CN118586375A - Text generation method, device and system, computing device, and readable storage medium - Google Patents

Abstract

Embodiments of the present specification provide a text generation method, apparatus and system, a computing device, and a readable storage medium, where the method includes: acquiring target data to be analyzed, and generating a text generation template according to the text generated based on the target data; determining a plurality of tasks to be executed for the target data based on each piece of configuration information in the text generation template; based on the dependency relationship of the plurality of tasks, performing topological ordering on the plurality of tasks, and determining a target task which supports parallel execution in the plurality of tasks; and executing the target tasks in parallel aiming at the target data, and integrating the execution results of the target tasks to generate a text conforming to the text generation template. The method can improve the text generation efficiency.

Description

Text generation method, device and system, computing device, and readable storage medium

Technical Field

The embodiments of this specification relate to the field of computer technology, and in particular to a text generation method; they also relate to a text generation device, a text generation system, a computing device, a computer-readable storage medium, and a computer program product.

Background Art

With the development of computer technology, many tasks in people's lives and work can be performed using computers.

For example, a computer may be used to analyze some basic data required for writing to generate text that meets the writing requirements. Usually, a corresponding text generation template is set for the text to be generated, and the text generation template may include multiple configuration information, which may indicate the theme of the text to be generated and the content to be included. The computer may analyze the basic data based on the text generation template to generate text that meets the text generation template.

However, the current text generation efficiency still needs to be improved.

Summary of the invention

The embodiments of this specification provide a text generation method, which can improve the efficiency of text generation. One or more embodiments of this specification also relate to a text generation device, a text generation system, a computing device, a computer-readable storage medium, and a computer program product.

According to one aspect of an embodiment of this specification, a text generation method is provided, the method comprising:

Acquire target data to be analyzed, and a text generation template required to generate text based on the target data;

Determining, based on each piece of configuration information in the text generation template, a plurality of tasks to be performed on the target data;

Based on the dependency relationship of the multiple tasks, topologically sort the multiple tasks to determine a target task that supports parallel execution among the multiple tasks;

The target tasks are executed in parallel for the target data, and the execution results of the target tasks are integrated to generate text that conforms to the text generation template.

According to another aspect of the embodiments of this specification, there is provided a text generation device, including:

An acquisition module, used to acquire target data to be analyzed, and a text generation template required to generate text based on the target data;

A first determination module, configured to determine a plurality of tasks to be performed on the target data based on each piece of configuration information in the text generation template;

A second determining module is used to topologically sort the multiple tasks based on the dependency relationship of the multiple tasks, and determine a target task that supports parallel execution among the multiple tasks;

The text generation module is used to execute the target tasks in parallel for the target data, and integrate the execution results of each target task to generate text that conforms to the text generation template.

According to another aspect of the embodiments of this specification, there is provided a text generation system, including a terminal and a server;

The terminal is used to: send to the server a text generation instruction for target data and a text generation template required to generate text based on the target data;

The server is used to: execute the above method in response to the text generation instruction, generate text that conforms to the text generation template, and send the generated text to the terminal.

According to another aspect of the embodiments of this specification, there is provided a computing device, including: a memory and a processor;

The memory is used to store computer programs/instructions, and the processor is used to execute the computer programs/instructions. When the programs/instructions are executed by the processor, the steps of the above method are implemented.

According to another aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores a computer program/instruction, and the steps of the above method are implemented when the computer program/instruction is executed by a processor.

According to another aspect of the embodiments of this specification, a computer program product is provided, including a computer program/instruction, and when the computer program/instruction in the computer program product is executed by a processor, the steps of the above method are implemented.

In one embodiment of the present specification, the target data to be analyzed and the text generation template required to generate the text can be obtained. In the process of generating the text according to the text generation template based on the target data to be analyzed, multiple tasks to be executed for the target data can be determined based on the various configuration information in the text generation template, and the tasks to be executed are topologically sorted based on their dependencies to determine the target tasks that support parallel execution and execute them in parallel, and then the execution results of each task are integrated to generate text that conforms to the text generation template. In this way, the process of generating patterned text (such as paradigm writing) can be made more flexible, and the efficiency of text generation can be improved by reasonably executing the target tasks in the text generation process in parallel.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a simplified flow chart of a template-based text generation method provided by the related art;

FIG2 is a schematic diagram of the structure of a text generation system provided in an embodiment of this specification;

FIG3 is a flow chart of a text generation method provided by an embodiment of this specification;

FIG4 is a schematic diagram of a result of topologically sorting multiple tasks provided in an embodiment of this specification;

FIG5 is a schematic diagram of a result of topologically sorting multiple subtasks in a target task provided by an embodiment of this specification;

FIG6 is a schematic diagram of the structure of a text generation device provided in an embodiment of this specification;

FIG. 7 is a structural block diagram of a computing device provided in an embodiment of the present specification.

DETAILED DESCRIPTION

Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.

The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms of "one", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and/or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items. The term "at least one" in one or more embodiments of this specification refers to "one or more", and "multiple" refers to "two or more". The term "including" is an open description and should be understood as "including but not limited to", and may include other content on the basis of the content already described.

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

In addition, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant standards and requirements, and corresponding operation entrances are provided for users to choose to authorize or refuse.

