Recording contradiction detection method and device, storage medium and equipment

Abstract

The invention discloses a record contradiction detection method and device, a storage medium and equipment. The method comprises the following steps: firstly, obtaining a to-be-detected target record text; then performing intention recognition on the target record text based on a pre-constructed record contradiction detection point set, and extracting case key elements and association relationships under the intention to form a contradiction problem to be detected; and then constructing a first prompt by using the to-be-detected contradictory question, inputting the first prompt into a preset large language model to obtain a reply result for the to-be-detected contradictory question output by the large language model, and determining a contradictory detection result corresponding to the target record text according to the reply result. Visibly, the contradiction points in the target record text are detected by using the preset large language model based on the constructed record contradiction detection point set, so that not only is the strong dependence on the knowledge graph construction effect broken through, but also the record contradiction point detection efficiency and accuracy are improved, and the detection effect is improved.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a method, device, storage medium and equipment for detecting contradictions in transcripts. Background Art

[0002] With the advent of the information age and the intelligent age, more and more information technology and artificial intelligence technologies have been applied to the judicial field, providing great convenience for the public and judicial staff. For telecommunications network-related cases involving a large number of people, a large number of transcripts are usually generated during the interrogation and questioning process. The key information in the transcripts, such as personnel information, company information, and information on the amount involved, is complex, which often brings a large workload and challenges to the staff in analyzing and handling the case. Therefore, whether the contradictions in the information in the transcripts can be detected quickly and accurately is of great significance to the staff in improving the efficiency of case handling.

[0003] The existing method for detecting contradictions in transcripts mainly uses the method of constructing a knowledge graph to detect contradictions in the transcript content. However, this method strongly relies on the knowledge graph construction effect, the construction process is complex, and it is difficult to implement. It is also impossible to enumerate all case business scenarios, which greatly reduces the accuracy of transcript contradiction detection. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to provide a method, device, storage medium and equipment for detecting contradictions in transcripts, which can improve the efficiency and accuracy of detecting contradictions in transcripts, thereby improving the detection effect of contradictions in transcripts.

[0005] The present application provides a method for detecting contradictions in a transcript, including:

[0006] Obtain the target transcript text to be detected;

[0007] Based on the pre-constructed transcript contradiction detection point set, the target transcript text is subjected to intent recognition, and the key elements and correlation relationships of the case under the intent are extracted to form the contradiction problem to be detected;

[0008] A first prompt instruction prompt is constructed using the contradiction problem to be detected, and is input into a preset large language model to obtain a response result for the contradiction problem to be detected output by the large language model, and a contradiction detection result corresponding to the target transcript text is determined according to the response result.

[0009] In a possible implementation, the method further includes:

[0010] Acquire sample transcripts; the sample transcripts include two transcripts with conflicting points; the conflicting points belong to a pre-constructed transcript conflict detection point set;

[0011] The sample transcript text is combined with the second prompt instruction to perform conflict point recognition training on the initial large language model to obtain the preset large language model.

[0012] In a possible implementation, the target transcript text is subjected to intent recognition based on the pre-constructed transcript contradiction detection point set, and key case elements and association relationships under the intent are extracted, including:

[0013] A third prompt instruction is constructed using a pre-constructed set of transcript contradiction detection points and the target transcript text, and the preset large language model is input to obtain the intention recognition result for the target transcript text output by the large language model, as well as the key elements and correlation relationships of the case under the identified intention.

[0014] In a possible implementation, the step of constructing a first prompt instruction prompt using the contradiction problem to be detected includes:

[0015] Sampling the problem set where the conflicting problem to be detected is located to obtain sampled problems;

[0016] Analyze the sampling problem using the preset large language model to obtain an analysis result of the reasoning process;

[0017] A first prompt instruction prompt is constructed using the sampling problem, the reasoning process analysis result and the contradictory problem to be detected.

[0018] In a possible implementation, the using the preset large language model to analyze the sampling problem to obtain the reasoning process analysis result includes:

[0019] The fourth prompt instruction is constructed using the sampled question, and is input into the preset large language model for analysis to obtain the answer result and reasoning process for the sampled question output by the large language model.

[0020] In one possible implementation, the reasoning process includes a question-answer pair text regarding the sampled question.

[0021] In one possible implementation, the pre-constructed set of transcript contradiction detection points includes at least one of personnel information contradiction, personnel affiliation company contradiction, personnel confession amount involved contradiction, company information contradiction, and company affiliation company contradiction.

