Table-to-text generation method based on multi-agent collaboration

Abstract

The invention discloses a table-to-text generation method based on multi-agent collaboration, which is characterized in that model training comprises two stages of supervised fine tuning and stylized preference optimization, enhanced training data is constructed through table replacement invariance, a large language model is efficiently fine-tuned by using PEFT, and the model is optimized; and a SimPO algorithm is adopted to carry out positive and negative sample contrast learning to generate a text conforming to a golden standard style. In error correction, through a multi-agent cooperation mechanism, table data is coded by using a BERT model and stored in a Faiss vector database, an agent is generated to generate a text based on a large language model, the agent is retrieved to construct a minimum related table, the agent is verified to compare unit declarations and classify errors, and the agent is scheduled to trigger an error correction strategy. And errors such as numerical values, knowledge missing and table conflicts are accurately corrected. The whole process is iteratively generated and verified, it is ensured that the generated text is loyalty to facts, consistent in style and free of illusion, and the quality and accuracy of table-to-text generation are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and natural language processing technology, and specifically relates to a table-to-text generation method based on multi-agent collaboration. Background Art

[0002] With the rapid development of natural language processing technology, large language models (LLMs) have demonstrated powerful capabilities in text generation tasks, and have been widely used in scenarios where natural language text is generated from structured data (such as tables). However, existing table-to-text generation methods face two major challenges: first, the style of the generated text is difficult to accurately align with specific requirements (such as classical Chinese style or professional terminology), resulting in the generated results deviating from expectations in tone and wording; second, large language models are prone to "hallucination" problems during the generation process, that is, the generated content may deviate from the facts in the table and contain fictitious or contradictory information. These problems are particularly prominent in scenarios that require high accuracy and a specific style (such as legal documents, academic reports, etc.).

[0003] To address the problem of style alignment, existing techniques often train models using supervised fine-tuning (SFT) or manually annotated data. However, due to limitations in training data diversity and the efficiency of model parameter adjustment, data synthesis relies heavily on powerful closed-source models, making efficient style optimization difficult. In recent years, preference optimization algorithms such as DPO (Direct Preference Optimization) and SimPO (Simple Preference Optimization) have shown promise in model alignment, but research on style preference optimization for table-to-text tasks remains limited. Furthermore, to mitigate the problem of hallucination, some methods proofread the generated text through post-processing or rule verification (e.g., [Jacob Meinick et al. Teaching language models to support answers with verified citations. arXiv preprint, 2203.11147, 2022]). However, these methods often rely on manual intervention or simple rules, struggle to address complex semantic errors, and lack real-time performance. Recently proposed general error detection large language model methods (such as the literature [Tang Liyan et al. Efficient fact-checking of large language models based on benchmark documents. Conference on Empirical Methods in Natural Language Processing, November 2024, pp. 8818-8847]) use large-scale data training and meet the requirements for common sense errors and simple factual conflicts. However, they are mostly limited to the single modality of text, cannot flexibly design error types, and lack a systematic multi-agent collaborative mechanism to comprehensively handle problems such as knowledge gaps and table conflicts.

[0004] Therefore, there is an urgent need for a table-to-text generation method that can simultaneously solve the style alignment and hallucination problems. Through efficient model training and real-time error correction mechanisms, it ensures that the generated text is both faithful to the tabular data and meets specific style requirements. Summary of the Invention

[0005] In response to the technical challenges of grid alignment and hallucination problems in existing table-to-text generation technologies, the present invention provides a table-to-text generation method based on multi-agent collaboration, which aims to ensure the complete transmission of the core information of the table while ensuring the accuracy of facts and the controllability of language style, thereby improving text generation quality and user satisfaction.

