WO2026001792A1 - Text generation method and apparatus, computer program product, electronic device and medium - Google Patents

Abstract

Provided in the present application are a text generation method and apparatus, a computer program product, an electronic device and a medium. The method comprises: acquiring N first text blocks that are associated with a user request question, satisfy a preset condition and are determined on the basis of retrieval across multiple dimensions; evenly distributing the N first text blocks to M pre-trained models for text rewriting, so as to obtain N target text blocks; after N pieces of text prompt information corresponding to the N target text blocks are determined, processing the N pieces of text prompt information on the basis of the M pre-trained models, so as to obtain a candidate response set, the text prompt information comprising the target text blocks, the user request question and a historical question-answer record; and, on the basis of an evaluation strategy, evaluating candidate responses in the candidate response set, so as to select response information matching the user request question. The present application can improve the retrieval accuracy, provide semantically coherent and natural text content, and provide response information matching user request questions, thus ensuring the quality of intelligent responses.

Description

Text generation methods, apparatus, computer program products, electronic devices and media

Cross-reference to related applications

This application claims priority to Chinese Patent Application No. 202410833500.7, filed on June 26, 2024, entitled “Text Generation Method, Apparatus, Computer Program Product, Electronic Device and Medium”, the entire contents of which are incorporated herein by reference.

Technical Field

This application relates to the field of artificial intelligence technology, and in particular to a text generation method, apparatus, computer program product, electronic device and medium.

Background Technology

In recent years, large-scale foundational models such as ChatGPT (Chat Generative Pre-trained Transformer) (e.g., LLM (Large Language Model)) have demonstrated impressive performance in tasks such as text generation and text-to-image generation. However, large language models have inherent limitations, including a tendency to generate illusions and a lack of interpretability in their responses. To fully leverage the powerful summarizing capabilities of large language models, enhancing their generated results by retrieving relevant external information is a common approach to addressing these issues. This is achieved through RAG (Retrieval-Augmented Generation) techniques, which provide more evidence-based and fact-dependent information.

The following drawbacks exist when using RAG technology to generate results for information-enhanced large language models: 1. Low retrieval precision: Not all document blocks in the retrieval set are relevant to the query content, potentially leading to erroneous or incoherent information. Secondly, low recall: Failure to retrieve all relevant document blocks prevents the large language model from acquiring sufficient background information to synthesize the answer. 2. Redundancy and repetition become problems when multiple retrieved texts contain similar information, potentially resulting in duplicate generated content. Submitting all relevant documents to the large language model at once may exceed the context window limit. Connecting a large number of documents to form lengthy search suggestions is ineffective, introducing noise and hindering the large language model's focus on key information. 3. The retrieved content may have a different writing style or tone than the query content, resulting in a less coherent or natural generated language structure.

It is evident that while introducing external knowledge into the RAG system can enrich and concretize the generated text content, it also presents drawbacks such as low retrieval accuracy, information redundancy, and inconsistent or unnatural generated language structures.

Summary of the Invention

In view of the above problems, embodiments of this application provide a text generation method, apparatus, computer program product, electronic device and medium that overcomes or at least partially solves the above problems.

In a first aspect, embodiments of this application provide a text generation method, including:

Obtain N first text blocks that are related to the user's request and meet preset conditions. The N first text blocks are determined based on retrieval from multiple dimensions.

N first text blocks are evenly distributed to M pre-trained models for text rewriting to obtain N target text blocks, where M and N are both integers greater than or equal to 1;

After determining the N text prompts corresponding to the N target text blocks, the N text prompts are processed based on M pre-trained models to obtain a candidate response set. The text prompts include the target text blocks, the user's request question, and the user's corresponding historical question and answer records.

The candidate responses in the candidate response set are evaluated based on the evaluation strategy, and the response information that matches the user's request is selected.

In some embodiments of this application, obtaining N first text blocks that are associated with a user request and meet preset conditions includes:

Retrieve text information that meets the semantic relevance requirements of the user's request question across multiple dimensions, and obtain a candidate text block set including K second text blocks;

Perform text block deduplication and sorting on the candidate text block set to obtain N first text blocks.

In some embodiments of this application, text information whose relevance to the user's request question meets semantic relevance requirements is retrieved across multiple dimensions to obtain a candidate text block set including K second text blocks, including:

Based on the user's request question and the user's corresponding historical Q&A records, obtain a keyword list and a target request question that is rewritten from the user's request question. The keyword list and the target request question are obtained based on the corresponding model service.

Based on the user's request question, the target request question, and the keyword list, text information is retrieved from multiple dimensions to obtain K second text blocks.

In some embodiments of this application, text information retrieval is performed across multiple dimensions based on the user request question, the target request question, and a keyword list to obtain K second text blocks, including:

Input the user request question, the target request question, and the keyword list into the target encoding model to obtain multiple text vectors;

In the target database, text blocks whose semantic relevance to each text vector meets the semantic relevance requirement are retrieved to determine K second text blocks;

The target database stores vector indexes and their corresponding text blocks.

In some embodiments of this application, the method further includes:

Construct a fine-tuned dataset consisting of multiple text data, including question and response information, and the text data is data from the target domain;

Adjust the initial coding model based on the fine-tuning dataset, and determine the target coding model;

Based on the determined text block size, the input document is divided into blocks, and the input document is generated based on information collected in the target domain.

The multiple text blocks obtained from the block processing are input into the target encoding model to determine the target database.

In some embodiments of this application, the method further includes:

The block adjustment range is determined based on the maximum number of words supported by the initial coding model;

Based on the constructed multiple test datasets, determine the text block size within the block adjustment range.

In some embodiments of this application, the candidate text block set is deduplicated and sorted to obtain N first text blocks, including:

Iterate through the candidate text block set and calculate the similarity between each pair of second text blocks;

Based on the similarity calculation results, target text block groups are determined. Any two second text blocks constitute a text block group, and text block groups in which the similarity between two second text blocks is greater than a preset threshold are target text block groups.

In the target text block group, identify the second text block that has a low relevance to the user's request question, and delete the identified second text block to perform deduplication of the second text block;

After performing second-level text block deduplication on the target text block group, obtain the target text block set;

Feature information is extracted from the second text block in the target text block set, and the second text block is scored based on the extracted feature information to obtain the first score corresponding to the second text block;

Based on the first score corresponding to the second text block, N first text blocks are determined in the target text block set.

In some embodiments of this application, based on the first score corresponding to the second text block, N first text blocks are determined in the target text block set, including:

Based on the first score corresponding to the second text block, the second text blocks in the target text block set are sorted to determine the sorting result. The sorting order of the second text blocks is positively correlated with the relevance of the second text blocks.

Based on the sorting results, the N second text blocks with the highest first score in the target text block set are determined, and these N second text blocks are designated as N first text blocks.

In some embodiments of this application, feature information is extracted from a second text block in a set of target text blocks, and a score is evaluated on the second text block based on the extracted feature information to obtain a first score corresponding to the second text block, including:

Feature information is extracted from the second text block to obtain at least one of the following: the first semantic similarity between the second text block and the user's request question, the second semantic similarity between the second text block and the preceding text of the user's request question, and text quality.

The first score corresponding to the second text block is obtained by evaluating the score based on at least one of the first semantic similarity, the second semantic similarity, and the text quality.

In some embodiments of this application, N first text blocks are evenly distributed to M pre-trained models for text rewriting to obtain N target text blocks, including:

According to the principle of balanced allocation, the N first text blocks are divided into M parts, and the pre-trained model corresponding to each part of the first text block is determined.

Based on the correspondence between the first text block and the pre-trained model, M copies of the first text block are concurrently input into the corresponding pre-trained model for text rewriting, resulting in N rewritten target text blocks.