In one or more embodiments of this specification, a large model refers to a deep learning model with large-scale model parameters, which usually contains hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than 10 trillion model parameters. A large model can also be called a foundation model/foundation model. The large model is pre-trained with large-scale unlabeled corpus to produce a pre-trained model with more than 100 million parameters. This model can adapt to a wide range of downstream tasks, and the model has good generalization ability, such as a large-scale language model (LLM), a multi-modal pre-training model, etc.

In actual application, the large model only needs a small amount of samples to fine-tune the pre-trained model and can be applied to different tasks. The large model can be widely used in natural language processing (Natural Language Processing, referred to as NLP), computer vision and other fields, and can be specifically applied to computer vision tasks such as visual question answering (Visual Question Answering, referred to as VQA), image description (Image Caption, referred to as IC), image generation, and text-based sentiment classification, text summary generation, machine translation and other natural language processing tasks. The main application scenarios of the large model include digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design, etc. The model involved in the embodiments of this specification includes but is not limited to the above-mentioned large model, which can be any type of machine learning model. For example, it can be a convolutional neural network (CNN, convolutional neural network) model, a Transformer model, etc.

With the development of computer technology and information digitization, various data are increasing, and the requirements for data efficiency are getting higher and higher. Many tasks that were originally performed manually can be performed by computers to reduce labor costs and improve task execution efficiency. For example, in many fields, it is necessary to analyze existing data and write the analysis results of the data into text according to a set template. This process can be called paradigm writing. For example, in the financial field, financial institutions need to analyze their financial reports to obtain a review text for the financial report. The type of content that needs to be included in the review text can be indicated by a set template. For example, the template can indicate that the review text must include the financial institution targeted by the review, the authenticity and completeness of the financial report, and the earnings trend of the financial structure.

At present, the application of natural language models is becoming more and more extensive. Natural language models can be applied to writing scenarios to generate texts that meet writing requirements. For example, the natural language model can include a large-scale language model LLM. Large-scale language models have demonstrated powerful capabilities in natural language processing, but they still face certain challenges in practical applications, such as the existence of hallucinations in large-scale language models, slow knowledge updates, and low information transparency in the process of outputting replies. In response to these problems, retrieval-augmented generation (RAG) technology is usually used to improve them. RAG refers to retrieving relevant information from an external knowledge base before using LLM to answer questions, so that LLM can answer questions in combination with the relevant information and the data that originally needs to be analyzed to answer the questions. Using RAG can significantly improve the accuracy of the answers output by LLM and reduce the hallucination phenomenon of LLM, especially for knowledge-intensive tasks. LLM can use RAG to display the reference source of the output answer, so that users can verify the accuracy of the answer and increase trust in the model output. The use of RAG also promotes knowledge updating and the introduction of specific domain knowledge, which can enable LLM to effectively combine its parameterized knowledge and non-parameterized external knowledge base to ensure the reliability of LLM work.

Paradigm writing can also be implemented using LLM based on RAG. For example, for the target data that needs to be analyzed in paradigm writing, LLM can be used to query external data sources to obtain relevant information of the target data, and then LLM can answer the corresponding questions (queries) or generate text. This ensures that the answers output by LLM are based on the retrieved information, which can significantly improve the accuracy and relevance of the writing content. In addition, the output of LLM can be converted into text output that meets the requirements of the paradigm through the text structure and generation rules indicated by the predefined template.

FIG1 is a simple flow chart of generating text based on a template provided by the related art. As shown in FIG1 , the text generating device can first obtain the template required for generating the text, and obtain the generated natural language text based on the configuration information in the template. If the text to be generated is a financial report review text, the template is a template for writing financial report reviews. Each line of configuration information in the template can correspond to a content that needs to be included in a generated text. The text generating device can rewrite the query statement (query) for each line of configuration information in the template in turn, and map each configuration information into a series of query statements. The query statement is also the retrieval condition corresponding to the natural language model.

The text generation device can recall information from the material library based on the obtained query statement to obtain multiple pieces of information associated with the query statement. For example, information recall can be performed through a vector recall method or a BM25 (Best Matching 25) algorithm. The material library may include the financial report to be reviewed, as well as historical data related to the financial report, such as annual reports and announcements of the same company as the financial report. The text generation device can rank the results of the recall, such as by ranking according to the degree of match with the query statement, to obtain a plurality of ranked materials.

After obtaining the template, the text generation device can also trigger the call instruction of the large language model to call the large language model to generate text based on the multiple materials after obtaining the sorted materials, and obtain a paragraph of review text corresponding to each query statement. The text generation device can also perform post-processing on each paragraph of review text obtained to obtain the final review text. For example, the text generation device verifies the review text; or determines whether the review text meets the requirements, and regenerates the review text if it does not meet the requirements; or performs certain processing to remove possible problems with the large language model itself.

However, the above process of generating text in the related art is highly linear and rigid, resulting in low text generation efficiency, that is, low paradigm writing efficiency, and lack of certain systematization. The present specification provides a text generation method, which uses a text generation process different from the related art to improve the efficiency and quality of text generation. The present specification also relates to a text generation device, a text generation system, a computing device, a computer-readable storage medium, and a computer program product. The text generation device and the computing device can be a server-side device, or can also be a terminal device with relatively powerful computing power.

FIG2 is a schematic diagram of the structure of a text generation system provided in an embodiment of the present specification. As shown in FIG2, the text generation system 10 may include a server 101 and a terminal 102, and the terminal 102 may establish a communication connection with the server 101. For example, the server 101 may be a cloud server or a server cluster, and the terminal 102 may be a smart phone, a desktop computer, a laptop computer, a tablet computer, or a smart wearable device.

