Text abstract consistency detection method and device, equipment, medium and product

Abstract

The invention provides a text summary consistency detection method and device, electronic equipment, a computer storage medium and a computer program product, and the method comprises the steps: carrying out the text summary generation of an original fact text, and obtaining a target text summary; determining each original fact entity corresponding to the original fact text and each abstract fact entity corresponding to the target text abstract; performing bidirectional consistency detection on the target text abstract according to each original fact entity and each abstract fact entity to obtain a consistency detection result; the consistency detection result represents whether the target text abstract has fact consistency or not.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular, to a method, device, electronic device, computer storage medium, and computer program product for detecting the consistency of text summaries. Background Art

[0002] With the rapid development of large language models, artificial intelligence technology has been widely applied. A new content creation method emerging from the development of large language models: Artificial Intelligence Generated Content (AIGC), its innovative applications in fields such as finance, healthcare, and e-commerce have effectively improved the production efficiency of all walks of life.

[0003] However, with the in-depth practice of AIGC, a series of problems have emerged, such as content duplication, inconsistency with facts, and hallucinations. In the process of financial field practice, the issue of factual consistency in AIGC text summaries is particularly prominent. The factual consistency problem manifests as inconsistencies with facts in aspects such as entity names, entity relationships, entity metrics, emotional expressions, and semantic logic. And text summaries with factual inconsistency problems do not conform to the review rules. Therefore, there is a need to provide a method for detecting the factual consistency of text summaries. Summary of the Invention

[0004] This application provides a method, device, electronic device, computer storage medium, and computer program product for detecting the consistency of text summaries.

[0005] The technical solution of this application is implemented as follows:

[0006] This application provides a method for detecting the consistency of text summaries, the method includes:

[0007] Generating a target text summary by generating a text summary of the original factual text;

[0008] Determining each original factual entity corresponding to the original factual text and each summary factual entity corresponding to the target text summary;

[0009] Performing bidirectional consistency detection on the target text summary according to the respective original factual entities and the respective summary factual entities to obtain a consistency detection result; the consistency detection result indicates whether the target text summary has factual consistency.

[0010] This application provides a device for detecting the consistency of text summaries, the device includes:

[0011] A generation module, configured to generate a target text summary by generating a text summary of the original factual text;

[0012] A determination module, configured to determine each original fact entity corresponding to the original fact text and each summary fact entity corresponding to the target text summary;

[0013] A detection module, configured to perform a two-way consistency detection on the target text summary according to each of the original fact entities and each of the summary fact entities, to obtain a consistency detection result; the consistency detection result indicates whether the target text summary has factual consistency.

[0014] The present application provides an electronic device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the text summary consistency detection method provided by the foregoing one or more technical solutions is implemented.

[0015] The present application provides a computer storage medium, which stores a computer program; after the computer program is executed, the text summary consistency detection method provided by the foregoing one or more technical solutions can be implemented.

[0016] The present application provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the text summary consistency detection method provided by the foregoing one or more technical solutions is implemented.

[0017] An embodiment of the present application provides a text summary consistency detection method, apparatus, electronic device, computer storage medium, and computer program product. The method includes: generating a text summary for an original fact text to obtain a target text summary; determining each original fact entity corresponding to the original fact text and each summary fact entity corresponding to the target text summary; performing a two-way consistency detection on the target text summary according to each of the original fact entities and each of the summary fact entities, to obtain a consistency detection result; the consistency detection result indicates whether the target text summary has factual consistency.