In some embodiments of this application, a candidate response set is obtained by processing N text prompts based on M pre-trained models, including:

After generating N text prompts based on N target text blocks, user request questions, and the user's corresponding historical question and answer records, the N text prompts are concurrently input into the corresponding pre-trained models based on the correspondence between the first text block and the pre-trained model to obtain a candidate response set.

Each text prompt corresponds to at least one candidate response.

In some embodiments of this application, candidate responses in the candidate response set are evaluated based on an evaluation strategy, and response information that matches the user's request is selected, including:

The candidate responses are evaluated based on at least one of response fidelity, response relevance, and contextual relevance to obtain at least one second score;

Based on at least one second score, determine the target score corresponding to the candidate response;

Based on the target score, select the response information that matches the user's requested question from the candidate responses in the candidate response set.

In some embodiments of this application, when evaluating candidate responses based on at least one of response fidelity, response relevance, and contextual relevance, the candidate responses are processed based on the pre-trained model corresponding to the candidate responses;

The pre-trained model corresponding to the candidate response is determined based on the correspondence between the first text block and the pre-trained model.

Secondly, embodiments of this application provide a text generation apparatus, including:

The acquisition module is used to acquire N first text blocks that are associated with the user's request and meet preset conditions. The N first text blocks are determined based on retrieval from multiple dimensions.

The allocation and acquisition module is used to evenly distribute N first text blocks to M pre-trained models for text rewriting and to acquire N target text blocks, where M and N are both integers greater than or equal to 1;

The processing and acquisition module is used to process the N text prompts corresponding to N target text blocks and obtain a set of candidate responses based on M pre-trained models after determining the N text prompts. The text prompts include the target text blocks, the user's request question, and the user's corresponding historical question and answer records.

The evaluation and selection module is used to evaluate candidate responses in the candidate response set based on the evaluation strategy and select the response information that matches the user's request.

Thirdly, embodiments of this application provide a computer program product, including a computer program/instructions, which, when executed by a processor, implement the text generation method of the first aspect.

Fourthly, embodiments of this application provide an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the text generation method of the first aspect.

Fifthly, embodiments of this application provide a computer non-volatile readable storage medium storing a computer program, which, when executed by a processor, implements the text generation method of the first aspect.

The technical solution provided in this application has at least the following beneficial effects:

This process involves retrieving N first text blocks related to the user's request from multiple dimensions, achieving a relatively comprehensive collection of content relevant to the user's request and improving retrieval accuracy. After determining the N first text blocks, they are evenly distributed to M pre-trained models for text rewriting, resulting in N target text blocks. The output of these target text blocks can be based on the pre-trained models, producing text content that is stylistically compatible, semantically coherent, and natural, thus improving the quality of subsequent text responses. After determining the text prompts, based on the correspondence between the pre-trained models and the first text blocks, the N text prompts are distributed to the M pre-trained models for processing to obtain a candidate response set. An evaluation strategy is used to evaluate each candidate response in the candidate response set. The candidate responses can be output using the appropriate pre-trained models, and the evaluation selects the candidate response that best matches the user's request, providing response information that matches the user's request, improving response accuracy, and ensuring the quality of intelligent responses.

Attached Figure Description

Figure 1 is a schematic diagram of the text generation method provided in an embodiment of this application;

Figure 2 is a schematic diagram illustrating the determination of three search requests according to an embodiment of this application;

Figure 3 is a flowchart of constructing a target database related to law provided in an embodiment of this application;

Figure 4 is a flowchart of determining N first text blocks based on 3 retrieval requests provided in an embodiment of this application;

Figure 5 is a flowchart of rewriting text and determining a set of candidate responses provided in an embodiment of this application;

Figure 6 is a schematic diagram of the overall implementation process of the text generation method provided in the embodiments of this application;

Figure 7 is a schematic diagram of the text generation device provided in an embodiment of this application.

Detailed Implementation

To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been presented in the various embodiments of this application to enable readers to better understand this application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can be implemented. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

Currently, when using large language models for intelligent responses, introducing RAG technology can provide more evidence-based and fact-dependent information. However, it also suffers from drawbacks such as low retrieval accuracy, information redundancy, and incoherent or unnatural generated language structures. To mitigate the problems associated with introducing RAG technology, this application provides a text generation method that offers strong support for intelligent responses in fields requiring question-and-answer interaction, such as law, medicine, and tourism.

As shown in Figure 1, the text generation method provided in this application includes:

Step 101: Obtain N first text blocks that are associated with the user's request and meet preset conditions. The N first text blocks are determined based on retrieval from multiple dimensions.

The text generation method provided in this application is applied to a server. In an intelligent question-answering scenario, a user asks a question on the client side. The server receives the user's request question from the client, performs a multi-dimensional search to determine the text content associated with the user's request question, and processes the retrieved text content to determine N first text blocks. The value of N is greater than or equal to 1, so as to determine at least one first text block associated with the user's request question through multi-dimensional search and processing of the search results. Furthermore, by searching for content related to the user's request question in multiple dimensions, relatively comprehensive collection of content related to the user's request question can be achieved, improving search accuracy.

In this application, the client is an application that supports intelligent question answering. The user request question obtained by the server can be a question related to a specific field, such as a legal, medical, or travel-related question. As an embodiment of this application, in an intelligent question answering scenario, a user asks a legal-related question on the client. The client sends the user request question to the server. The server performs relevant content retrieval based on the received user request question across multiple dimensions to determine one or more first text blocks.

Step 102: Distribute the N first text blocks evenly to the M pre-trained models for text rewriting to obtain N target text blocks.

After obtaining N first text blocks related to the user's request, the server evenly distributes these N first text blocks to M pre-trained models deployed on the server. These pre-trained models then rewrite the first text blocks. By rewriting the first text blocks using the pre-trained models, the server outputs text content that matches the style of the pre-trained models, is semantically coherent, and natural, thus obtaining N target text blocks that have undergone style rewriting.

The value of N is greater than or equal to 1, and correspondingly, the value of M is greater than or equal to 1. When the values of N and M are the same, the first text block can be directly assigned to the pre-trained model in a one-to-one correspondence manner; when the value of N is greater than M, the N first text blocks are relatively evenly distributed among the M pre-trained models; when the value of N is less than M, the N first text blocks are evenly distributed among some of the M pre-trained models. As an embodiment of this application, the server deploys pre-trained models A, B, C, and D on different GPUs (Graphics Processing Units), and evenly distributes the 8 first text blocks among the above 4 pre-trained models.

By inputting the first text block into a pre-trained model for text rewriting, the model can output text content with an adapted style to transform the first text block, providing semantically coherent and natural text content, thereby improving the semantic coherence of the response information generated in response to user requests.

Step 103: After determining the N text prompts corresponding to the N target text blocks, process the N text prompts based on M pre-trained models to obtain a candidate response set. The text prompts include the target text blocks, the user's request question, and the user's corresponding historical question and answer records.

After rewriting the first text block based on the text style of the pre-trained model to obtain N target text blocks, corresponding text prompts are generated for each target text block. These prompts include the target text block itself, the user's request question, and the user's corresponding historical question-and-answer records. Then, based on the correspondence between the target text blocks and the pre-trained models, the N text prompts are assigned to M pre-trained models. These M models then process the N text prompts to determine the candidate response set.

The text prompts can be generated based on a combination of the target text block, the user's request question, and the user's corresponding historical question-and-answer records. N text prompts are provided to the corresponding pre-trained models according to the previous allocation strategy. The pre-trained models then process the text prompts and output candidate responses that match the prompts. A candidate response set is generated based on the candidate responses provided by the M pre-trained models.

Step 104: Evaluate the candidate responses in the candidate response set based on the evaluation strategy, and select the response information that matches the user's request.