The text generation method provided in this specification can be applied to the server 101. The terminal 102 can interact with the server 101, such as sending data to the server 101, or obtaining data from the server 101. For example, the terminal 102 can send the target data to be analyzed and the text generation template according to which the generated text is to be generated to the server 101, and the server 101 can process the received target data, such as executing the text generation method provided in the embodiment of this specification, obtaining the text that conforms to the text generation template, and feeding back the generated text to the terminal 102. The terminal 102 can send a text generation request to the server 101, and the request can carry the target data and the text generation template. The text generation template can exist in the form of a document, or the terminal 102 can directly send the content in the template to the server 101 in the form of a dialogue.

In some implementations, if the computing capability of the terminal 102 is relatively strong, the terminal 102 may not be connected to the server 101 , and the terminal 102 may directly execute the text generation method provided in the embodiments of this specification.

FIG3 is a flow chart of a text generation method provided in an embodiment of the present specification. The following is an introduction to the method applied to a text generation device, and the text generation device is taken as a server. As shown in FIG3, the method may include the following steps:

Step 302: Obtain target data to be analyzed and a text generation template required to generate text based on the target data.

In the embodiments of this specification, the text generation device can analyze certain target data to generate the required text based on the target data. The target data can be specified by the terminal, and the user can send the target data to be analyzed to the text generation device through the terminal. For example, the target data can be a financial report, and the text to be generated for the target data can be a comment text of the financial report. The target data can also be data in other fields, such as relevant data in the project execution process of the enterprise, and the text to be generated can be a summary text of the project execution process, etc.

The text generation device can generate text for the target data according to a certain text generation template, and the text generation template can be used to specify the content required to be included in the text to be generated. For example, the text generation template can include the theme of the text, evaluation indicators, and questions that need to be answered. The text generation template can also be specified by the terminal, such as being sent by the terminal when sending the target data. Optionally, the text generation template can also be stored in the text generation device, and the terminal can only indicate the type of text to be generated, and the text generation device determines the corresponding text generation template based on the text type.

Optionally, the terminal may not directly send the target data and the text generation template to the text generation device, but may send indication information of the target data to the text generation device, and the text generation device may obtain the target data and the text generation template from other data storage devices based on the indication information.

Step 304: Determine multiple tasks to be performed on the target data based on each configuration information in the text generation template.

Different configuration information in the text generation template may correspond to different contents in the text to be generated. For example, each configuration information may occupy a line of the text generation template, which may be the basic unit of the text generation template, corresponding to a basic content in the final generated text.

The text generation device may execute at least one task for each piece of configuration information to generate text corresponding to the configuration information. Accordingly, at least one corresponding task may be determined based on each piece of configuration information, and the at least one task may be a task to be executed for the target data.

In the embodiments of this specification, the text generation device may also need to perform tasks unrelated to the text generation template to generate the required text. The text generation device can also determine these tasks as tasks to be performed for the target data. For example, for the generation process of various texts, the text generation device may need to perform certain fixed tasks, such as data formatting tasks, index creation tasks, etc. The text generation device can also determine these tasks as tasks to be performed for the target data.

Step 306: Topologically sort the multiple tasks based on the dependency relationship of the multiple tasks, and determine the target task that supports parallel execution among the multiple tasks.

Among the tasks to be executed for the target data, some tasks have dependencies, while others do not. This dependency is the order of execution. For example, if the execution of task A requires the execution result of task B, then task A and task B have a dependency relationship, and task A depends on task B.

After determining the tasks to be executed for the target data, the text generation device can also analyze the dependencies of the tasks, and then topologically sort the tasks based on the dependencies, and generate a directed acyclic graph (DAG) for each task. Based on the result of the topological sorting, the execution process of each task can be intuitively determined, and the tasks that need to be executed in sequence can be determined, as well as the target tasks that do not have a sequential execution order and support parallel execution.

For example, the task generation device can retrieve the target data only after formatting the target data, and the formatting task and the retrieval task have a sequence. The formatting of different data may not have a sequence of execution, and parallel execution is supported.

Step 308: Execute target tasks in parallel for target data, and integrate the execution results of each target task to generate text that conforms to the text generation template.

After the text generation device performs topological sorting on each task and determines the execution order of each task, it can execute each task on the target data in this order. For tasks with a sequential execution order, they are executed in sequence in this order, and for target tasks that support parallel execution, they are executed in parallel to improve the efficiency of task execution. For each task, the task generation device starts executing the task after the task on which it depends is completed.

Since the execution results of each target task executed in parallel are independent, but a complete text needs to be generated ultimately, the text generation device can also integrate the execution results of each target task according to the relationship between the target tasks and the relationship between the target tasks and the generation of the overall text to generate text that conforms to the text generation template.

In some embodiments, the step of obtaining the complete text in the text generation process only relies on multiple target tasks executed in parallel, and the execution results of each target task are directly integrated according to the set integration method to obtain the final generated text. In other embodiments, after multiple target tasks are executed in parallel, other tasks need to be executed. At this time, after integrating the execution results of multiple target tasks, it is necessary to continue processing based on the execution results until the step of generating the complete text is completed to obtain the text that conforms to the text generation template. In the text generation process, there may be multiple groups of target tasks, each group of target tasks may include multiple target tasks, multiple target tasks in the same group may be executed in parallel, and target tasks in different groups may have dependencies, or may also be executed in parallel.