[0018] It can be seen that in the embodiment of the present application, after obtaining the target text summary, by performing a two-way consistency detection on the target text summary, a more comprehensive detection can be achieved, ensuring the accuracy of the consistency detection result, and it can be widely applied to business scenarios in different fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of a text summary consistency detection method provided by an embodiment of the present application;

[0020] Figure 2 It is a flowchart of a method for determining an entity knowledge base provided by an embodiment of the present application;

[0021] Figure 3Flowchart of a method for determining an original fact text provided by an embodiment of the present application;

[0022] Figure 4 Flowchart of a method for generating a target text summary provided by an embodiment of the present application;

[0023] Figure 5 Flowchart of a method for segmenting an original fact text provided by an embodiment of the present application;

[0024] Figure 6 Flowchart of a method for segmenting a target text summary provided by an embodiment of the present application;

[0025] Figure 7 Flowchart of a method for determining a first detection result provided by an embodiment of the present application;

[0026] Figure 8 Flowchart of a method for determining a second detection result provided by an embodiment of the present application;

[0027] Figure 9 Flowchart of another method for detecting text summary consistency provided by an embodiment of the present application;

[0028] Figure 10 Flowchart of yet another method for detecting text summary consistency provided by an embodiment of the present application;

[0029] Figure 11 Schematic diagram of the composition structure of a text summary consistency detection device provided by an embodiment of the present application;

[0030] Figure 12 Schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0031] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments provided herein are only used to explain the present application and are not used to limit the present application. In addition, the embodiments provided below are partial embodiments for implementing the present application, rather than all embodiments for implementing the present application. Without conflict, the technical solutions described in the embodiments of the present application can be implemented in any combination.

[0032] It should be noted that, in the embodiments of the present application, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that a method or system including a series of elements not only includes the elements explicitly recited, but also includes other elements not explicitly listed, or further includes elements inherent in the implementation of the method or system. Without further limitation, the element defined by the statement "including one..." does not exclude the existence of other related elements in the method or system including the element (such as steps in the method or units in the system, and the unit can be a part of the processor, a part of the program or software, etc.).

[0033] For example, the text abstract consistency detection method provided in the present application includes a series of steps, but the text abstract consistency detection method provided in the present application is not limited to the recited steps. Similarly, the text abstract consistency detection device provided in the present application includes a series of modules, but the text abstract consistency detection device provided in the present application is not limited to including the explicitly recited modules, and may further include modules required for obtaining relevant information or processing based on the information.

[0034] In some embodiments of the present application, the text abstract consistency detection method can be implemented using a processor in the text abstract consistency detection device. The above-mentioned processor can be at least one of an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, and a microprocessor.

[0035] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as commonly understood by those skilled in the technical field to which the present application belongs. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0036] In the practical application of the relevant data collection and processing in the embodiments of the present application, the informed consent or separate consent of the personal information subject should be obtained strictly in accordance with the requirements of relevant national laws and regulations, and subsequent data use and processing should be carried out within the scope authorized by laws and regulations and the personal information subject.

[0037] Figure 1 This is a flowchart of a text summary consistency detection method provided by an embodiment of the present application. As Figure 1 shown, this process may include:

[0038] Step 101: Generate a text summary for the original factual text to obtain a target text summary.

[0039] In the embodiment of the present application, the text summary consistency detection method may be applied to a target field. Here, the target field may be fields such as finance, medical care, and e-commerce, which are not specifically limited herein. Subsequently, the finance field is taken as an example for illustration.

[0040] Here, the original factual text refers to the factual text associated with the target business scenario; the target business scenario is any one of at least one business scenario included in the target field; for example, when the target field is the finance field, the business scenario may include: fund product recommendation scenario, insurance product recommendation scenario, bond business scenario, etc.

[0041] In the embodiment of the present application, before generating a text summary for the original factual text, the method may include: preprocessing the basic database of the target field to obtain an entity knowledge base; filtering the entity knowledge base according to a predetermined filtering rule to obtain the original factual text.

[0042] Here, the basic database refers to the database known in the target field, and the number of basic databases may be one or more. For example, when the target field is the finance field, the basic database may include a fund database, a manager database, a consulting database, a research report database, etc.

[0043] Exemplarily, the preprocessing may include processing procedures such as data cleaning, word segmentation, and entity extraction, which are not specifically limited herein; by preprocessing the basic database of the target field, an entity knowledge base can be obtained. Exemplarily, Python may be used for preprocessing. Python, as a mainstream data development and processing technology, is designed specifically for processing and analyzing data.