After obtaining the candidate response set based on the pre-trained model, an evaluation strategy is used to evaluate each candidate response in the candidate response set. Through the evaluation of the candidate responses, the candidate response that best matches the user's request is selected, and the selected candidate response is used as the response information that matches the user's request.

When evaluating each candidate response in the candidate response set using an evaluation strategy, the candidate responses can be evaluated based on one or more evaluation indicators. For the case of using multiple evaluation indicators, a weight is set for each evaluation indicator, and then the final evaluation result is determined based on the weight and the score corresponding to the evaluation indicator.

The above-described implementation scheme of this application retrieves N first text blocks related to the user's request question based on multi-dimensional retrieval, achieving a relatively comprehensive collection of content related to the user's request question and improving retrieval accuracy. After determining the N first text blocks, the N first text blocks are evenly distributed to M pre-trained models for text rewriting to obtain N target text blocks. The output of text content that is adapted to the style of the pre-trained models, semantically coherent, and natural can be based on the pre-trained models, so as to improve the quality of subsequent text responses. After determining the text prompt information, based on the correspondence between the pre-trained models and the first text blocks, the N text prompt information is distributed to the M pre-trained models for processing to obtain a candidate response set. An evaluation strategy is used to evaluate each candidate response in the candidate response set. The candidate responses can be output using the adapted pre-trained models, and the candidate responses are evaluated to select the candidate responses that best match the user's request question, so as to provide response information that matches the user's request question, thereby improving response accuracy and ensuring the quality of intelligent responses.

The following describes the specific process of retrieving N first text blocks related to a user's request question based on multiple dimensions. As an optional embodiment of this application, retrieving N first text blocks that are associated with the user's request question and meet preset conditions includes:

Retrieve text information that meets the semantic relevance requirements of the user's request question across multiple dimensions, and obtain a candidate text block set including K second text blocks;

Perform text block deduplication and sorting on the candidate text block set to obtain N first text blocks.

After obtaining the user's request question, text information that meets the semantic relevance requirement of the user's request question is retrieved across multiple dimensions. The retrieved text information can be text information whose relevance to the user's request question is greater than a pre-set first threshold, or it can be a preset number of text information items ranked high in relevance to the user's request question, where higher ranking indicates higher relevance. During the retrieval, the search is performed in a pre-generated database storing text information.

After obtaining a candidate text block set containing K second text blocks related to the user's request question through multi-dimensional retrieval, the second text blocks in the candidate text block set are deduplicated to streamline the content of the candidate text block set by filtering out redundant information and retaining only necessary text content. After filtering the second text blocks in the candidate text block set through deduplication, the remaining second text blocks in the candidate text block set are sorted. The sorting determines the second text blocks in the candidate text block set that meet the set conditions, thereby obtaining N first text blocks related to the user's request question.

The above implementation process obtains a set of candidate text blocks through multi-dimensional retrieval, which can acquire relatively comprehensive text information related to the user's request. After determining the set of candidate text blocks through retrieval, the content in the set of candidate text blocks is deduplicated and sorted, which can provide text blocks that are highly relevant to the user's request while simplifying the content and avoiding information redundancy.

In some embodiments of this application, when retrieving text information that meets the semantic relevance requirements to the user's request question across multiple dimensions to obtain a candidate text block set including K second text blocks, the process includes:

Based on the user's request question and the user's corresponding historical Q&A records, obtain a keyword list and a target request question that is rewritten from the user's request question. The keyword list and the target request question are obtained based on the corresponding model service.

Based on the user's request question, the target request question, and the keyword list, text information is retrieved from multiple dimensions to obtain K second text blocks.

In intelligent question-answering scenarios, users engage in multi-turn dialogues with pre-trained models. Sometimes, the user's request may be incomplete. Since pre-trained models demonstrate excellent request rewriting capabilities in conversational dense retrieval, they can rewrite the current user request to provide a complete and compliant response. Rewriting a user request can be understood as expanding upon it, making the initial question more complete and clearer through content enrichment.

The request rewriting capability of the pre-trained model is based on the request rewriting model service supported by the pre-trained model. While supporting the request rewriting model service, the pre-trained model also supports the keyword extraction model service, and the two model services support concurrent calls. Therefore, the user's request question can be rewritten and keyword information can be obtained by calling the corresponding model service.

In other words, when performing searches across multiple dimensions, the system retrieves the user's historical question-and-answer records based on the user's request question. These historical records represent the user's past conversations within the intelligent question-and-answer scenario, using pre-trained models deployed on the client and server. Then, two model services are invoked, taking the user's request question and corresponding historical question-and-answer records as input. The two model services then perform request rewriting and keyword extraction respectively, yielding the rewritten target request question and a list of keywords. To improve efficiency, a multi-process concurrent approach can be used to simultaneously invoke the two model services for request rewriting and keyword extraction.

In some embodiments of this application, the request rewriting model service and the keyword extraction model service can also be merged into one model service. In this case, two functions are implemented through one model service. In some embodiments of this application, the request rewriting capability can be implemented based on the cooperation of at least two model services, and the keyword extraction capability can be implemented based on the cooperation of at least two model services. This will not be elaborated further here.

It should be noted that the model service can be a service supported by any pre-trained model deployed on the server, or it can be a service supported by a specific pre-trained model. The server can obtain the keyword list and rewrite the user request by calling the model service, or it can use other methods to extract keywords and obtain the rewritten request.

The following is an embodiment of this application illustrating the process of rewriting requests and extracting keywords. For example, the input text is composed of the following template, and the corresponding model service rewrites the request: As a legal Q&A assistant, my job is to understand the user's true intent. Please complete the user's current request question based on the above scenario. Here, the above scenario is: {dialogue record}, and the user's current request question is: {current question}. Based on the above content, the user's request question is rewritten to make its meaning clear. Finally, please output a rewritten request question without any further explanation or intermediate processes.

For example, consider the following template for input text, with keyword extraction performed by the corresponding model service: "As a legal question-and-answer assistant, your job is to understand the user's true intent. Please extract keywords based on the context. The context is: {dialogue history}, {current question}. Please output the three most important keywords, without any further explanation or intermediate steps."

After obtaining the keyword list and determining the target request question by rewriting the user request question, to ensure retrieval accuracy, the user request question is retained as retrieval request 1 (represented by Q1), the target request question is retained as retrieval request 2 (represented by Q2), and the keyword list is retained as retrieval request 3 (represented by Q3). Based on Q1, Q2, and Q3, text information is retrieved across multiple dimensions to obtain K second text blocks. This is achieved using a hybrid retrieval strategy to adapt to different query types and information needs, ensuring consistent retrieval of highly relevant and context-rich information.

As an embodiment of this application, the process of determining the three requests to be retrieved—request 1, request 2, and request 3—based on the user's request question and the corresponding historical question-and-answer records is shown in Figure 2. The user's request question and the corresponding historical question-and-answer records are input into model service 1, which outputs a rewritten target request question. The user's request question and the corresponding historical question-and-answer records are input into model service 2, which outputs a keyword list. By using the user's request question as request 1, the target request question as request 2, and the keyword list as request 3, three requests to be retrieved are determined based on the user's request question and the corresponding historical question-and-answer records.

The above implementation process, by rewriting the user request question and extracting keywords based on the dialogue record and the current user request question, can obtain a complete and clear rewritten target request question and extract key information points. Furthermore, by using a multi-process concurrent approach to call two model services to perform request rewriting and keyword extraction operations simultaneously, processing efficiency can be improved. Based on the user request question, the target request question, and the extracted keywords, text information retrieval can be performed in multiple dimensions, and mixed retrieval can adapt to different query types and information needs.