The text generation method of the embodiment of this specification is equivalent to splitting the paradigm writing process into multiple independent tasks. According to the dependencies between these tasks, the target tasks that support parallel processing are processed in parallel, which can effectively optimize the processing flow and improve the efficiency of paradigm writing.

In some embodiments, executing the target task in parallel for the target data in the above step 308 includes: for the target data, using the coroutine pool to call multiple coroutines to execute the target tasks in parallel. The coroutine runs on the thread. When a coroutine is executed, it can choose to give up actively and let another coroutine run on the current thread. For example, a coroutine can be called for each target task, and the thread running by the coroutine is specifically used to execute the target task. The use of coroutines can effectively control the concurrent execution of multiple tasks. The embodiments of this specification can also use the coroutine pool to limit and reuse each coroutine to avoid unlimited opening of coroutines and waste of resources.

Coroutines do not increase the number of threads. They just run multiple coroutines on the basis of threads through time-sharing multiplexing. Moreover, the switching of coroutines is completed in user state, and the cost of switching is much smaller than the cost of switching threads from user state to kernel state.

In some implementations, a maximum number of concurrent tasks may be set. Accordingly, executing target tasks in parallel for target data in step 308 may include: for the target data, based on the set number of concurrent tasks, controlling target tasks less than or equal to the number of concurrent tasks to be executed in parallel.

For example, if the number of target tasks that support parallel execution is greater than the set number of concurrent tasks m, the text generation device may first execute m target tasks in parallel, and then execute the remaining target tasks. Optionally, the text generation device may also divide each target task into multiple groups of target tasks, with the number of target tasks in each group being slightly different and less than or equal to m, and then execute the multiple groups of target tasks in sequence, and execute the target tasks in each group in parallel.

In the embodiments of this specification, a natural language model (such as a large language model LLM) can be used to implement text generation, and the efficiency of paradigm writing based on LLM can be improved accordingly. For example, in step 306, target tasks that can be executed in parallel by the natural language model can be determined, and in step 308, the natural language model can be called to execute one or more tasks for the target data.

In summary, in the text generation method provided by the embodiment of this specification, the target data to be analyzed and the text generation template required to generate the text can be obtained. In the process of generating the text based on the target data to be analyzed according to the text generation template, multiple tasks to be executed for the target data can be determined based on the various configuration information in the text generation template, and the tasks to be executed are topologically sorted based on their dependencies to determine the target tasks that support parallel execution and execute them in parallel, and then the execution results of each task are integrated to generate text that conforms to the text generation template. In this way, the process of generating patterned text (such as paradigm writing) can be made more flexible, and the efficiency of text generation can be improved by reasonably executing the target tasks in the text generation process in parallel.

In some implementations, the above step 304 determines multiple tasks to be performed on the target data based on each configuration information in the text generation template, including:

Based on each configuration information in the text generation template, determine at least one data query task corresponding to each configuration information, and an integration task for query results of each data query task;

For each data source in the target data, determine a data formatting task and at least one index building task corresponding to each data source;

Based on the data query task, the integration task, the data formatting task and the index building task, multiple tasks to be executed for the target data are obtained.

In the embodiment of the present specification, the tasks to be performed in the text generation process may include: a data formatting task and an index building task for the target data, and may also include a data query task for the target data, and a combination task for the query results of each data query task. The combination task may belong to the tasks in the post-processing process in the text generation.

Each configuration information in the text generation template may correspond to at least one data query task, and the data query task may be a task of querying required data from the target data and its related data (such as from the above-mentioned material library). The data query task may be a task executed using a natural language model. The text generation device may integrate the execution results of each data query task to generate the required text, and accordingly may determine the integration task that depends on the data query task.

The data in the target data to be analyzed may be relatively complex, such as data from multiple data sources, and the data formats of different data sources may be different. The data from different data sources described in the embodiments of this specification may correspond to data collected from different devices, or may correspond to data collected from different functional modules of the same device, and the content, type or form of the data from different data sources may be different. In order to facilitate the analysis of the target data (such as query), the text generation device may format the target data before executing the data query task to convert the target data into data in a unified format. Accordingly, the formatting process may correspond to a data formatting task. The text generation device may perform formatting processing on different data sources respectively, and accordingly, each data source may correspond to a data formatting task.

After formatting the data from different data sources, the text generation device can also build an index for the data from each data source, so as to quickly determine the required data for analysis based on the index. Accordingly, an index building task can be set for each data source. Optionally, only one index building task can be established for one data source; or one data source may include different types of data, and different index building tasks can be established for different types of data to directly determine the corresponding type of data based on the index.

The data formatting tasks, index building tasks, data query tasks and integration tasks for the same data source have sequential dependencies, so these tasks for the unified data source do not support parallel execution, but these tasks for different data sources can be executed in parallel. For example, the above step 306 topologically sorts the multiple tasks based on the dependencies of the multiple tasks to determine the target tasks that support parallel execution among the multiple tasks, which may include: topologically sorting the multiple tasks based on the dependencies of the multiple tasks to determine that the data formatting tasks, index building tasks and data query tasks for different data sources are target tasks that support parallel execution.