[0044] Exemplarily, referring to Figure 2 , if the basic databases such as the fund database, the manager database, the consulting database, and the research report database are preprocessed, entity knowledge bases such as a fund entity knowledge base, a manager entity knowledge base, and an institutional entity knowledge base can be obtained.

[0045] Here, the predefined filtering rule refers to the filtering rule that conforms to the target business scenario; after obtaining the entity knowledge base, the predefined filtering rule that conforms to the target business scenario can be obtained; furthermore, the entity knowledge base is filtered according to the predefined filtering rule to obtain the original factual text.

[0046] Exemplarily, the predefined filtering rule may include a key information set, and the key information set may include at least one key information; among them, the key information may be determined correspondingly according to the business type or business characteristics of the business scenario; for example, when the business scenario is a fund product recommendation scenario, the predefined filtering rule may include key information such as fund product name, institution name, fund information, and fund market performance.

[0047] Exemplarily, see Figure 3 , in the case where the business scenario is a fund product recommendation scenario and the entity knowledge base includes entity knowledge bases such as a fund entity knowledge base, a manager entity knowledge base, and an institution entity knowledge base, after being filtered by the predefined filtering rule, the following original factual texts can be obtained: fund product promotion original factual text, public opinion research original factual text, institutional research and analysis original factual text, etc.

[0048] It can be understood that since the basic database in the financial field includes real financial data from multiple dimensions such as managers, fund products, company institutions, and research reports, a reliable and objective entity knowledge base can be obtained through preprocessing of the basic database in this field; if the entity knowledge base is filtered according to the predefined filtering rule that conforms to the business scenario, it can ensure that the subsequent generated text summary is more in line with the business scenario and ensure the accuracy of the subsequent detection results.

[0049] In some embodiments, before generating the text summary of the original factual text, the above method may further include: constructing the constraint conditions of the prompt engineering; using the constraint conditions to guide the output of the initial large language model to obtain the preset large language model; correspondingly, generating the text summary of the original factual text to obtain the target text summary, which may include: using the preset large language model to generate the text summary of the original factual text to obtain the target text summary.

[0050] Exemplarily, the prompt engineering is also called prompt engineering; the constraint conditions of the prompt engineering can be constructed before or after obtaining the original factual text, and the constraint conditions are used to guide the output of the initial large language model to obtain the preset large language model.

[0051] Exemplarily, a prompt is the starting point for a user to interact with a large language model. It can be an instruction, a question, or a statement used to guide or direct a large language model to generate specific content. By carefully designing the constraints of prompt engineering, the intention and expectation of the current task can be clearly conveyed to the model, enabling the large model to fully understand the task significance and generate more accurate and useful content as expected. Prompt engineering involves role design, clear task objectives, and prompt testing and optimization to help users effectively interact with the large language model. The embodiments of this application apply prompt engineering to generate a target text summary and detect and analyze the factual consistency of the target text summary.

[0052] Here, the type of the initial large language model is not limited. For example, it can be a ChatGPT model or other large language models.

[0053] In the embodiments of this application, the constraints are used to guide the large language model to generate a target text summary based on the original factual text. Here, the content of the constraints is not specifically limited. For example, the constraints can include the word count limit and format limit of the target text summary. In addition, the constraints can also include a list of original reference citations provided by the model for each factual entity in the generated text summary.

[0054] Exemplarily, a factual subject refers to the entity involved in the text content, which can include company names, fund product names, personal names, place names, event names, etc. Information such as the name, logical relationship, characteristic indicators, and sentiment tendency of the entity constitutes entity facts.

[0055] Exemplarily, refer to Figure 4 , the constraints of prompt engineering can be pre-constructed. Through prompt engineering, the constraints and the original factual text are submitted to the initial large language model. The constraints are used to guide the output of the initial large language model to obtain a preset large language model. The preset large language model is used to generate a text summary of the original factual text to obtain the target text summary and output it. It should be noted that if the constraints include a list of original reference citations provided by the model for each factual entity in the generated text summary, the preset large language model will also output a list of factual basis in addition to the target text summary. The list of factual basis includes the list of original reference citations for each factual entity in the target text summary.