[0006] A table-to-text generation method based on multi-agent collaboration includes the following steps: (1) Use the Sentence-BERT (Bidirectional Encoder Representations from Transformers) model to encode the rows and columns of the tabular data, generate representation vectors and store them in the Faiss (Facebook AI Similarity Search Library) vector database, and establish an index mapping from the representation vectors to the original rows and columns; (2) Calling the LLM-based generation agent to generate natural language text based on the semantic information of the table and synchronously perform error detection and correction during the text generation process; (3) During the text generation process, the output text is truncated according to the set termination condition (output period or generated content length reaches the preset threshold), and sentence units that can independently express information are extracted from the truncated sentences and recorded as unit statements; (4) Calling the retrieval agent, which uses the Sentence-BERT model to encode the unit statement into a vector, and retrieves the rows and columns with similarity to the vector higher than the preset threshold in the Faiss vector database to form a minimum concern table; (5) Call the verification agent to compare and analyze the unit statement with the content in the minimum concern table to determine whether the unit statement is supported by the minimum concern table data: if supported, the generation agent will continue to generate subsequent content; if not, potential errors will be identified and classified; (6) Call the scheduling agent and assign the corresponding strategy to correct the error according to the error type; (7) Each time a sentence is generated, the verification and error correction process is performed according to steps (4) to (6), and then the process returns to step (2) to continue generating the next sentence until all the generated text contents pass the verification.

[0007] Furthermore, the training process of generating the agent in step (2) includes two stages: supervised fine-tuning and stylized preference optimization. The specific process is as follows: 2.1 Construct an enhanced training dataset for supervised fine-tuning; 2.2 Using the enhanced training dataset, the large language model is fine-tuned using the PEFT (Parameter-Efficient Fine-Tuning) strategy; 2.3 Construct a style preference training dataset for style preference optimization training; 2.4 Based on the style preference training dataset, the SimPO algorithm is used to optimize the style preference of the large language model to obtain a generative intelligent agent based on LLM.

[0008] Furthermore, the specific implementation method of step 2.1 is: obtaining a certain amount of tabular data, performing conversion operations on the tabular data including row shuffling, column shuffling and transposition, and expanding a set of tabular data into multiple sets of tabular data with different arrangement orders as training samples, thereby obtaining an enhanced training data set to enhance the model's ability to recognize core facts.

[0009] Furthermore, the specific implementation method of step 2.3 is as follows: first, the manually annotated gold standard is used as a training positive sample. Based on the manually annotated gold standard, the large language model is used to delete part of the gold standard content or add table content that is not included in the gold standard to the gold standard as a fine-grained style negative sample; then, the large language model is used to rewrite the gold standard in a more colloquial and popular language style as a language style negative sample, thereby obtaining a style preference training dataset.

[0010] Furthermore, the SimPO algorithm in step 2.4 guides preference training through the Bradley-Terry model, constructs a preference probability function for positive and negative sample pairs, and combines the KL divergence to constrain the stability of the large language model. The large language model is optimized through the corresponding objective function, thereby improving the ability of the large language model to generate factually consistent text while satisfying style preferences.

[0011] Furthermore, in step (5), the verification agent identifies potential errors and classifies the errors, and provides a detailed description. The error types are divided into two categories, namely, non-inclusion errors and unsupported errors. Non-inclusion errors refer to the complete absence of the unit declaration content in the minimum concern table, which is model fictitious information; unsupported errors refer to the inconsistency between the unit declaration content and the data in the minimum concern table, which is the introduction of external error information by the model.

[0012] Furthermore, the specific implementation of step (6) is as follows: For those that do not contain errors, the scheduling agent instructs the generating agent to regenerate the content and requires it to generate it based on the current table data; For unsupported errors, the scheduling agent calls a specific correction agent to make corrections based on the detailed description of the error to ensure that the generated content is consistent with the table.

[0013] Furthermore, the specific execution method of the correction agent is: If the generated text content conflicts with the value in the minimum concern table, the correct value in the table is replaced in the text to keep the context consistent; If the error is due to missing terms or misunderstanding of abbreviations, call the DuckDuckGo search library for external search and supplement the large language model with relevant background knowledge to assist in correction; If there is a data conflict in the minimum concern table, the large language model attempts to handle it based on built-in knowledge; if it cannot make a decision, the conflict information is retained and logged for manual review; For errors that cannot be classified, construct prompts to guide the large language model to generate correct content.

[0014] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor is used to execute the computer program to implement the above-mentioned table-to-text generation method based on multi-agent collaboration.

[0015] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned table-to-text generation method based on multi-agent collaboration.