FIG4 is a schematic diagram of a result of topological sorting of multiple tasks provided by an embodiment of the present specification. As shown in FIG4 , the target data may include data from two data sources, such as data from one data source may be some financial data, and data from another data source may be some other descriptive information. A data formatting task may be performed for the two types of data to obtain financial data and a PDF parsing file in a standard format shown in FIG4 . Index building tasks may be performed for the two types of data, respectively. FIG4 takes an example of performing an index building task on financial data to obtain a financial data index; performing two different index building tasks on a block parsing file to obtain an index of table data and an index of text data, respectively. Optionally, the text generation device may also obtain other indexes other than the target data, such as the other index may be an index of external data related to the target data.

The text generation device can call multiple data query instances to perform data query tasks based on the indexes of each data. Figure 4 takes the multiple data query instances corresponding to the financial data query task, the problem query task and the feature query task as an example. The data targeted by different query tasks may be intersecting or overlapping. As shown in Figure 4, the financial data query task can perform data query based on the financial data index and the table data index, the problem query task can perform data query based on the text data index and the external data index, and the feature query task can perform data query based on the table data index, the text data index and the external data index. After the multiple data query tasks, post-processing can be performed, such as integrating the query results of each data query task.

For example, as shown in Figure 4, the data formatting task and the index building task for the two types of data can be executed in parallel, and each data query task can be executed in parallel after the index building task is completed. The task of obtaining the external data index can also be executed in parallel with the data formatting task and the index building task.

In some embodiments, the above step 306 topologically sorts the multiple tasks based on the dependency relationship of the multiple tasks, and determines the target task that supports parallel execution among the multiple tasks, which may include: topologically sorting the multiple tasks based on the dependency relationship of the multiple tasks, and determining that different index building tasks for the same data source are target tasks that support parallel execution. For example, as shown in FIG4, two index building tasks executed for block parsing files are target tasks that support parallel execution. After the data formatting process is performed to obtain the block parsing file, the two index building tasks can be executed in parallel.

In the embodiments of this specification, each task can support mutual nesting. For example, any task in the text generation process can include multiple subtasks, and the multiple subtasks can also be executed in parallel in a manner similar to the execution of the aforementioned tasks. Taking the target task as an example, the text generation device can topologically sort the multiple subtasks based on the dependency relationship of the multiple subtasks included in the target task, and determine the target subtask that supports parallel execution among the multiple subtasks. Furthermore, the target subtasks can be executed in parallel, and the execution results of each target subtask can be integrated to obtain the execution result of the target task.

For example, the target task is a data query task, which can be a task to be executed based on the configuration information in the text generation template. The data query task can also be a query task for data of other set types, such as a task that directly queries data that meets certain characteristic conditions. The target task can include query statement rewriting subtasks, data recall subtasks, analysis subtasks, and post-processing subtasks that need to be executed in sequence.

The process of executing the target task for the target data in the above step 308 may include:

Step s2: For each configuration information in the text generation template, execute the query statement rewriting subtask to obtain prompt information of the natural language model.

Each configuration information in the text generation template may indicate some content that needs to be included in the text to be generated. The text generation device may map the configuration information to prompt information (prompt) required to generate text using a natural language model. For example, the mapping is achieved by executing a query statement rewriting subtask.

Step s4: executing the subtask of recalling the relevant data of the prompt information for the target data.

For any prompt information, the text generation device can recall the relevant data of the prompt information from the material library to obtain multiple materials. The relevant information can be determined based on the degree of matching with the prompt information. For the recall subtask, please refer to the relevant introduction about information recall in Figure 1, which will not be repeated here. For example, when the target data includes multiple data sources, the recall subtask of the relevant data of the prompt information can be executed in parallel for different data sources.

Optionally, the text generation device may recall relevant data based on the data index, and may recall data from different storage locations for different data indexes. Recall subtasks may be executed in parallel for different data indexes.

Step s6: Utilize the natural language model to perform analysis subtasks in parallel on multiple related data to obtain the relevance between each related data and the prompt information.

After recalling multiple relevant data of the prompt information, the text generation device can analyze each relevant data using a natural language model to generate text corresponding to the prompt information based on the relevant data. This process can refer to the introduction of generating the review text in Figure 1. The text generation device can also determine the accuracy of the analysis result based on the analysis result of the relevant data (such as the generated text) to determine the correlation between the relevant data and the prompt information. The correlation can be used as a score of the relevant data.

Step s8: Execute the post-processing subtask based on the relevant data and the correlation between the data and the prompt information to obtain the result text corresponding to the prompt information, and determine the execution result of the data query task based on the result text.

The text generation device may perform post-processing subtasks, such as integrating the analysis results based on the relevant data, or further processing or adjusting the analysis results to obtain a result text corresponding to the prompt information.

FIG5 is a schematic diagram of a result obtained by topologically sorting multiple subtasks in a target task provided by an embodiment of the present specification, and FIG5 is explained by taking the problem query task in FIG4 corresponding to the target task as an example. As shown in FIG5, the text generation device rewrites the configuration information to obtain corresponding prompt information. Afterwards, the recall subtask can be executed based on the text data index and the external data index respectively, and the prompt information can be recalled from the text data and the external data, such as recalling file 1 and file 2 from the text data, and recalling file 3 and file 4 from the external data. For each file, the large language model LLM service can be used to execute the analysis subtask to obtain the analysis result, and each file can be scored based on the analysis result. Afterwards, post-processing can be performed based on the analysis results of each file to obtain the result text generated by the target task.

For the target task shown in FIG5 , the process of executing the recall subtask based on the text data index and the external data index can be executed in parallel, and the process of calling the LLM service for analyzing each recalled file can also be executed in parallel.