[0056] It can be seen that the embodiments of this application utilize the powerful semantic understanding and analysis capabilities of the large language model to introduce a clearly itemized list of factual basis from the source of the target text summary. In this way, the quality of the target text summary can be improved.

[0057] Step 102: Determine each original fact entity corresponding to the original fact text and each summary fact entity corresponding to the target text summary.

[0058] In the embodiment of the present application, after obtaining the target text summary according to the above steps, each original fact entity corresponding to the original fact text and each summary fact entity corresponding to the target text summary can be determined for subsequent two-way consistency detection of the target text summary.

[0059] In some embodiments, determining each original fact entity corresponding to the original fact text and each summary fact entity corresponding to the target text summary may include: segmenting the original fact text using a first segmentation method to obtain each original fact entity corresponding to the original fact text; segmenting the target text summary using a second segmentation method to obtain each summary fact entity corresponding to the target text summary.

[0060] Exemplarily, the first segmentation method may be a method of segmenting according to punctuation marks; here, the punctuation marks may be semicolons, commas, or full stops, etc.; for example, see Figure 5 , assuming that the original fact text can be expressed as SourceTextContent = {"src text xxx"}, if it is segmented according to semicolons, an original fact entity list can be obtained, and this list includes each original fact entity corresponding to the original fact text, which can be correspondingly expressed as SourceFactList = {Sf1, Sf2, Sf3...}; here, Sf1, Sf2, Sf3 are each original fact entity after segmentation.

[0061] Exemplarily, the second segmentation method may be a method of segmenting according to punctuation marks or a method of segmenting according to a large language model; here, the method of segmenting according to punctuation marks is similar to the above first segmentation method, and to avoid repeated description, it will not be elaborated here; the following will be combined with Figure 6 to give an exemplary description of the method of segmenting according to a large language model.

[0062] Exemplarily, see Figure 6, after obtaining the target text summary, the target text summary can be submitted to the large language model through prompt engineering. Here, prompt engineering is used to constrain the large language model to identify and split from the text summary, which cannot be changed; using this large language model to split the target text summary, a list of summary fact entities can be obtained and output. This list includes each summary fact entity corresponding to the target text summary, which can be correspondingly expressed as AbstractFactList={Af1, Af2, Af3...}; here, Af1, Af2, Af3 are each summary fact entity after splitting. Here, the large language model can be of the same type as the above-mentioned initial large prediction model or different.

[0063] Step 103: Perform bidirectional consistency detection on the target text summary according to each original fact entity and each summary fact entity to obtain a consistency detection result.

[0064] In the embodiment of the present application, after obtaining each original fact entity corresponding to the original fact text and each summary fact entity corresponding to the target text summary according to the above steps, bidirectional consistency detection can be performed on the target text summary according to each original fact entity and each summary fact entity to obtain a consistency detection result; here, the consistency detection result characterizes whether the target text summary has factual consistency.

[0065] In some embodiments, performing bidirectional consistency detection on the target text summary according to each original fact entity and each summary fact entity to obtain a consistency detection result includes: detecting the reference relationship between each original fact entity and the target text summary to obtain a first detection result; detecting the basis relationship between each original fact entity and each summary fact entity to obtain a second detection result; and obtaining a consistency detection result according to the first detection result and the second detection result.

[0066] Exemplarily, the first detection result may include the result of whether each original fact entity is cited by the target text summary; the second detection result may include the result of whether each summary fact entity is based on any one original fact entity; here, there is no specific limitation on the order of obtaining the first detection result and the second detection result.

[0067] It should be noted that both the first detection result and the second detection result are obtained by using the large language model. The two can be processed using the same large language model or different large language models; for ease of understanding, the processes of obtaining the first detection result and the second detection result will be exemplarily described below with reference to the accompanying drawings.