In some embodiments of this application, when retrieving K second text blocks by performing text information retrieval across multiple dimensions based on the user request question, the target request question, and a keyword list, the process includes:

Input the user request question, the target request question, and the keyword list into the target encoding model to obtain multiple text vectors;

In the target database, text blocks that meet the semantic relevance requirements with each text vector are retrieved to determine K second text blocks. The target database stores the vector index and the corresponding text block.

After determining the target request question and keyword list, the user request question, target request question, and keyword list are input into the target encoding model to obtain multiple text vectors. For example, the first text vector E1 is obtained based on the user request question, the second text vector E2 is obtained based on the target request question, and the third text vector E3 is obtained based on the keyword list. Then, in the target database used to store vector indexes and corresponding text blocks, text blocks semantically related to the first text vector E1, text blocks semantically related to the second text vector E2, and text blocks semantically related to the third text vector E3 are retrieved. Based on the retrieved text blocks, a candidate text block set including K second text blocks is determined.

In some embodiments of this application, the number of text vectors obtained based on the target encoding model can also be other, such as obtaining 4 text vectors, with the keyword list corresponding to two text vectors; or obtaining 2 text vectors, with the user request question and the target request question corresponding to the same text vector; other numbers of text vectors can also be obtained, which will not be listed here.

The text blocks retrieved from the target database that are semantically related to the text vector can be text blocks with a correlation degree greater than a threshold, or a specific number of text blocks that rank highly in terms of correlation degree with the text vector. In this case, the higher the ranking of the text block, the higher its correlation degree with the text vector.

The target encoding model is a model determined by adjusting the initial encoding model, and the size of the text blocks in the target database is the optimal block size determined through testing. Determining the text block size includes: determining the block adjustment range based on the maximum number of words supported by the initial encoding model; and determining the text block size within the block adjustment range based on multiple constructed test datasets.

When building a RAG system, chunk size is a key parameter. Generally, it's desirable to retain some overlap between chunks to ensure semantic context isn't lost. In most cases, fixed-size chunking is the optimal approach. Compared to other forms of chunking, fixed-size chunking is computationally inexpensive and easy to use because it doesn't require any NLP (Natural Language Processing) libraries. This application's embodiment adjusts chunks from smallest to largest. After selecting an initial encoding model, the chunk adjustment range is determined based on the maximum number of words supported by the initial encoding model. Then, based on multiple constructed test datasets, chunk sizes are tested within this adjustment range in ascending order to determine the optimal chunk size.

As an embodiment of this application, bge (BAAI General Embedding)-base-zh-v1.5 (智源一般嵌模-基本模型-中文-版1.5) is used as the initial encoding model. This model supports a maximum of 512 words (tokens). Therefore, the block adjustment range is designed as [128, 256, 384, 512], and for example, 100 test sets are constructed. Based on the constructed 100 test sets, the block size is tested in ascending order, and the optimal block size is determined through testing.

This application embodiment can also employ other methods, such as reinforcement learning, to determine the optimal block size. After determining the block adjustment range, multiple codes are determined based on the multiple block sizes corresponding to the block adjustment range, with each code corresponding to a block size. The core of searching for the optimal block size based on reinforcement learning is: using a policy network to obtain the code for the text block size, calculating the accuracy of text retrieval based on the block size, and using the accuracy as the reward for updating the reinforcement learning policy network. That is, when performing text retrieval based on the block size within the block adjustment range, the accuracy is determined, the policy network is updated based on the determined accuracy to continuously optimize the policy network, and the optimal block size is determined when the policy network optimization is complete.

Wherein, given a determined text block size, the text generation method provided in this application embodiment further includes:

Construct a fine-tuned dataset consisting of multiple text data, including question and response information, and the text data is data from the target domain;

Adjust the initial coding model based on the fine-tuning dataset, and determine the target coding model;

Based on the determined text block size, the input document is divided into blocks, and the input document is generated based on information collected in the target domain.

The multiple text blocks obtained from the block processing are input into the target encoding model to determine the target database.

In this embodiment, in order to improve retrieval quality and ensure that the retrieved content is highly relevant to the query content, the initial encoding model is fine-tuned to determine the target encoding model. After determining the target encoding model, a target database is constructed based on the target encoding model and the input documents corresponding to the target domain to store vector indexes and corresponding text blocks. This allows for the storage of evolving or rare terms in the target domain in the target database, thereby ensuring the comprehensiveness of the data stored in the target database and improving retrieval quality.

Before fine-tuning the initial encoding model, a fine-tuning dataset consisting of multiple text data points is constructed. Each text data point represents the target domain and includes both question and response information. The text data can be generated based on a pre-trained model, and its size is not specifically limited. Then, the model parameters of the initial encoding model are adjusted based on the fine-tuning dataset to determine the target encoding model. For example, a fine-tuning dataset of 10,000 text data points is constructed. The bge-base-zh-v1.5 model is fine-tuned based on this dataset to determine the target encoding model, which outputs text vectors based on the input text content.

After determining the target encoding model by fine-tuning the initial encoding model, the input document is divided into blocks according to the determined text block size, resulting in multiple text blocks. These multiple text blocks are then input into the target encoding model to determine the corresponding text vectors. Vector indices are set for the text vectors, and a target database is constructed based on the correspondence between vector indices and text blocks. The input document is generated based on information collected from the target domain. This collected information may include text, audio, and/or video information. Audio and video information needs to be converted into text to form the input document.

The above implementation process involves selecting the optimal block size, constructing a target database based on the optimal block size, and retrieving text vectors corresponding to the user request question, the target request question, and the keyword list from the target database. This can provide text blocks that are adapted to multiple dimensions, thereby generating a candidate text block set including K second text blocks.

As an optional embodiment of this application, the target field is the legal field, and the constructed target database is a law-related target database. Based on the constructed database, answers are provided to law-related questions raised by users. Figure 3 shows the implementation process of constructing the law-related target database. Law-related books, official website links, and detailed regulations are collected to form law-related documents, which serve as the input document. Based on a determined text block size (optimal text block size), the input document is divided into blocks to obtain multiple text blocks. These multiple text blocks are input into the target encoding model to obtain text vectors and set vector indices. The law-related target database is constructed based on the vector indices and the corresponding text blocks. The constructed law-related target database can include various aspects of legal knowledge to ensure the comprehensiveness of the information covered by the target database.

After constructing a target database related to law, appropriate text content can be retrieved from this database for users' legal-related questions, and responses can be provided based on the constructed target database.

The process of determining N first text blocks from a candidate text block set is described below. In some embodiments of this application, after generating a candidate text block set including K text blocks, the candidate text block set is subjected to text block deduplication and sorting processing to obtain N first text blocks, including:

Iterate through the candidate text block set and calculate the similarity between each pair of second text blocks;

Based on the similarity calculation results, target text block groups are determined. Any two second text blocks constitute a text block group, and text block groups in which the similarity between two second text blocks is greater than a preset threshold are target text block groups.

In the target text block group, identify the second text block that has a low relevance to the user's request question, and delete the identified second text block to perform deduplication of the second text block;

After performing second-level text block deduplication on the target text block group, obtain the target text block set;

Feature information is extracted from the second text block in the target text block set, and the score of the second text block is evaluated based on the extracted feature information to obtain the first score corresponding to the second text block;

Based on the first score corresponding to the second text block, N first text blocks are determined in the target text block set.

When deduplicating text blocks in a candidate text block set, the similarity between two second text blocks is calculated based on a strategy of pairwise combination of each second text block with other second text blocks that are distinct from the current second text block. Two combined second text blocks form a text block group, and combinations of second text block A and second text block B, and combinations of second text block B and second text block A, belong to the same text block group. For example, if the candidate text block set includes second text block A, second text block B, and second text block C, then based on pairwise combinations, the following text block groups can be obtained: a text block group including second text block A and second text block B, a text block group including second text block A and second text block C, and a text block group including second text block B and second text block C.