In the embodiment of this specification, the text generation device may set a retry mechanism, and may re-execute any task after the task fails. The above step s8 may include:

Perform post-processing subtasks based on the relevant data and their relevance to the prompt information;

When the relevance between the processing result of the post-processing subtask and the prompt information does not meet the relevance condition, the recall subtask, the analysis subtask and the post-processing subtask are re-executed;

When the correlation between the processing result of the post-processing subtask and the prompt information meets the correlation condition, the processing result is determined as the result text corresponding to the prompt information.

For the post-processing subtask, please refer to the relevant introduction about post-processing in the aforementioned FIG. 1. By way of example, during the process of executing the post-processing subtask, the text generation device may integrate the analysis results of the relevant data with a correlation greater than a threshold value to obtain the processing result of the post-processing subtask. By way of another example, the text generation device may further adjust the analysis results of the relevant data with a correlation greater than a threshold value to obtain the processing result of the post-processing subtask.

The text generation device can also determine whether the processing result of the post-processing subtask meets the text generation requirements corresponding to the prompt information, such as by determining whether the correlation between the processing result and the prompt information meets the correlation condition (such as whether it is greater than the correlation threshold) to determine whether the text generation requirements are met. If the correlation condition is not met, the recall subtask and the analysis subtask can be re-executed based on the prompt information; if the correlation condition is met, the post-processing result can be determined as the result text corresponding to the prompt information, that is, it is determined as the execution result of the target task. As shown in Figure 5, if the post-processing result is invalid, a retry can be performed to return to the process after the rewrite.

Optionally, the number of retries and the retry time interval may be set for the retry mechanism. The text generation device may re-execute a task only when the number of retries for a task is less than a set number threshold and the time interval between the last execution of the task and the last execution of the task is greater than the set retry time interval. For example, the recall subtask and the analysis subtask may be re-executed.

In the embodiment of this specification, the text generation device can call the DAG engine to implement the above steps 306 and 308. The DAG engine can define a runnable interface, a basic task class and a DAG class, and the basic task class and the DAG class correspond to executable entities.

Among them, the Runnable interface defines the basic structure of all executable entities. As an abstract base class, the Runnable interface ensures that different types of execution entities follow a unified behavior pattern. The basic task class inherits from the Runnable interface and can provide the basic implementation of task execution and additional functions (such as a retry mechanism). The DAG class is the core of the DAG engine. Its class extends the Runnable interface so that it can be used not only as an independent task execution unit, but also as a collection of tasks. The DAG class supports nesting, allowing a DAG to be added as a subtask to a larger DAG. This nesting function provides great flexibility and powerful organizational capabilities for building complex and multi-level task execution processes.

For example, during the text generation process, the text generation device can first define the subclass instance of the basic task class and determine the specific task logic implemented by each subclass instance. A DAG instance can also be created to add each subtask to be executed to the DAG instance to achieve DAG nesting. After that, the asynchronous execution method of the DAG class can be called to perform topological sorting of tasks and manage the execution order and concurrency of each subtask in the DAG instance. The nesting capability of DAG can be used to build a complex task structure and optimize the organization and execution of the text generation process.

In the embodiment of this specification, the text generation device can split the entire paradigm writing process into several basic task instances and DAG instances. The basic task instance may include a data formatting task instance, an index creation task instance, and a post-processing task instance. The DAG instance may include various data query task instances.

In some embodiments, for the above interfaces and classes, corresponding properties and methods to be executed can be defined for them, and when calling the corresponding instance to execute the task, the properties and methods need to be followed. For example, the properties of the Runnable interface may include: an identifier (name, used to identify and track the corresponding instance), an execution result (output), a list of dependent tasks (dependencies), the number of unresolved dependencies (dependency_count) of other tasks (dependents) that depend on this task, and the current status (status, such as not started, running, completed, and failed).

The methods corresponding to the Runnable interface may include: initializing basic properties, such as can be represented by the function initialize(); adding dependent tasks, such as can be represented by the function add_dependencies(tasks); checking task readiness, such as can be represented by the function is_ready(); and defining asynchronous execution logic, such as can be represented by the function async_run().

The properties of the basic task class may include: task name, task input parameters, task output results, list of dependent tasks, list of other tasks that depend on this task, number of task retries (max_retries) and retry interval (retry_delay), and current task status. The corresponding methods of the basic task class may include: adding a set of dependent tasks (add_dependencies), checking whether the task is ready to execute (is_ready), asynchronously executing tasks (including processing retry logic and updating dependent task status), and setting the specific task logic that the subclass needs to implement (execute).

The properties of the DAG class can include: the maximum number of concurrently executed tasks (max_workers), a list of all tasks in the DAG, and a property indicating whether the tasks have been topologically sorted (Sorted). The corresponding methods of the DAG class can include: adding a single task (add_task) or a list of tasks (add_tasks) to the DAG, and asynchronously executing all subtasks in the DAG.

In the embodiments of this specification, a new text generation mode is adopted to optimize the generation link of paradigm writing, which can fill the technical gap and make the text generation process more efficient and controllable. Optionally, the text generation device can also use other task scheduling algorithms or parallel processing frameworks other than the DAG engine to optimize the text generation process.