[0068] Exemplarily, the detection of the reference relationship is also called top-down detection; see Figure 7, Prompt engineering is used to guide the output of large language models. By submitting each original fact entity corresponding to the original fact text and the target text summary to the large language model through prompt engineering, the large language model infers the reference relationship between each original fact entity and the target text summary, determines which original fact entities are cited by the target text summary, and outputs the first detection result.

[0069] Exemplarily, assume that each original fact entity corresponding to the original fact text can be represented as {Sf1, Sf2, Sf3...}. Submitting each original fact entity {Sf1, Sf2, Sf3...} and the target text summary to the large language model, the first detection result obtained includes: [original fact entity Sf1, 'abstract fragment 1'], [original fact entity Sf2, null]; At this time, it indicates that the original fact entity Sf1 is cited by the abstract fragment 1 in the target text summary, and the abstract fragment 1 is included in the target text summary, that is, Sf1 has an abstract reference; Sf2 is not cited by the target text summary, that is, Sf2 has no abstract reference.

[0070] Exemplarily, the detection of the basis relationship is also called bottom-up detection; see Figure 8 , Prompt engineering is used to guide the output of large language models. By submitting each original fact entity corresponding to the original fact text and each abstract fact entity corresponding to the target text summary to the large language model through prompt engineering, the large language model infers the basis relationship between each original fact entity and each abstract fact entity, determines which abstract fact entities are based on the original fact entities, and outputs the second detection result.

[0071] Exemplarily, assume that each original fact entity corresponding to the original fact text can be represented as {Sf1, Sf2, Sf3...}, and each abstract fact entity corresponding to the target text summary can be represented as {Af1, Af2, Af3...}. Submitting each original fact entity {Sf1, Sf2, Sf3...} and each abstract fact entity to the large language model, the second detection result obtained includes: [abstract fact entity Af1, original fact entity Sf1], [abstract fact entity Af2, null]; At this time, it indicates that the abstract fact entity Af1 is based on the original fact entity Sf1, that is, Af1 has a factual basis; the abstract fact entity Af2 is not based on the original fact entity, that is, Af2 has no factual basis.

[0072] Exemplarily, according to the first detection result obtained by the above steps, it can be determined whether each original fact entity has an abstract reference. According to the second detection result obtained by the above steps, it can be determined whether each abstract fact entity has a factual basis; Further, the consistency detection result can also be obtained based on the first detection result and the second detection result.

[0073] In some embodiments, obtaining a consistency detection result based on the first detection result and the second detection result may include: performing type evaluation on the first detection result and the second detection result according to a preset consistency standard definition to obtain an evaluation result; determining the consistency detection result according to the consistency type of each summary fragment corresponding to each summary fact entity.

[0074] Here, the partitioning result may include the consistency type of each summary fragment corresponding to each summary fact entity, and each summary fragment is included in the target text summary; the summary fragment corresponding to the summary fact entity means that the summary fact entity is included in the summary fragment.

[0075] Exemplarily, the consistency type may include four types: two-way consistency, one-way consistency, factual uncertainty, and false suspiciousness, and the order of the factual consistency of these four types from high to low is as follows:

[0076] 1) Two-way consistency. Suppose the summary fact entity Af1 is included in the summary fragment 1, and the summary fact entity Af1 has a factual basis [summary entity fragment Af1, original fact entity Sf1]; and the original fact entity Sf1 has a summary reference [original fact entity Sf1,'summary fragment 1'], then the consistency type of this summary fragment is defined as two-way consistency. At this time, the factual consistency of this summary fragment is the highest.

[0077] 2) One-way consistency. Suppose the summary fact entity is included in the summary fragment, and the summary fact entity has a factual basis, but the original fact entity corresponding to this factual basis does not have a summary reference, then the consistency type of this summary fragment is defined as one-way consistency. At this time, the factual consistency of this summary fragment is good.