[0016] Based on the above technical solution, the present invention has the following beneficial technical effects: 1. This invention, through a style-controllable training mechanism and a multi-agent collaborative error correction mechanism, comprehensively improves the style expression capability and factual consistency assurance capabilities of the table-to-text generation system in both the training and reasoning stages. It can be widely applied to structured data-driven natural language generation tasks and has significant industrial application value and promotion prospects.

[0017] 2. This invention achieves the coordinated optimization of the factual fidelity of tabular data and the controllability of language style by introducing mechanisms such as structural perturbation enhancement, efficient parameter fine-tuning, and style preference comparative learning. It has strong practical value and promotion prospects.

[0018] 3. The present invention provides a table-to-text generation error correction method based on multi-agent collaboration, which is used to detect and correct factual errors in real time during the text generation process, significantly enhancing the semantic consistency and data reliability of the generated results.

[0019] 4. This invention solves the hallucination problem of large language models in table-to-text conversion through a multi-agent collaborative mechanism. It has strong semantic consistency control capabilities and is suitable for scenarios with high requirements for data accuracy, such as financial statement analysis and data report generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the overall flow of the table feature encoding step in an embodiment of the present invention.

[0021] Figure 2 This is a flow chart of a two-stage training method for a table-to-text generation model based on style preference optimization in an embodiment of the present invention.

[0022] Figure 3 2. Schematic diagram of the flow of the error correction method for multi-agent collaboration in an embodiment of the present invention.

[0023] Figure 4 Schematic diagram of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] This embodiment provides a table-to-text generation method based on multi-agent collaboration, including the following steps: (1) Use the Sentence-BERT model to encode the rows and columns of the table data, generate representation vectors, and store them in the Faiss vector database. At the same time, establish an index mapping from the vector to the original rows and columns. The specific operation process is as follows: Figure 1 shown.

[0026] Sentence-BERT is a model based on the Transformer architecture. Its core is its ability to efficiently map input text into an embedding vector in a high-dimensional semantic space. This allows semantically similar texts to have a high degree of similarity in this vector space, providing a powerful semantic representation foundation for numerous natural language processing tasks. Sentence-BERT is an improved variant of BERT, employing optimization strategies such as contrastive learning and pooling to further enhance the semantic quality and discriminability of the generated text embedding vectors, making it particularly adept at sentence embedding and semantic similarity calculation tasks.

[0027] Specifically, the Sentence-BERT model converts the text (rows and columns) in the table into a token sequence. After processing through the Transformer encoder, it obtains the contextual representation of each token. Finally, it integrates the representation of each token through the average pooling operation to form the embedding vector of the complete sentence. The calculation formula is as follows:

[0028] in: h i Indicates the i The context representation of a token, n Indicates the number of tokens in the text sequence, s This is the high-dimensional embedding vector of the sentence. The above vector reflects the semantic information of the table content and can be used for subsequent efficient retrieval and verification.

[0029] The Faiss vector database is responsible for storing and retrieving the above-mentioned embedded vectors, and adopts an indexing mechanism based on cosine similarity to support fast matching of high-dimensional vectors. The Faiss vector database is responsible for storing and retrieving the above-mentioned embedded vectors, and adopts an indexing mechanism based on cosine similarity to support fast matching of high-dimensional vectors. In actual implementation, the vector dimension is set according to the output of the selected Sentence-BERT model (such as 384 dimensions or 768 dimensions), and the index type can be selected from IVF (inverted file index) or HNSW (hierarchical neighbor graph) structure, depending on the retrieval efficiency and accuracy requirements. To improve query performance, the original vectors are normalized during the index establishment phase to ensure the consistency of the similarity measurement.

[0030] (2) Calling a generative agent based on a large language model to generate natural language text based on the table semantic information, and synchronously performing error detection and correction during the text generation process.

[0031] The large language model used in this embodiment is Qwen2.5-7B-Instruct, developed by the Alibaba Cloud Tongyi Qianwen team. It is based on a decoder-only architecture and is similar to modern large language models (such as the GPT series). It is suitable for a variety of application scenarios and is applicable to multilingual processing and long text generation tasks.

[0032] The model training process of the generative agent adopts a two-stage training method of table-to-text generation model based on style preference optimization, including supervised fine-tuning (SFT) and stylistic preference optimization (StylePO). The specific process is as follows: Figure 2 As shown: Step S1: Construct enhanced training data for supervised fine-tuning.