In summary, in the text generation method provided by the embodiment of this specification, the target data to be analyzed and the text generation template required to generate the text can be obtained. In the process of generating the text based on the target data to be analyzed according to the text generation template, multiple tasks to be executed for the target data can be determined based on the various configuration information in the text generation template, and the tasks to be executed are topologically sorted based on their dependencies to determine the target tasks that support parallel execution and execute them in parallel, and then the execution results of each task are integrated to generate text that conforms to the text generation template. In this way, the process of generating patterned text (such as paradigm writing) can be made more flexible, and the efficiency of text generation can be improved by reasonably executing the target tasks in the text generation process in parallel.

Corresponding to the above method embodiment, this specification also provides a text generation device embodiment, which can be a server. FIG6 is a structural diagram of a text generation device provided in an embodiment of this specification. As shown in FIG6, the text generation device includes:

An acquisition module 602 is used to acquire target data to be analyzed and a text generation template required to generate text based on the target data;

A first determination module 604, configured to determine a plurality of tasks to be performed on the target data based on each piece of configuration information in the text generation template;

A second determining module 606 is used to perform topological sorting on the multiple tasks based on the dependency relationship of the multiple tasks, and determine a target task that supports parallel execution among the multiple tasks;

The text generation module 608 is used to execute target tasks in parallel for target data, and integrate the execution results of each target task to generate text that conforms to the text generation template.

Optionally, the first determining module 604 is used to:

Based on each configuration information in the text generation template, determine at least one data query task corresponding to each configuration information, and an integration task for query results of each data query task;

For each data source in the target data, determine a data formatting task and at least one index building task corresponding to each data source;

Based on the data query task, the integration task, the data formatting task and the index building task, multiple tasks to be executed for the target data are obtained.

Optionally, the data formatting task, index building task, data query task and integration task for the same data source have a sequential dependency relationship; the second determining module 606 is used to:

Based on the dependency relationship of the multiple tasks, the multiple tasks are topologically sorted to determine data formatting tasks, index building tasks and data query tasks for different data sources, which are target tasks that support parallel execution.

Optionally, the second determining module 606 is used to:

Based on the dependency relationship of the multiple tasks, the multiple tasks are topologically sorted to determine that different index building tasks for the same data source belong to target tasks that support parallel execution.

Optionally, the target task includes a data query task, and the text generation module 608 is used to:

For each configuration information in the text generation template, execute the query statement rewriting subtask to obtain the prompt information of the natural language model;

Perform the recall subtask of the relevant data of the prompt information for the target data;

Using a natural language model to perform analysis subtasks in parallel on the multiple related data, to obtain a correlation between each related data and the prompt information;

The post-processing subtask is executed based on each relevant data and its correlation with the prompt information to obtain the result text corresponding to the prompt information, and the execution result of the data query task is determined based on the result text.

Optionally, the text generation module 608 is used to:

In the case that the target data includes multiple data sources, the subtask of recalling data related to the prompt information is executed in parallel for different data sources.

Optionally, the text generation module 608 is used to:

Perform post-processing subtasks based on the relevant data and their relevance to the prompt information;

When the relevance between the processing result of the post-processing subtask and the prompt information does not meet the relevance condition, the recall subtask, the analysis subtask and the post-processing subtask are re-executed;

When the correlation between the processing result of the post-processing subtask and the prompt information meets the correlation condition, the processing result is determined as the result text corresponding to the prompt information.

Optionally, the text generation module 608 is used to:

For the target data, based on the set number of concurrent tasks, control the parallel execution of target tasks that are less than or equal to the number of concurrent tasks.

Optionally, the text generation module 608 is used to:

For the target data, the coroutine pool is used to call multiple coroutines to execute the target tasks in parallel.

In summary, in the text generation device provided by the embodiment of the present specification, the acquisition module can acquire the target data to be analyzed and the text generation template required to generate the text. In the process of generating the text based on the target data to be analyzed according to the text generation template, the first determination module can determine multiple tasks to be executed for the target data based on each configuration information in the text generation template, and the second determination module topologically sorts the tasks to be executed based on their dependencies to determine the target tasks that support parallel execution and execute them in parallel, and then the text generation module integrates the execution results of each task to generate text that conforms to the text generation template. In this way, the process of generating patterned text (such as paradigm writing) can be made more flexible, and the efficiency of text generation can be improved by reasonably executing the target tasks in the text generation process in parallel.

Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the text generation device, since it is basically similar to the text generation method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the text generation method embodiment.

Fig. 7 is a block diagram of a computing device provided in an embodiment of the present specification. The components of the computing device 700 include but are not limited to a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and the database 750 is used to store data.

The computing device 700 also includes an access device 740 that enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 740 may include one or more of any type of network interface (e.g., a network interface card (NIC)) that is wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a world-wide interoperability for microwave access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, and a near field communication (NFC).

In one embodiment of the present specification, the above components of the computing device 700 and other components not shown in FIG. 7 may also be connected to each other, for example, through a bus. It should be understood that the computing device structure block diagram shown in FIG. 7 is only for illustrative purposes and is not intended to limit the scope of the present specification. Those skilled in the art may add or replace other components as needed.

The computing device 700 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 700 may also be a mobile or stationary server.

The processor 720 is used to execute computer programs/instructions, which implement the steps in the above-mentioned text generation method when executed by the processor.

As for the computing device embodiment, since it is basically similar to the text generation method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the text generation method embodiment.

One embodiment of the present specification also provides a computer-readable storage medium, which stores computer instructions, and when the computer instructions are executed by the processor, the steps of the above-mentioned text generation method are implemented. The computer instructions include computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable storage media do not include electric carrier signals and telecommunication signals.