[0078] 3) Factual uncertainty. Suppose the summary fact entity is included in the summary fragment, the original fact entity has a summary reference, but the summary fact entity does not have a factual basis, then the consistency type of this summary fragment is defined as factual uncertainty. At this time, the factual consistency of this summary fragment is average. For example, there is a relationship of [original fact entity Sf1,'summary fragment 1'] && summary fact entity Af1 Summary fragment 1 && [summary fact entity Af1, null], indicating that this summary fragment 1 fails to be directly supported by the original fact entity and there is factual uncertainty. At this time, manual review and inspection can be performed.

[0079] 4) False suspiciousness. Assume that the summary fact entity is included in the summary fragment, the summary fact entity does not have a factual basis, and the original fact entity corresponding to this factual basis does not have a summary reference, then this summary fragment is defined as false suspiciousness. At this time, the factual consistency evaluation of this summary fragment is suspicious.

[0080] In the embodiments of the present application, after obtaining the evaluation result, the consistency detection result can be determined according to the consistency types of the respective abstract segments corresponding to each abstract fact entity in the evaluation result, that is, it can be determined whether the target text abstract has factual consistency; here, the specific determination method for the consistency detection result is not limited.

[0081] In one embodiment, the number of abstract segments with a two-way consistent consistency type can be determined according to the evaluation result, and the first ratio of this number to the total number of abstract segments can be determined. If it is determined that this first ratio reaches the set value, the target text abstract has factual consistency; otherwise, it is determined that the target text abstract does not have factual consistency.

[0082] In another embodiment, the sum of the numbers of abstract segments with a two-way consistent and a one-way consistent consistency type can be determined according to the evaluation result, and the second ratio of this sum to the total number of abstract segments can be determined. If it is determined that this second ratio reaches the set value, the target text abstract has factual consistency; otherwise, it is determined that the target text abstract does not have factual consistency.

[0083] It can be understood that assuming there are abstract segments defined as factually uncertain or suspiciously false in the target text abstract, manual review can also be introduced to determine whether the corresponding abstract segments have factual consistency, update the above evaluation result according to the manual review result, and use the updated evaluation result to determine the consistency detection result; the corresponding determination process is similar to the above embodiments, and to avoid repeated description, it will not be elaborated here.

[0084] It can be seen that in the embodiments of the present application, by performing two-way consistency detection on the target text abstract generated by the preset large language model from top to bottom and from bottom to top, the problems of incomplete detection and low accuracy faced in the detection practice process can be fully solved. In addition, the problems of overgeneralization and unconscious logical contradictions of the large language model can also be effectively solved.

[0085] In order to better reflect the purpose of the present application, on the basis of the above embodiments of the present application, the financial field is taken as an example for further illustration.

[0086] Exemplarily, Figure 9 is a flowchart of another text abstract consistency detection method provided by the embodiments of the present application, as Figure 9As shown, the method includes: performing Python preprocessing on basic databases such as a fund database, a manager database, a consulting database, and a research report database to obtain an entity knowledge base; filtering the entity knowledge base according to predetermined filtering rules to obtain original fact texts; submitting the pre-constructed constraint conditions and the original fact texts to an initial large language model through prompt engineering, where the constraint conditions are used to guide the output of the initial large language model to obtain a preset large language model, using the preset large language model to generate a text summary of the original fact texts to obtain a target text summary and output it; continuing to use the large language model to perform a bidirectional consistency detection on the target text summary to obtain a consistency detection result; these steps can be specifically referred to the Figures 2 to 8 process shown above and will not be elaborated here.

[0087] Exemplarily, Figure 10 is a flowchart of another text summary consistency detection method provided by an embodiment of the present application. As Figure 10 shown, the method includes: preprocessing a basic database in the financial field to obtain an entity knowledge base; filtering the entity knowledge base according to predetermined filtering rules to obtain original fact texts, generating a text summary of the original fact texts to obtain a target text summary; performing a bidirectional consistency detection on the target text summary to obtain a consistency detection result corresponding to the target text summary; the consistency detection result indicates whether the target text summary has factual consistency.

