summary, the contract appendix attribution method, device 100, system 200, and computer-readable storage medium 300 of this application construct a multi-stage, sequential, and collaborative language model processing flow. This flow progressively completes the process from identifying discounted contracts and extracting evidence to determining a multi-level discount attribution catalog. This achieves efficient and automatic identification of discounted contract situations in contract appendices and accurate extraction of relevant evidence information, as well as automated, standardized, and multi-level attribution analysis of the reasons for discounted contracts. This application can improve the efficiency and quality of large-scale contract appendix attribution analysis, providing enterprises with more accurate and in-depth business insights.
[0169] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0170] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.
[0171] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, a computer-readable storage medium can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable storage medium could be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.
[0172] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0173] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments. Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc.
[0174] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application, the scope of which is defined by the claims and their equivalents.
Claims
1. A method for attributing contract remarks, characterized in that, include: The original contract appendix text is preprocessed to obtain the current contract appendix text; The current contract appendix text is analyzed using a predefined language model to identify discounted contracts and extract discounted evidence, thereby obtaining discounted contract evidence information and generating preliminary reasons for discounted attribution. The predetermined language model determines a multi-level discount attribution catalog based on the current contract appendix text, the preliminary discount attribution reasons, and the discount contract evidence information; The discount attribution result of the original contract appendix text is determined based on the original contract appendix text, the discounted contract evidence information, and the multi-level discount attribution directory.
2. The contract appendix attribution method according to claim 1, characterized in that, The preprocessing of the original contract appendix text to obtain the current contract appendix text includes: The original contract appendix text is cleaned to filter out noise content, resulting in the current contract appendix text. The noise content includes any one or more of the following: leased resource space, leased area, renovation period project, project conditions, and deduction points.
3. The contract appendix attribution method according to claim 1, characterized in that, The predetermined language model includes a first language model module. The process of identifying discounted contracts and extracting discounted evidence from the current contract appendix text using the predetermined language model to obtain discounted contract evidence information and generate preliminary reasons for discounted attribution includes: The first language model module determines whether the current contract appendix text is a discounted contract. If the current contract appendix text is a discounted contract, extract the discounted contract evidence information from the current contract appendix text; The preliminary reasons for the discount attribution are generated based on the evidence information regarding the discounted contract.
4. The contract appendix attribution method according to claim 1, characterized in that, The predetermined language model includes a second language model module, a third language model module, and a fourth language model module. The step of determining a multi-level discount attribution directory based on the current contract appendix text, the preliminary discount attribution reasons, and the discount signing evidence information using the predetermined language model includes: A predefined primary discount attribution catalog, which includes market factors, store location factors, brand factors, and negotiation factors; Using the second language model module, based on the current contract appendix text, the preliminary discount attribution reasons, and the discount signing evidence information, the current contract appendix text is categorized into the corresponding first-level discount attribution directory in the predefined first-level discount attribution category, and the corresponding first-level discount attribution directory is used as the current first-level discount attribution directory; Through the third language model module, based on the current contract appendix text, the current first-level discount attribution directory, the preliminary discount attribution reasons, and the discount contract evidence information, the current contract appendix text is generated or mapped to the current second-level discount attribution directory under the current first-level discount attribution directory; The fourth language model module generates or maps the current contract postscript text to the current tertiary discount attribution directory under the current tertiary discount attribution directory based on the current contract postscript text, the current secondary discount attribution directory, the preliminary discount attribution reasons, and the discount contract evidence information, and determines the corresponding current tertiary discount attribution description.
5. The contract appendix attribution method according to claim 4, characterized in that, The step of determining the multi-level discount attribution result of the original contract appendix text based on the original contract appendix text, the discounted contract evidence information, and the multi-level discount attribution directory includes: The original contract appendix text, the discounted contract evidence information, the current first-level discounted price attribution directory, the current second-level discounted price attribution directory, the current third-level discounted price attribution directory, and the corresponding current third-level discounted price attribution description are structurally integrated to form the third-level discounted price attribution result of the original contract appendix text.
6. The contract appendix attribution method according to claim 1, characterized in that, The contract appendix attribution method also includes: Obtain reasonable attribution examples from manually annotated files; The reasonable attribution examples are dynamically added to the prompt words of the predetermined language model, wherein the prompt words are used to guide the predetermined language model to determine the multi-level discount attribution catalog.
7. The contract appendix attribution method according to claim 1, characterized in that, The contract appendix attribution method also includes: Traverse multiple original contract appendix texts to determine the multi-level discount attribution directory corresponding to each original contract appendix text; Based on the multi-level discount attribution directory corresponding to each original contract appendix text, update the global discount attribution directory corresponding to multiple original contract appendix texts according to the hierarchical nesting rules.
8. A contract appendix attribution device, characterized in that, include: The preprocessing module is used to preprocess the original contract appendix text to obtain the current contract appendix text; The preliminary attribution module is used to identify discounted contracts and extract discounted evidence from the current contract appendix text using a predefined language model, obtain discounted contract evidence information, and generate preliminary attribution reasons for the discounted contracts. The multi-level attribution module is used to determine a multi-level discount attribution catalog based on the current contract appendix text, the preliminary discount attribution reasons, and the discount contract evidence information through the predetermined language model; The attribution determination module is used to determine the discount attribution result of the original contract appendix text based on the original contract appendix text, the discounted contract evidence information, and the multi-level discount attribution directory.
9. A contract appendix attribution system, characterized in that, The contract appendix attribution system includes one or more processors and a memory, the memory storing a computer program that, when executed by the processor, implements the contract appendix attribution method according to any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the contract appendix attribution method according to any one of claims 1-7.