[0022] The present application also provides a device for detecting inconsistencies in transcripts, including:

[0023] A first acquisition unit, used to acquire a target transcript text to be detected;

[0024] An identification unit, configured to perform intent recognition on the target transcript text based on a pre-constructed transcript contradiction detection point set, and extract key case elements and association relationships under the intent to form a contradiction problem to be detected;

[0025] The input unit is used to construct a first prompt instruction prompt using the contradiction problem to be detected, input it into a preset large language model, obtain a response result for the contradiction problem to be detected output by the large language model, and determine the contradiction detection result corresponding to the target transcript text according to the response result.

[0026] In a possible implementation manner, the device further includes:

[0027] A second acquisition unit is used to acquire a sample transcript text; the sample transcript text includes two transcript texts with conflicting points; the conflicting points belong to a pre-constructed transcript conflict detection point set;

[0028] The training unit is used to use the sample transcript text in combination with the second prompt instruction to perform conflict point recognition training on the initial large language model to obtain the preset large language model.

[0029] In a possible implementation, the identification unit is specifically used to:

[0030] A third prompt instruction is constructed using a pre-constructed set of transcript contradiction detection points and the target transcript text, and the preset large language model is input to obtain the intention recognition result for the target transcript text output by the large language model, as well as the key elements and correlation relationships of the case under the identified intention.

[0031] In a possible implementation, the input unit includes:

[0032] A sampling subunit, used for sampling the problem set where the conflicting problem to be detected is located, to obtain a sampled problem;

[0033] An analysis subunit, used to analyze the sampling problem using the preset large language model to obtain an analysis result of the reasoning process;

[0034] The construction subunit is used to construct a first prompt instruction prompt by using the sampling problem, the reasoning process analysis result and the contradictory problem to be detected.

[0035] In a possible implementation, the analysis subunit is specifically used for:

[0036] The fourth prompt instruction is constructed using the sampled question, and is input into the preset large language model for analysis to obtain the answer result and reasoning process for the sampled question output by the large language model.

[0037] In one possible implementation, the reasoning process includes a question-answer pair text regarding the sampled question.

[0038] In one possible implementation, the pre-constructed set of transcript contradiction detection points includes at least one of personnel information contradiction, personnel affiliation company contradiction, personnel confession amount involved contradiction, company information contradiction, and company affiliation company contradiction.

[0039] The embodiment of the present application also provides a transcript contradiction detection device, including: a processor, a memory, and a system bus;

[0040] The processor and the memory are connected via the system bus;

[0041] The memory is used to store one or more programs, and the one or more programs include instructions. When the instructions are executed by the processor, the processor executes any one of the implementation methods of the above-mentioned transcript contradiction detection method.

[0042] An embodiment of the present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any one of the implementations of the above-mentioned transcript contradiction detection method.

[0043] The embodiment of the present application also provides a computer program product. When the computer program product is run on a terminal device, the terminal device executes any one of the implementations of the above-mentioned transcript contradiction detection method.

[0044] The embodiments of the present application provide a transcript contradiction detection method, apparatus, storage medium and device, which first obtain a target transcript text to be detected; then, based on a pre-constructed transcript contradiction detection point set, perform intent recognition on the target transcript text, and extract key case elements and association relationships under the intent to form a contradiction problem to be detected; then, use the contradiction problem to be detected to construct a first prompt instruction prompt, input it into a preset large language model, obtain a reply result for the contradiction problem to be detected output by the large language model, and determine the contradiction detection result corresponding to the target transcript text according to the reply result.

[0045] It can be seen that since this application utilizes a preset large language model to detect contradictions in the target transcript text based on a constructed set of transcript contradiction detection points, it not only breaks through the strong dependence on the knowledge graph construction effect and solves the limitations of the detection scenario, but also improves the efficiency and accuracy of transcript contradiction detection, thereby improving the detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 A flowchart of a method for detecting inconsistencies in transcripts provided in an embodiment of the present application;

[0048] Figure 2 A schematic diagram of the composition of a transcript contradiction detection device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] With the development of science and technology, a series of technical tools have been developed and used, which has enabled some people to use mobile phones, landlines, the Internet and other communication tools and modern technologies to carry out contactless telecommunications network-related fraud, causing great losses to the masses. Such telecommunications network-related fraud cases often involve many people. During the interrogation and questioning process, a large number of transcripts are usually generated. The key information in the transcripts, such as personnel information, company information, and information on the amount involved, is complicated, which often brings a large workload and challenges to the staff in analyzing and handling the case. Therefore, how to quickly and accurately detect the contradictions in the information in the transcript is of great significance to the staff in improving the efficiency of case handling.