[0088] Figure 11 is a schematic diagram of the composition structure of a text summary consistency detection device according to an embodiment of the present application. As Figure 11 shown, the device may include:

[0089] A generation module 200, configured to generate a text summary of the original fact texts to obtain a target text summary;

[0090] A determination module 201, configured to determine each original fact entity corresponding to the original fact texts and each summary fact entity corresponding to the target text summary;

[0091] A detection module 202, configured to perform a bidirectional consistency detection on the target text summary according to each original fact entity and each summary fact entity to obtain a consistency detection result; the consistency detection result indicates whether the target text summary has factual consistency.

[0092] In some embodiments, the detection module 202 is further configured to:

[0093] detect the reference relationship between each original fact entity and the target text summary to obtain a first detection result; the first detection result includes the result of whether each original fact entity is referenced by the target text summary;

[0094] Detect the basis relationship between each original fact entity and each summary fact entity to obtain a second detection result; the second detection result includes the result of whether each summary fact entity is based on any one of the original fact entities;

[0095] Obtain a consistency detection result according to the first detection result and the second detection result.

[0096] In some embodiments, the detection module 202 is further configured to:

[0097] Perform type evaluation on the first detection result and the second detection result according to a preset consistency standard definition to obtain an evaluation result; the evaluation result includes the consistency type of each summary segment corresponding to each summary fact entity, and each summary segment is included in the target text summary;

[0098] Determine the consistency detection result according to the evaluation result.

[0099] In some embodiments, the determination module 201 is further configured to:

[0100] Segment the original fact text using a first segmentation method to obtain each original fact entity corresponding to the original fact text;

[0101] Segment the target text summary using a second segmentation method to obtain each summary fact entity corresponding to the target text summary.

[0102] In some embodiments, the generation module 200 is further configured to:

[0103] Construct the constraint conditions of the prompt engineering;

[0104] Use the constraint conditions to guide the output of the initial large language model to obtain a preset large language model;

[0105] Use the preset large language model to generate a text summary of the original fact text to obtain the target text summary.

[0106] In some embodiments, before generating the text summary of the original fact text, the generation module 200 is further configured to:

[0107] Preprocess the basic database of the target domain to obtain an entity knowledge base;

[0108] Filter the entity knowledge base according to a predetermined filtering rule to obtain the original fact text.

[0109] In practical applications, the above-mentioned generation module 200, determination module 201, and detection module 202 can all be implemented by a processor located in an electronic device, and the processor can be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, and microprocessor.

[0110] In addition, each functional module in this embodiment can be integrated in a processing unit, can also exist physically separately for each unit, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional module.

[0111] If the integrated unit is implemented in the form of a software functional module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to the related technology, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method of this embodiment. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

[0112] Specifically, the computer program instructions corresponding to a text summary consistency detection method in this embodiment can be stored on storage media such as optical discs, hard disks, and USB flash drives. When the computer program instructions corresponding to a text summary consistency detection method in the storage medium are read or executed by an electronic device, any one of the text summary consistency detection methods of the foregoing embodiments is implemented.

[0113] Based on the same technical concept as the foregoing embodiments, see Figure 12 , which shows the electronic device 300 provided in the embodiment of the present application, and can include: a memory 301 and a processor 302; wherein,

[0114] The memory 301 is used to store computer programs and data;

[0115] The processor 302 is used to execute the computer program stored in the memory to implement any one of the text summary consistency detection methods of the foregoing embodiments.

[0116] In practical applications, the above-mentioned memory 301 can be a volatile memory, such as RAM; or a non-volatile memory, such as ROM, flash memory, Hard Disk Drive (HDD), or Solid-State Drive (SSD); or a combination of the above types of memories, and provide instructions and data to the processor 302.

[0117] The above-mentioned processor 302 can be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, and microprocessor. It can be understood that for different text summary consistency detection devices, the electronic devices used to implement the functions of the above-mentioned processor can also be others, and the embodiments of the present application do not make specific limitations.