[0033] Specifically, based on the permutation invariance of the table, transformation operations are performed on the table, including row shuffling, column shuffling and transposition, to generate training samples with multiple arrangement orders to enhance the model's ability to recognize core facts.

[0034] Step S2: Use the enhanced training data to fine-tune the large language model. In this step, the PEFT efficient parameter fine-tuning method is used to fine-tune the model while changing only a small number of parameters.

[0035] Step S3: Construct style preference training data for style preference optimization training, specifically: S3.1: Use manually annotated gold standards as training positive samples; S3.2: Based on the manually annotated gold standard, use a large language model to delete part of the gold standard content or add table content not included in the gold standard as fine-grained style negative samples; S3.3: Use the large language model to rewrite the gold standard using a more colloquial and popular language style as a negative sample of language style; S3.4: Use the SimPO algorithm to perform stylistic preference optimization to generate text that aligns with the gold standard style preferences.

[0036] Specifically, the SimPO algorithm uses positive and negative sample pairs for preference guidance training, where the positive sample and negative samples The preference probability p is expressed by the Bradley-Terry model:

[0037] in: represents the reward score of the positive sample, Represents the reward score of negative samples.

[0038] The objective function of SimPO is:

[0039] Where: E represents the expected function, Indicates action a In state x The rewards under,are based on preference data; Represents the current strategy, θ are model parameters, Representative reference strategy; D KL Represents KL divergence, which is used to constrain the difference between the new and old strategies; βRepresents the penalty coefficient of KL divergence, which controls the conservatism of the policy update.

[0040] Through comparative learning of positive and negative samples, the large language model is guided to generate text that is both faithful to the facts and has a given style. Then, based on the multi-agent collaborative table-to-text generation error correction method, factual errors are detected and corrected in real time during the text generation process. The specific process is as follows: Figure 3 As shown: (3) Each text generation process is terminated when a period is output or the length of the generated content reaches a preset threshold. The system will truncate the output text after the period or punctuation mark (such as a comma) and extract sentence units that can independently express information, called "unit statements", which refer to text fragments that can fully express a fact or information and serve as the basic unit for subsequent verification and error correction.

[0041] (4) Call the retrieval agent and use the Sentence-BERT model to encode the "unit statement" into a vector. Retrieve the table rows and columns with similarity higher than the preset threshold in the Faiss database and construct the "minimum concern table", which refers to the subset of tables containing the most relevant information about the current unit statement, so as to improve the verification efficiency and accuracy.

[0042] The "minimum concern table" is a subset of tables containing data that is highly relevant to the "unit statement", providing an accurate reference for subsequent verification. The setting of the cosine similarity threshold can be flexibly adjusted according to the application scenario.

[0043] (5) The verification agent is called to compare and analyze the "unit statement" with the content in the "minimum concern table" to determine whether the statement is supported by the table data. If it is supported, the generation agent will continue to generate subsequent content. If it is not supported, the verification agent will identify potential errors and classify them as "not included error" or "not supported error", and provide a detailed description. The system must clearly indicate the corresponding data conflict points in the error description, such as inconsistent values ​​or semantics.

[0044] Specifically, "unincluded error" means that the declared content is completely missing from the table, which is fictitious information of the model (internal illusion); "unsupported error" means that the declared content is inconsistent with the table data, which is external error information introduced by the model (external illusion).

[0045] (6) Call the scheduling agent and assign the corresponding error correction strategy according to the error type. If no error is detected, continue to the next round of generation; if an error is identified, enter the following error correction process: For "no error included", the scheduling agent instructs the generating agent to regenerate this part of the content and requires it to generate it based on the current table data.

[0046] For "unsupported errors", the scheduling agent calls a specific correction agent based on the error description to make corrections to ensure that the generated content is consistent with the table, including: Step S1: Numerical error correction agent: If the generated content conflicts with the table value, the correct value in the table is replaced in the text to keep the context consistent.

[0047] Step S2: Knowledge-missing error correction agent: If the error is caused by missing terms or misunderstanding of abbreviations, call the duckduckgo_search library for external search and supplement relevant background knowledge to assist in correction.