Within the constructed text block group, based on the similarity calculation results, target text block groups are identified where the similarity between two second text blocks exceeds a preset threshold. Then, for each target text block group, second text blocks with low relevance to the user's request are identified and deleted to perform text block deduplication. After deduplication of the target text block group, a target text block set is determined based on the deduplicated target text block group and other text block groups. This target text block set does not include duplicate second text blocks; if a second text block is retained multiple times, the target text block set includes only one instance of that second text block.

After determining the target text block set based on deduplication, feature information is extracted from each second text block in the target text block set. Then, the second text block is evaluated based on the extracted feature information to obtain the first score corresponding to the second text block.

Specifically, when extracting feature information from the second text block and obtaining a first score based on the extracted feature information, at least one of the following is obtained: the first semantic similarity between the second text block and the user's request question, the second semantic similarity between the second text block and the preceding text of the user's request question, and the text quality; a score evaluation is performed based on at least one of the first semantic similarity, the second semantic similarity, and the text quality to obtain the first score corresponding to the second text block.

In this embodiment, the feature information obtained by extracting feature information from the second text block includes at least one of first semantic similarity, second semantic similarity, and text quality. The first similarity measures the relevance of the second text block to the user's request question, and the second similarity measures the relevance of the second text block to historical question-and-answer records. Text quality can be evaluated based on metrics such as PPL (perplexity) and Distinct. Perplexity measures whether a sentence is fluent, and Distinct measures the diversity of responses. The definition of perplexity is as follows:

Where P(x _{_i} | x_₁, x _₂ , ..., x _{_i-1} ) represents the probability of predicting the i-th word based on the preceding words, and N represents the sentence length. The smaller the PPL value, the more natural and fluent the text. Evaluating text using PPL can avoid situations where the text is out of order or has inverted sentences.

The Distinct metric is specifically used to determine whether there is a large amount of generic or repetitive content. The definition of Distinct is as follows:

In the above formula, Count(uniquengram) represents the number of unique ngrams in the text, Count(word) represents the total number of ngram words in the text, and the larger Distinct(n) is, the higher the diversity.

After extracting at least one of the following based on feature information—first semantic similarity, second semantic similarity, and text quality—a score is evaluated based on at least one of these features to obtain a first score for the second text block. If the score evaluation is based on at least two of the aforementioned feature information, parameters for each feature information need to be set to determine the first score based on the parameters and feature information.

As an embodiment of this application, the first semantic similarity Score ₁ , the second semantic similarity Score ₂ , and the text quality Score _PPL and Score _Distinct of the second text block are obtained based on feature information extraction. Then, the corresponding first score is determined based on the scoring formula: Score = α1Score ₁ + β1Score ₂ + γ1(Score _PPL + Score _Distinct ), where α1, β1, and γ1 are hyperparameter coefficients and are set according to the experimental results.

After obtaining the first score corresponding to each second text block in the target text block set, when determining N first text blocks in the target text block set based on the first score corresponding to the second text blocks, the process includes: sorting the second text blocks in the target text block set based on the first score corresponding to the second text blocks, and determining the sorting result, wherein the order of the second text blocks is positively correlated with the relevance of the second text blocks; and based on the sorting result, determining the N second text blocks with the highest first scores in the target text block set, and defining the N second text blocks as N first text blocks.

After obtaining the first score for each second text block, the second text blocks in the target text block set are sorted based on the first score. The second text blocks are sorted from highest to lowest first score to obtain the sorting result. A higher first score indicates a higher relevance between the second text block and the user's request question and/or historical Q&A records, and/or higher text quality. That is, the second text block with higher relevance and higher text quality appears earlier. After sorting, the top N second text blocks with the highest first scores in the target text block set are determined to obtain N first text blocks, thus filtering out text blocks with high relevance and high text quality to the user's request question and/or historical Q&A records.

The above implementation process involves determining a set of candidate text blocks, then deduplicating the set to remove redundant information and simplify the content. After deduplication, feature information is extracted for score evaluation, and text blocks are selected based on the scores. This provides high-quality text blocks and optimizes content related to the query while deduplicating.

As an embodiment of this application, the process of determining the first N text blocks based on three search requests is described below using a flowchart, as shown in Figure 4. The user request question is taken as search request 1, the target request question as search request 2, and the keyword list as search request 3. Search request 1, search request 2, and search request 3 are respectively input into the target encoding model to obtain a first text vector, a second text vector, and a third text vector. Based on the first, second, and third text vectors, a search is performed in the target database to obtain q text blocks associated with the first text vector, q text blocks associated with the second text vector, and q text blocks associated with the third text vector, thus determining K second text blocks. The q text blocks associated with the text vectors are the selected text blocks with high relevance.

Then, text block deduplication is performed on the candidate text block set including K second text blocks. After deduplication, feature information is extracted from the remaining second text blocks, a first score is calculated, and the second text blocks are sorted according to the first score. Finally, N first text blocks are provided.

As shown in Figure 4, the process of searching, deduplication, scoring, and sorting filters out highly relevant text blocks, thus providing high-quality text blocks and ensuring search results.

The following describes an implementation scheme for determining the response information matching the user's request based on the first text block N. This implementation scheme includes the processes of obtaining the target text block, obtaining a candidate response set, and filtering matching candidate responses from the candidate response set.

Specifically, the process of obtaining target text blocks, in which N first text blocks are evenly distributed to M pre-trained models for text rewriting to obtain N target text blocks, includes:

According to the principle of balanced allocation, the N first text blocks are divided into M parts, and the pre-trained model corresponding to each part of the first text block is determined. Based on the correspondence between the first text blocks and the pre-trained model, the M parts of the first text blocks are concurrently input into the corresponding pre-trained model for text rewriting, and N rewritten target text blocks are obtained.

In this embodiment, M pre-trained models are deployed on the server. When the value of M is greater than 1, the M pre-trained models support concurrent invocation. After determining N first text blocks, based on the principle of balanced allocation, the N first text blocks are divided into M parts, and the pre-trained model corresponding to each part of the first text blocks is determined, thus establishing a correspondence between the M parts of the first text blocks and the M pre-trained models.

In one embodiment of this application, if N is 10 and M is 4, the 10 first text blocks are divided into 4 parts, with the number of text blocks in each part being 2, 2, 3, and 3 respectively, to achieve a relatively balanced division of text blocks. Alternatively, if N is 15 and M is 3, the 15 first text blocks are divided into 3 equal parts, with the number of text blocks in each part being 5.

The pre-trained model corresponding to each first text block can be randomly determined or determined according to a preset strategy. For example, if N first text blocks are divided in descending order of relevance, the resulting M first text blocks are arranged in descending order of priority, with relevance being positively correlated with priority; that is, the higher the relevance, the higher the priority. For the M pre-trained models, their corresponding priority order is determined based on parameters such as usage frequency and audience reach. The correspondence between the M first text blocks and the M pre-trained models is established either by assigning higher-priority first text blocks to higher-priority pre-trained models, or by assigning higher-priority first text blocks to lower-priority pre-trained models.

After establishing the correspondence between the first text block and the pre-trained model, M copies of the first text block are concurrently input into the corresponding pre-trained model. The first text blocks are then rewritten based on the text style of the pre-trained model to obtain N target text blocks. This achieves efficient acquisition of the required information through multi-threaded concurrency. Furthermore, by rewriting the text style of the first text block, issues such as incoherent or unnatural responses generated for user requests can be reduced or avoided.

The process of obtaining a candidate response set, when processing N text prompts based on M pre-trained models to obtain the candidate response set, includes:

After generating N text prompts based on N target text blocks, user request questions, and corresponding historical question-and-answer records, the N text prompts are concurrently input into the corresponding pre-trained models based on the correspondence between the first text block and the pre-trained model to obtain a set of candidate responses; each text prompt corresponds to at least one candidate response.