[0118] The present application provides a computer program product, including a computer program, which when executed by a processor implements any one of the text summary consistency detection methods in the foregoing embodiments.

[0119] In some embodiments, the functions or modules included in the device provided in the embodiments of the present application can be used to execute the methods described in the foregoing method embodiments. The specific implementation can refer to the description of the foregoing method embodiments. For the sake of brevity, it will not be repeated here.

[0120] The descriptions of the foregoing embodiments tend to emphasize the differences between the embodiments. The same or similar parts can be referred to each other. For the sake of brevity, they will not be repeated here. The methods disclosed in the method embodiments provided in the present application can be combined arbitrarily without conflict to obtain new method embodiments. The features disclosed in the product embodiments provided in the present application can be combined arbitrarily without conflict to obtain new product embodiments. The features disclosed in the method or device embodiments provided in the present application can be combined arbitrarily without conflict to obtain new method embodiments or device embodiments.

[0121] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0122] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device generate a means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or a means for implementing the functions specified in one or more of the blocks.

[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or a means for implementing the functions specified in one or more of the blocks.

[0124] The above is only a preferred embodiment of the present application and is not intended to limit the protection scope of the present application.

Claims

1. A method for detecting text summary consistency, characterized in that, The method includes: Performing text summarization on the original factual text to obtain a target text summary; Determining each original factual entity corresponding to the original factual text and each summary factual entity corresponding to the target text summary; Performing bidirectional consistency detection on the target text summary according to each of the original factual entities and each of the summary factual entities to obtain a consistency detection result; the consistency detection result indicates whether the target text summary has factual consistency.

2. The method according to claim 1, wherein The performing bidirectional consistency detection on the target text summary according to each of the original factual entities and each of the summary factual entities to obtain a consistency detection result includes: Detecting the citation relationship between each of the original factual entities and the target text summary to obtain a first detection result; the first detection result includes the result of whether each of the original factual entities is cited by the target text summary; Detecting the basis relationship between each of the original factual entities and each of the summary factual entities to obtain a second detection result; the second detection result includes the result of whether each of the summary factual entities is based on any one of the original factual entities; Obtaining the consistency detection result according to the first detection result and the second detection result.

3. The method according to claim 1, wherein The obtaining the consistency detection result according to the first detection result and the second detection result includes: Performing type evaluation on the first detection result and the second detection result according to a preset consistency standard definition to obtain an evaluation result; the evaluation result includes the consistency type of each summary segment corresponding to each of the summary factual entities, and each summary segment is included in the target text summary; Determining the consistency detection result according to the evaluation result.

4. The method according to any one of claims 1 to 3, characterized in that, Determining each original factual entity corresponding to the original factual text and each summary factual entity corresponding to the target text summary includes: Segmenting the original factual text by a first segmentation method to obtain each original factual entity corresponding to the original factual text; Segmenting the target text summary by a second segmentation method to obtain each summary factual entity corresponding to the target text summary.

5. The method according to claim 1, wherein The method further includes: Constructing the constraint conditions of the prompt engineering; Guiding the output of the initial large language model using the constraint conditions to obtain the preset large language model; Correspondingly, the performing text summarization on the original factual text to obtain a target text summary includes: Performing text summarization on the original factual text using the preset large language model to obtain the target text summary.

6. The method according to claim 1, wherein Before performing text summarization on the original factual text, the method further includes: Preprocessing the basic database of the target domain to obtain an entity knowledge base; Filtering the entity knowledge base according to a predetermined filtering rule to obtain the original factual text.

7. A text summary consistency detection device, characterized in that, The device includes: A generation module, configured to perform text summarization on the original factual text to obtain a target text summary; A determination module, configured to determine each original factual entity corresponding to the original factual text and each summary factual entity corresponding to the target text summary; A detection module, configured to perform bidirectional consistency detection on the target text summary according to the respective original fact entities and the respective summary fact entities, so as to obtain a consistency detection result; the consistency detection result indicates whether the target text summary has factual consistency.

8. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

9. A computer storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by the processor, the method according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images