[0048] Step S3: Table conflict error correction agent: If there is a data conflict in the table, the model attempts to handle it based on built-in knowledge; if it cannot be determined, the conflict information is retained and a log is recorded for manual review.

[0049] Step S4: Other error correction agents: For errors that cannot be classified, construct prompts to guide the large language model to generate correct content.

[0050] (7) Each time a sentence is generated, return to step (2) to continue generating, and execute the cyclic generation and verification process until all generated content passes verification.

[0051] This embodiment also provides an electronic device designed to achieve table-to-text generation based on style preference optimization and multi-agent collaborative error correction. The electronic device supports functions such as supervised fine-tuning, stylized preference optimization, and multi-agent collaborative error correction through efficient hardware architecture and optimized computing power, ensuring high accuracy, style consistency, and hallucination-free output of generated text. Figure 4 As shown, the electronic device is equipped with one or more high-performance processors (central computing modules) for performing tasks such as data processing, model training, text generation, and error correction. Processor types include one or a combination of central processing units, graphics processing units (GPUs), neural network processing units (NPUs), or tensor processing units (TPUs). The CPU is responsible for general computing tasks such as data preprocessing, task scheduling, and system management. The GPU supports parallel computing to optimize the training and inference of deep learning models. The NPU or TPU is designed to provide low-power, high-throughput computing capabilities for efficient inference and training of large language models.

[0052] The electronic device also includes a storage module for storing computer programs, model parameters, enhanced training data, Faiss vector databases, and intermediate calculation results. The storage module includes non-volatile memory and volatile memory. Non-volatile memory, such as solid-state drives, hard disk drives, or flash memory, is used to persistently store model weights, training data, and program code; volatile memory, such as dynamic random access memory or static random access memory, is used to cache temporary data, encoding results, and intermediate states during multi-agent collaboration. In addition, the cache is integrated into the processor or deployed independently to accelerate data access to optimize the real-time performance of multi-agent tasks.

[0053] To support efficient input and output of external data, the electronic device is equipped with a communication interface module, including a universal serial bus interface for importing tabular data or exporting generated text, a network interface such as an Ethernet port or a high-speed fiber optic interface to support data interaction with cloud servers, databases or client devices, and a serial communication interface to meet the low-bandwidth communication requirements with embedded devices or IoT devices. The communication interface module further provides wired or wireless data transmission capabilities to ensure seamless interaction between the device and external systems such as cloud training clusters or edge computing nodes. The communication interface module includes a network adapter supporting high-speed Ethernet connection, suitable for distributed model training and inference, and a wireless communication module such as Wi-Fi, 5G module or Bluetooth to meet the low-latency data transmission requirements in mobile devices or IoT scenarios.

[0054] Depending on the application scenario, the electronic device can also integrate auxiliary hardware modules to enhance performance and stability. Efficient cooling systems, such as air-cooled radiators, liquid-cooled modules, or heat pipes, ensure stability during long periods of high-load computing. Power management units optimize energy consumption distribution to support the continuous operation of large language model training and inference. Hardware security modules protect the confidentiality and integrity of model parameters, training data, and generated text, making them particularly suitable for sensitive data processing scenarios. These hardware components work together to enable electronic devices to efficiently and stably implement the full-process tasks of table-to-text generation based on style preference optimization and multi-agent collaborative error correction.

[0055] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It is apparent that those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without requiring creative effort. Therefore, the present invention is not limited to the above embodiments. Any improvements or modifications made by those skilled in the art based on the disclosure of the present invention should fall within the scope of protection of the present invention.

Claims

1. A table-to-text generation method based on multi-agent collaboration, characterized in that: The steps include: (1) Use the Sentence-BERT model to encode the rows and columns of the tabular data, generate representation vectors and store them in the Faiss vector database, and establish an index mapping from the representation vectors to the original rows and columns; (2) Calling the LLM-based generation agent to generate natural language text based on the semantic information of the table and synchronously perform error detection and correction during the text generation process; (3) During the text generation process, the output text is truncated by the set termination condition, and sentence units that can independently express information are extracted from the truncated sentences and recorded as unit statements; (4) Calling the retrieval agent, which uses the Sentence-BERT model to encode the unit statement into a vector, and retrieves the rows and columns with similarity to the vector higher than the preset threshold in the Faiss vector database to form a minimum concern table; (5) Call the verification agent to compare and analyze the unit statement with the content in the minimum concern table to determine whether the unit statement is supported by the minimum concern table data: if supported, the generation agent will continue to generate subsequent content; if not, potential errors will be identified and classified; (6) Call the scheduling agent and assign the corresponding strategy to correct the error according to the error type; (7) Each time a sentence is generated, the verification and error correction process is performed according to steps (4) to (6), and then the process returns to step (2) to continue generating the next sentence until all the generated text contents pass the verification.