After rewriting the first text block using a pre-trained model to obtain N target text blocks, for each target text block, corresponding text prompts are generated based on the combination of the current target text block, the user's request question, and the user's corresponding historical question-and-answer records, resulting in N text prompts for the N target text blocks. Then, using the correspondence between the first text block and the pre-trained model, the N text prompts are concurrently input into the corresponding pre-trained models. The pre-trained models process the text prompts to obtain candidate responses determined based on the text prompts, and then aggregate these candidate responses to determine a candidate response set.

By concurrently inputting N text prompts into M pre-trained models, the M pre-trained models can process the text prompts concurrently, efficiently outputting candidate responses in a multi-threaded concurrent manner. For any text prompt, after it is input into the corresponding pre-trained model, the pre-trained model outputs at least one candidate response. The candidate responses corresponding to the N text prompts are aggregated to determine a candidate response set including at least N candidate responses.

The following describes an embodiment of this application illustrating the process of rewriting the first text block and determining the candidate response set based on the rewritten target text block. As shown in Figure 5, N first text blocks are relatively evenly distributed among four pre-trained models (pre-trained model A, pre-trained model B, pre-trained model C, and pre-trained model D). The four pre-trained models concurrently process the text rewriting of the first text blocks, outputting N target text blocks, thus efficiently providing the target text blocks in a multi-threaded concurrent manner.

After obtaining N target text blocks, for each target text block, text prompt information is generated based on the target text block, the user's request question, and the user's corresponding historical question and answer records. The N text prompt information are input into the corresponding pre-trained model, and the four pre-trained models process the text prompt information concurrently to output the candidate responses corresponding to the text prompt information, so as to efficiently determine the candidate response set in a concurrent processing manner.

In the process shown in Figure 5, a series of steps, including text rewriting, text prompt generation, and output of candidate responses, determine the set of candidate responses. This provides a basis for filtering responses that match the user's request, and the concurrent processing method improves efficiency.

The process of filtering matching candidate responses from a candidate response set, when evaluating candidate responses based on an evaluation strategy and selecting the response information that matches the user's request, includes:

The candidate responses are evaluated based on at least one of response fidelity, response relevance, and contextual relevance to obtain at least one second score; based on the at least one second score, the target score corresponding to the candidate response is determined; and based on the target score, the response information that matches the user's request question is selected from the candidate responses corresponding to the candidate response set.

When selecting candidate responses from the candidate response set as responses that match the user's request, each candidate response in the candidate response set is evaluated based on at least one of response fidelity, response relevance, and contextual relevance to determine at least one second score corresponding to the candidate response.

When evaluating candidate responses based on response fidelity, a pre-trained model corresponding to each candidate response is used to perform semantic decomposition, resulting in multiple reference responses. The consistency of each reference response with the context is then examined to determine whether each reference response supports the user's requested question. A second score corresponding to the response fidelity is determined based on the ratio of the number of reference responses that support the user's requested question to the total number of reference responses.

When evaluating candidate responses based on their relevance, a pre-trained model corresponding to each candidate response is used to create potential questions that fit the candidate response. The similarity between these questions and the user's requested question is then analyzed. The score corresponding to the response relevance is obtained by calculating the average similarity between all generated questions and the user's requested question. The specific calculation formula is as follows:

For a given candidate response as(q), the pre-trained model is prompted to generate n potential questions q _i based on as(q). For each q _i , the similarity sim(q,q _i ) with the user's request question (original question) q is calculated. The final result is determined based on the mean of the sum of the n similarities and is used as the second score corresponding to the relevance of the response.

When evaluating candidate responses based on contextual relevance, a pre-trained model corresponding to the candidate response is used to filter statements directly related to the user's request from the user's historical question-and-answer records. The second score corresponding to contextual relevance is determined by the proportion of these statements to the total number of statements in the context. The specific calculation formula is as follows:
CR = Number of filtered statements / Total number of context statements

By filtering statements directly related to the user's question from the user's historical question and answer records, a second score is determined based on the ratio of the filtered statements to the total number of statements, thus measuring contextual relevance.

Since the pre-trained model corresponding to the candidate response is used to process the candidate response when evaluating response fidelity, response relevance and context relevance, it is necessary to determine the appropriate pre-trained model in advance based on the correspondence between the candidate response and the pre-trained model.

After determining at least one second score for any candidate response, a target score for the candidate response is determined based on the at least one second score. Finally, based on the target scores for each candidate response, response information for the user's requested question is selected from the candidate response set.

As an embodiment of this application, candidate responses are evaluated based on response fidelity, response relevance, and contextual relevance to obtain a second score Score _F corresponding to response fidelity, a second score _AR corresponding to response relevance, and a second score _CR corresponding to contextual relevance. Then, the target score is determined based on the scoring formula: Score = α²Score _F + β²Score _AR + γ²Score _CR , where α², β², and γ² are hyperparameter coefficients set according to experimental results.

After determining the target score for each candidate response, the candidate responses are sorted in descending order of score. The candidate response with the highest score is selected as the response to the user's request. This allows for the automatic evaluation of multiple candidate responses to select the best response, thereby improving the accuracy of the generated response information.

After determining the response information for the user's request, the server returns the determined response information to the client for display, so that the user can obtain high-quality response information.

The overall implementation process of this application is described below through a specific implementation procedure, as shown in Figure 6:

Before retrieval, the target encoding model and text block size are determined. Based on the text block size, the input document is divided into multiple text blocks. A target database is constructed based on the target encoding model and the multiple text blocks. Based on the received user request question, the corresponding historical question and answer records are obtained. Based on the user request question and historical question and answer records, a keyword list and a target request question rewritten from the user request question are obtained.

Retrieval Phase: Based on the user request question, target request question, and keyword list, a multi-dimensional search is performed to obtain a set of candidate text blocks, including K second text blocks. Specifically, the retrieval process involves inputting the user request question, target request question, and keyword list into the target encoding model to obtain multiple text vectors; then, text blocks whose semantic relevance to each text vector meets the semantic relevance requirement are retrieved from the target database to determine the K second text blocks.

Deduplication and sorting stage: The second text block in the candidate text block set is deduplicated and sorted. Based on the sorting results, N first text blocks are selected. For details, please refer to the relevant descriptions in the above embodiments, which will not be repeated here.

Generation phase: The pre-trained model is used to rewrite the text of N first text blocks to determine the target text block. Based on the target text block, the text prompt information is determined and input into the corresponding pre-trained model to obtain the candidate response corresponding to the text prompt information.

Screening phase: Score candidate responses based on at least one evaluation metric, and select responses that match the user's question based on the scores.

The above implementation process, through a series of operations including target database construction, multi-dimensional retrieval, information deduplication, text block sorting, text block rewriting, candidate response generation, and candidate response filtering, provides the client with response information that matches the user's request question, ensuring the response quality of intelligent question answering based on the pre-trained model.

The above describes the overall implementation process of the text generation method provided in this application. Based on multi-dimensional retrieval, N first text blocks related to the user's request are obtained, achieving a relatively comprehensive collection of content related to the user's request and improving retrieval accuracy. After determining the N first text blocks, they are evenly distributed to M pre-trained models for text rewriting to obtain N target text blocks. Based on the pre-trained models, text content that is style-adapted, semantically coherent, and natural can be output, thus improving the quality of subsequent text responses. After determining the text prompt information, based on the correspondence between the pre-trained models and the first text blocks, the N text prompt information is distributed to the M pre-trained models for processing to obtain a candidate response set. An evaluation strategy is used to evaluate each candidate response in the candidate response set. Adapted pre-trained models can be used to output candidate responses, and the candidate responses are evaluated to select the most suitable candidate response to the user's request, providing response information that matches the user's request, thus improving response accuracy and ensuring the quality of intelligent responses.