An embodiment of the present specification also provides a computer program product, including a computer program/instruction, which implements the steps of the above-mentioned text generation method when the computer program/instruction is executed in a processor.

As for the computer-readable storage medium embodiment and the computer program product embodiment, since they are basically similar to the text generation method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the text generation method embodiment.

The above is a description of a specific embodiment of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

It should be noted that the above is a description of a specific embodiment of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of the present specification.

In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can understand and use this specification well.

Claims (14)

1. A text generation method, characterized in that the method comprises: Acquire target data to be analyzed, and a text generation template required to generate text based on the target data; Determining, based on each piece of configuration information in the text generation template, a plurality of tasks to be performed on the target data; Based on the dependency relationship of the multiple tasks, topologically sort the multiple tasks to determine a target task that supports parallel execution among the multiple tasks; The target tasks are executed in parallel for the target data, and the execution results of the target tasks are integrated to generate text that conforms to the text generation template. 2. The method according to claim 1, characterized in that the step of determining a plurality of tasks to be performed on the target data based on each piece of configuration information in the text generation template comprises: Based on each piece of configuration information in the text generation template, determining at least one data query task corresponding to each piece of configuration information, and an integration task for query results of each data query task; For each data source in the target data, determining a data formatting task and at least one index building task corresponding to each data source; Based on the data query task, the integration task, the data formatting task and the index building task, a plurality of tasks to be executed for the target data are obtained. 3. The method according to claim 2, characterized in that the data formatting task, the index building task, the data query task and the integration task for the same data source have a sequential dependency relationship; the topological sorting of the multiple tasks based on the dependency relationship of the multiple tasks to determine the target task that supports parallel execution among the multiple tasks comprises: Based on the dependency relationship of the multiple tasks, the multiple tasks are topologically sorted to determine that the data formatting tasks, the index building tasks, and the data query tasks for different data sources are target tasks that support parallel execution. 4. The method according to claim 2, characterized in that the topological sorting of the multiple tasks based on the dependency relationship of the multiple tasks and determining the target tasks that support parallel execution among the multiple tasks comprises: Based on the dependency relationship of the multiple tasks, the multiple tasks are topologically sorted to determine that different index building tasks for the same data source belong to target tasks that support parallel execution. 5. The method according to any one of claims 1 to 4, characterized in that the target task includes a data query task, and the parallel execution of the target task for the target data comprises: For each configuration information in the text generation template, execute the query statement rewriting subtask to obtain prompt information of the natural language model; Execute a subtask of recalling data related to the prompt information for the target data; Using a natural language model to perform analysis subtasks in parallel on the plurality of related data, to obtain a correlation between each related data and the prompt information; A post-processing subtask is executed based on each relevant data and its relevance to the prompt information to obtain a result text corresponding to the prompt information, and an execution result of the data query task is determined based on the result text. 6. The method according to claim 5, characterized in that the recall subtask of the relevant data of the prompt information executed for the target data comprises: In the case that the target data includes multiple data sources, the subtask of recalling the relevant data of the prompt information is executed in parallel for different data sources. 7. The method according to claim 5, characterized in that the performing of the post-processing subtask based on each relevant data and its relevance to the prompt information to obtain the result text corresponding to the prompt information comprises: Performing a post-processing subtask based on each relevant data and its relevance to the prompt information; If the correlation between the processing result of the post-processing subtask and the prompt information does not meet the correlation condition, re-execute the recall subtask, the analysis subtask and the post-processing subtask; When the correlation between the processing result of the post-processing subtask and the prompt information meets the correlation condition, the processing result is determined as the result text corresponding to the prompt information. 8. The method according to any one of claims 1 to 4, characterized in that the parallel execution of the target tasks for the target data comprises: For the target data, based on the set number of concurrent tasks, target tasks less than or equal to the number of concurrent tasks are controlled to be executed in parallel. 9. The method according to any one of claims 1 to 4, characterized in that the parallel execution of the target tasks for the target data comprises: For the target data, a coroutine pool is used to call multiple coroutines to execute the target task in parallel. 10. A text generation device, characterized in that the text generation device comprises: An acquisition module, used to acquire target data to be analyzed, and a text generation template required to generate text based on the target data; A first determination module, configured to determine a plurality of tasks to be performed on the target data based on each piece of configuration information in the text generation template; A second determining module is used to topologically sort the multiple tasks based on the dependency relationship of the multiple tasks, and determine a target task that supports parallel execution among the multiple tasks; The text generation module is used to execute the target tasks in parallel for the target data, and integrate the execution results of each target task to generate text that conforms to the text generation template. 11. A text generation system, characterized by comprising a terminal and a server; The terminal is used to: send to the server a text generation instruction for target data and a text generation template required to generate text based on the target data; The server is used to: execute the method described in any one of claims 1 to 9 in response to the text generation instruction, generate text that conforms to the text generation template, and send the generated text to the terminal. 12. A computing device, comprising: a memory and a processor; The memory is used to store computer programs/instructions, and the processor is used to execute the computer programs/instructions. When the programs/instructions are executed by the processor, the method according to any one of claims 1 to 9 is implemented. 13. A computer-readable storage medium, characterized in that a computer program/instruction is stored therein, and when the computer program/instruction is executed by a processor, the method according to any one of claims 1 to 9 is implemented. 14. A computer program product, characterized in that it comprises a computer program/instruction, and when the computer program/instruction in the computer program product is executed by a processor, the method according to any one of claims 1 to 9 is implemented.