2. The table-to-text generation method based on multi-agent collaboration according to claim 1 is characterized by: The training process of generating the agent in step (2) includes two stages: supervised fine-tuning and stylized preference optimization. The specific process is as follows: 2.1 Construct an enhanced training dataset for supervised fine-tuning; 2.2 Using the enhanced training dataset, the PEFT strategy is used to supervise fine-tune the large language model; 2.3 Construct a style preference training dataset for style preference optimization training; 2.4 Based on the style preference training dataset, the SimPO algorithm is used to optimize the style preference of the large language model to obtain a generative intelligent agent based on LLM.

3. The table-to-text generation method based on multi-agent collaboration according to claim 2 is characterized by: The specific implementation method of step 2.1 is: obtain a certain amount of tabular data, perform conversion operations on the tabular data including row shuffling, column shuffling and transposition, expand a set of tabular data into multiple sets of tabular data with different arrangement orders as training samples, thereby obtaining an enhanced training data set.

4. The table-to-text generation method based on multi-agent collaboration according to claim 2 is characterized by: The specific implementation method of step 2.3 is as follows: first, the manually annotated gold standard is used as a training positive sample. Based on the manually annotated gold standard, the large language model is used to delete part of the gold standard content or add table content that is not included in the gold standard to the gold standard as a fine-grained style negative sample; then, the large language model is used to rewrite the gold standard in a more colloquial and popular language style as a language style negative sample, thereby obtaining a style preference training dataset.

5. The table-to-text generation method based on multi-agent collaboration according to claim 2 is characterized by: The SimPO algorithm in step 2.4 guides preference training through the Bradley-Terry model, constructs a preference probability function for positive and negative sample pairs, and combines the KL divergence to constrain the stability of the large language model. The large language model is optimized through the corresponding objective function, thereby improving the ability of the large language model to generate factually consistent text while satisfying style preferences.

6. The table-to-text generation method based on multi-agent collaboration according to claim 1 is characterized by: In step (5), the verification agent identifies potential errors and classifies the errors, and provides a detailed description. The error types are divided into two categories, namely, non-inclusion errors and unsupported errors. Non-inclusion errors refer to the complete absence of the unit declaration content in the minimum concern table, which is model fictitious information; unsupported errors refer to the inconsistency between the unit declaration content and the data in the minimum concern table, which is the introduction of external error information by the model.

7. The table-to-text generation method based on multi-agent collaboration according to claim 6 is characterized by: The specific implementation of step (6) is as follows: For those that do not contain errors, the scheduling agent instructs the generating agent to regenerate the content and requires it to generate it based on the current table data; For unsupported errors, the scheduling agent calls a specific correction agent to make corrections based on the detailed description of the error to ensure that the generated content is consistent with the table.

8. The table-to-text generation method based on multi-agent collaboration according to claim 7 is characterized by: The specific execution method of the correction agent is: If the generated text content conflicts with the value in the minimum concern table, the correct value in the table is replaced in the text to keep the context consistent; If the error is due to missing terms or misunderstanding of abbreviations, call the DuckDuckGo search library for external search and supplement the large language model with relevant background knowledge to assist in correction; If there are data conflicts in the minimum concern table, the large language model attempts to handle them based on built-in knowledge; If a determination cannot be made, the conflict information is retained and a log is recorded for manual review; For errors that cannot be classified, construct prompts to guide the large language model to generate correct content.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: The processor is used to execute the computer program to implement the table-to-text generation method based on multi-agent collaboration as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the table-to-text generation method based on multi-agent collaboration as described in any one of claims 1 to 8.

Images