After determining the candidate text block set, the content in the candidate text block set is deduplicated and sorted. This can reduce the amount of content and avoid information redundancy while providing text blocks that are highly relevant to the user's request. By adopting a multi-process concurrent processing method, processing efficiency can be improved.

This application embodiment also provides a text generation apparatus, as shown in FIG7, including:

The acquisition module 701 is used to acquire N first text blocks that are associated with the user's request and meet preset conditions. The N first text blocks are determined based on retrieval from multiple dimensions.

The allocation and acquisition module 702 is used to evenly distribute N first text blocks to M pre-trained models for text rewriting and to acquire N target text blocks, where M and N are both integers greater than or equal to 1;

The processing and acquisition module 703 is used to process the N text prompts based on M pre-trained models after determining the N text prompts corresponding to the N target text blocks to obtain a set of candidate responses. The text prompts include the target text blocks, the user's request question, and the user's corresponding historical question and answer records.

The evaluation and selection module 704 is used to evaluate the candidate responses in the candidate response set based on the evaluation strategy and select the response information that matches the user's request question.

In some embodiments of this application, the acquisition module includes:

The retrieval submodule is used to retrieve text information that meets the semantic relevance requirements of the user's request question across multiple dimensions, and to obtain a candidate text block set including K second text blocks;

The processing and acquisition submodule is used to perform text block deduplication and sorting on the candidate text block set to obtain N first text blocks.

In some embodiments of this application, the retrieval submodule includes:

The first acquisition unit is used to acquire a keyword list and a target request question rewritten from the user's request question, based on the user's request question and the user's corresponding historical question and answer records. The keyword list and the target request question are acquired based on the corresponding model service.

The second acquisition unit is used to retrieve text information across multiple dimensions based on the user request question, the target request question, and the keyword list, and to acquire K second text blocks.

In some embodiments of this application, the second acquisition unit includes:

The input acquisition subunit is used to input the user request question, the target request question, and the keyword list into the target encoding model to obtain multiple text vectors;

The retrieval determines the sub-unit, which is used to retrieve text blocks in the target database that meet the semantic relevance requirements of each text vector, so as to determine K second text blocks;

The target database stores vector indexes and their corresponding text blocks.

In some embodiments of this application, the apparatus further includes:

The building module is used to construct a fine-tuning dataset that includes multiple text data, including question and response information, and the text data is data from the target domain;

The adjustment and determination module is used to adjust the initial coding model based on the fine-tuning dataset and determine the target coding model.

The block processing module is used to block the input document according to the determined text block size. The input document is generated based on information collected in the target domain.

The first determining module is used to input the multiple text blocks obtained from the block processing into the target encoding model to determine the target database.

In some embodiments of this application, the apparatus further includes:

The second determining module is used to determine the block adjustment range based on the maximum number of words supported by the initial encoding model;

The third determination module is used to determine the text block size within the block adjustment range based on the multiple test datasets constructed.

In some embodiments of this application, the processing acquisition submodule includes:

The calculation unit is used to iterate through and calculate the similarity between each pair of second text blocks in the candidate text block set;

The determining unit is used to determine the target text block group based on the similarity calculation result. Any two second text blocks constitute a text block group, and the text block group in which the similarity between two second text blocks is greater than a preset threshold is the target text block group.

The deduplication unit is used to identify a second text block in the target text block group that has a low relevance to the user's request question, and delete the identified second text block to perform deduplication of the second text block;

The third acquisition unit is used to acquire the target text block set after performing the second text block deduplication on the target text block group;

The evaluation and acquisition unit is used to extract feature information from the second text block in the target text block set, and to evaluate the score of the second text block based on the extracted feature information to obtain the first score corresponding to the second text block;

The sorting determination unit is used to determine N first text blocks in the target text block set based on the first score corresponding to the second text block.

In some embodiments of this application, the sorting determination unit is further configured to:

Based on the first score corresponding to the second text block, the second text blocks in the target text block set are sorted to determine the sorting result. The sorting order of the second text blocks is positively correlated with the relevance of the second text blocks.

Based on the sorting results, the N second text blocks with the highest first score in the target text block set are determined, and these N second text blocks are designated as N first text blocks.

In some embodiments of this application, the evaluation acquisition unit is further used for:

Feature information is extracted from the second text block to obtain at least one of the following: the first semantic similarity between the second text block and the user's request question, the second semantic similarity between the second text block and the preceding text of the user's request question, and text quality.

The first score corresponding to the second text block is obtained by evaluating the score based on at least one of the first semantic similarity, the second semantic similarity, and the text quality.

In some embodiments of this application, the allocation and acquisition module includes:

The division and determination submodule is used to divide the N first text blocks into M parts according to the principle of balanced distribution, and determine the pre-trained model corresponding to each part of the first text block.

The rewrite acquisition submodule is used to rewrite the text by concurrently inputting M copies of the first text block into the corresponding pre-trained model based on the correspondence between the first text block and the pre-trained model, and obtaining N rewritten target text blocks.

In some embodiments of this application, the processing acquisition module is further used for:

After generating N text prompts based on N target text blocks, user request questions, and the user's corresponding historical question and answer records, the N text prompts are concurrently input into the corresponding pre-trained models based on the correspondence between the first text block and the pre-trained model to obtain a candidate response set.

Each text prompt corresponds to at least one candidate response.

In some embodiments of this application, the evaluation selection module includes:

The evaluation acquisition submodule is used to evaluate candidate responses for at least one of response fidelity, response relevance, and contextual relevance, and to obtain at least one second score.

The determination submodule is used to determine the target score corresponding to the candidate response based on at least one second score;

The selection submodule is used to select the response information that matches the user's requested question from the candidate responses in the candidate response set based on the target score.

In some embodiments of this application, when evaluating candidate responses based on at least one of response fidelity, response relevance, and contextual relevance, the candidate responses are processed based on the pre-trained model corresponding to the candidate responses;

The pre-trained model corresponding to the candidate response is determined based on the correspondence between the first text block and the pre-trained model.

This application also provides a computer program product, including a computer program/instruction. When the computer program/instruction is executed by a processor, it implements the various processes of the above-described text generation method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

This application also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the various processes of the above-described text generation method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here.

This application also provides a computer non-volatile readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the various processes of the above-described text generation method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, computer program products, electronic devices, and storage media. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products embodied on one or more computer-non-volatile readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

This application describes embodiments of methods, apparatus, computer program products, electronic devices, and storage media according to embodiments of this application using flowcharts and/or block diagrams. It should be understood that each block of the flowcharts and/or block diagrams, and combinations of blocks in the flowcharts and/or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal equipment to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal equipment, create means for implementing the functions specified in one or more flowchart blocks and/or one or more block diagrams. These computer program instructions can also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing terminal equipment to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowchart blocks and/or one or more block diagrams. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing steps for implementing the functions specified in one or more flowcharts and/or one or more block diagrams.

Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

The foregoing has described a text generation method, apparatus, computer program product, electronic device, and medium provided by this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims (20)

A text generation method, characterized by comprising: Obtain N first text blocks that are associated with the user's request and meet preset conditions, wherein the N first text blocks are determined based on retrieval from multiple dimensions; The N first text blocks are evenly distributed to M pre-trained models for text rewriting to obtain N target text blocks, where M and N are both integers greater than or equal to 1; After determining the N text prompts corresponding to the N target text blocks, the N text prompts are processed based on the M pre-trained models to obtain a candidate response set. The text prompts include the target text blocks, the user request question, and the user's corresponding historical question and answer records. The candidate responses in the candidate response set are evaluated based on the evaluation strategy, and the response information that matches the user's request question is selected. The text generation method according to claim 1, characterized in that, obtaining N first text blocks associated with the user's request question and satisfying preset conditions includes: Retrieve text information that meets the semantic relevance requirement to the user's request question across multiple dimensions, and obtain a candidate text block set including K second text blocks; The candidate text block set is subjected to text block deduplication and sorting processing to obtain the N first text blocks. The text generation method according to claim 2, characterized in that, the step of retrieving text information whose relevance to the user request question meets the semantic relevance requirement in multiple dimensions, and obtaining a candidate text block set including K second text blocks, includes: Based on the user's request question and the user's corresponding historical question and answer records, a keyword list and a target request question rewritten from the user's request question are obtained. The keyword list and the target request question are obtained based on the corresponding model service. Based on the user request question, the target request question, and the keyword list, text information is retrieved across multiple dimensions to obtain the K second text blocks. The text generation method according to claim 3, characterized in that, the step of retrieving text information from multiple dimensions based on the user request question, the target request question, and the keyword list to obtain the K second text blocks includes: Input the user request question, the target request question, and the keyword list into the target encoding model to obtain multiple text vectors; The K second text blocks are determined by retrieving text blocks from the target database that meet the semantic relevance requirements of each text vector. The target database stores vector indexes and corresponding text blocks. The text generation method according to claim 4, characterized in that it further includes: Construct a fine-tuning dataset comprising multiple text data, including question information and response information, and the text data being data from the target domain; The initial coding model is adjusted based on the fine-tuning dataset, and the target coding model is determined. The input document is divided into blocks based on the determined text block size, and the input document is generated based on information collected in the target domain. The multiple text blocks obtained from the block processing are input into the target encoding model to determine the target database. The text generation method according to claim 5, characterized in that it further includes: The block adjustment range is determined based on the maximum number of words supported by the initial encoding model. Based on the constructed multiple test datasets, determine the text block size within the block adjustment range. According to claim 2, the text generation method is characterized in that, the step of performing text block deduplication and sorting processing on the candidate text block set to obtain the N first text blocks includes: The similarity between each pair of second text blocks in the candidate text block set is calculated by iterating through the set. Based on the similarity calculation results, target text block groups are determined, wherein any two second text blocks constitute a text block group, and the text block group in which the similarity between two second text blocks is greater than a preset threshold is the target text block group. In the target text block group, identify a second text block that has a low relevance to the user's request question, and delete the identified second text block to perform deduplication of the second text block; After performing a second text block deduplication on the target text block group, a target text block set is obtained; Feature information is extracted from the second text block in the target text block set, and a score is evaluated on the second text block based on the extracted feature information to obtain the first score corresponding to the second text block; Based on the first score corresponding to the second text block, N first text blocks are determined in the target text block set. The text generation method according to claim 7, characterized in that, determining N first text blocks in the target text block set based on the first score corresponding to the second text block includes: Based on the first score corresponding to the second text block, the second text blocks in the target text block set are sorted to determine the sorting result, wherein the sorting order of the second text blocks is positively correlated with the relevance of the second text block; Based on the sorting results, the N second text blocks with the highest first scores in the target text block set are determined, and the N second text blocks are identified as the N first text blocks. The text generation method according to claim 7, characterized in that, the step of extracting feature information from the second text block in the target text block set, and evaluating the score of the second text block based on the extracted feature information to obtain a first score corresponding to the second text block, includes: Feature information is extracted from the second text block to obtain at least one of the following: the first semantic similarity between the second text block and the user request question, the second semantic similarity between the second text block and the preceding text of the user request question, and text quality. A score is evaluated based on at least one of the first semantic similarity, the second semantic similarity, and the text quality to obtain the first score corresponding to the second text block. According to claim 1, the text generation method is characterized in that, the step of evenly distributing the N first text blocks to M pre-trained models for text rewriting to obtain N target text blocks includes: According to the principle of balanced allocation, the N first text blocks are divided into M parts, and the pre-trained model corresponding to each part of the first text block is determined. Based on the correspondence between the first text block and the pre-trained model, M copies of the first text block are concurrently input into the corresponding pre-trained model for text rewriting, resulting in N rewritten target text blocks. According to the text generation method of claim 10, the step of processing the N text prompts based on the M pre-trained models to obtain a candidate response set includes: After generating N text prompts based on N target text blocks, the user's request question, and the user's corresponding historical question and answer records, the N text prompts are concurrently input into the corresponding pre-trained model based on the correspondence between the first text block and the pre-trained model to obtain the candidate response set. Each text prompt corresponds to at least one candidate response. The text generation method according to claim 10 or 11, characterized in that, the step of evaluating the candidate responses in the candidate response set based on the evaluation strategy and selecting the response information that matches the user's request question includes: The candidate responses are evaluated based on at least one of response fidelity, response relevance, and contextual relevance to obtain at least one second score; Based on the at least one second score, determine the target score corresponding to the candidate response; Based on the target score, the response information that matches the user's requested question is selected from the candidate responses corresponding to the candidate response set. The text generation method according to claim 12 is characterized in that, When evaluating the candidate responses based on at least one of response fidelity, response relevance, and contextual relevance, the candidate responses are processed based on the pre-trained model corresponding to the candidate responses. The pre-trained model corresponding to the candidate response is determined based on the correspondence between the first text block and the pre-trained model. The text generation method according to claim 13, characterized in that the method further includes: When evaluating candidate responses based on the response fidelity, the candidate responses are semantically decomposed based on the pre-trained model corresponding to the candidate responses to obtain several reference responses. The consistency between any reference response and the context is checked to determine whether any reference response is a response that matches the user's request question. The proportion of the number of reference responses that match the user's request question to the total number of reference responses is used as the second score corresponding to the response fidelity. When evaluating the candidate responses based on the response relevance, a potential question that matches the candidate response is generated based on the pre-trained model corresponding to the candidate response, and the average similarity between the potential question and the user request question is used as the second score corresponding to the response relevance; When evaluating the candidate responses based on the context relevance, the pre-trained model corresponding to the candidate responses filters out statements related to the user's request question from the user's historical question-and-answer records, and uses the proportion of statements related to the user's request question to the total number of context statements as the second score corresponding to the context relevance. The text generation method according to claim 1 is characterized in that the N text prompts are evenly distributed to the M pre-trained models based on the correspondence between the first text block and the pre-trained model. The text generation method according to claim 1 is characterized in that the text prompt information is generated based on a combination of the target text block, the user request question, and the user's corresponding historical question and answer records. A text generation device, characterized in that it comprises: The acquisition module is configured to acquire N first text blocks that are associated with the user's request question and meet preset conditions. The N first text blocks are determined based on retrieval from multiple dimensions. The allocation and acquisition module is configured to evenly distribute the N first text blocks to M pre-trained models for text rewriting, and acquire N target text blocks, where M and N are both integers greater than or equal to 1; The processing and acquisition module is configured to, after determining the N text prompts corresponding to the N target text blocks, process the N text prompts based on the M pre-trained models to obtain a candidate response set, wherein the text prompts include the target text blocks, the user request question, and the user's corresponding historical question and answer records; The evaluation and selection module is configured to evaluate the candidate responses in the candidate response set based on an evaluation strategy, and select the response information that matches the user's request question. A computer program product comprising a computer program/instructions, characterized in that, when executed by a processor, the computer program/instructions implement the text generation method according to any one of claims 1 to 16. An electronic device, characterized in that it includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the text generation method as described in any one of claims 1 to 16. A computer non-volatile readable storage medium, characterized in that a computer program is stored on the computer non-volatile readable storage medium, and the computer program, when executed by a processor, implements the text generation method as described in any one of claims 1 to 16.