[0050] Existing transcript contradiction detection methods generally use the method of constructing a knowledge graph to detect contradictions in transcript content. Specifically, first, a knowledge graph related to telecommunications network cases is constructed, that is, information extraction is carried out based on the existing transcript text, and after extracting entities, attributes, and the relationships between entities, an ontological knowledge expression is formed. Then, relying on business experts, knowledge fusion and knowledge processing are carried out to finally form a knowledge graph related to telecommunications network cases. Secondly, through deductive knowledge graph reasoning and inductive knowledge graph reasoning, the knowledge contradictions in the transcript are analyzed.

[0051] It can be seen that the existing contradiction detection methods do not consider the effect of information extraction in various case business scenarios, and strongly rely on the construction effect of knowledge graph construction. In the implementation process, due to the complexity of the knowledge graph construction process, it will be difficult to implement it, which greatly reduces the accuracy of transcript contradiction detection.

[0052] To address the above-mentioned defects, the present application provides a transcript contradiction detection method, which first obtains a target transcript text to be detected; then, based on a pre-constructed transcript contradiction detection point set, performs intent recognition on the target transcript text, and extracts key case elements and association relationships under the intent to form a contradiction problem to be detected; then, a first prompt instruction prompt is constructed using the contradiction problem to be detected, and is input into a preset large language model to obtain a response result for the contradiction problem to be detected output by the large language model, and the contradiction detection result corresponding to the target transcript text is determined according to the response result.

[0053] It can be seen that since this application utilizes a preset large language model to detect contradictions in the target transcript text based on a constructed set of transcript contradiction detection points, it not only breaks through the strong dependence on the knowledge graph construction effect and solves the limitations of the detection scenario, but also improves the efficiency and accuracy of transcript contradiction detection, thereby improving the detection effect.

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0055] First embodiment

[0056] See also Figure 1 , is a flow chart of a method for detecting contradictions in transcripts provided in this embodiment, the method comprising the following steps:

[0057] S101: Obtain the target transcript text to be detected.

[0058] In this embodiment, any transcript text that needs to be detected for contradictions in the intelligent interactive software or device such as the large language model (LLM) preset by the user is defined as the target transcript text to be detected, and this embodiment does not limit the language type of the target transcript text, for example, the target transcript text can be a Chinese transcript text or an English transcript text, etc. This embodiment does not limit the length of the target transcript text, that is, the target transcript text can be a sentence text (i.e., a collection of words) or a chapter text (i.e., a collection of sentences), etc. And this embodiment does not limit the content of the target transcript text, for example, the target transcript text can be a text question-answer pair (QApairs) formed by the staff of the judicial department during the interrogation and questioning of the persons involved in the case, that is, it can include a pair of question (Question) text and answer (Answer) text, etc. In addition, this embodiment does not limit the method of obtaining the target transcript text, for example, the target transcript text can be the audio recorded by the staff of the judicial department during the interrogation and questioning of the persons involved in the case, and the transcript text is recognized by speech recognition and other methods.

[0059] Among them, the target transcript text to be tested refers to the transcript text generated by the staff of the judicial department during the interrogation and questioning of the persons involved in the case. Each transcript complies with relevant specifications, and the content of the transcript uses the original words of the person under investigation and is faithful to the actual situation of the conversation. The conversation information includes the time and place of the interrogation or questioning, the name of the investigator, the notification of rights and obligations, simultaneous audio and video recording, the time, place, circumstances, results of the facts confessed, and other relevant facts.

[0060] For example, the target transcript text to be detected can be a text question-answer pair (QApairs) formed by judicial staff during the interrogation of the suspect "Zhang San". The following is an example:

[0061] Question: We are police officers from the case review center of the XX Branch of the XX City. We are now interrogating you in accordance with the law. You should answer truthfully and must not conceal anything or make false statements. Otherwise, you will bear corresponding legal responsibilities. You have the right to refuse to answer questions that are not related to the case. Do you understand?

[0062] Answer: I heard it clearly.

[0063] Q: Did you know that the entire interrogation will be recorded in audio and video?

[0064] A: I understand.

[0065] Q: What is your basic situation?

[0066] Answer: My name is Zhang San, male, Han nationality, born on December 6, 1996, with junior high school education, unmarried, unemployed, ID number is **, registered address is No. xx, xx Road, xx District, xx City, xx Province, current address: No fixed place of residence.

[0067] Q: When and why were you criminally detained?

[0068] Answer: I was criminally detained on suspicion of cyber fraud.

[0069] ask:……

[0070] S102: Based on the pre-constructed transcript contradiction detection point set, the target transcript text is subjected to intent recognition, and the key elements and correlation relationships of the case under the intent are extracted to form the contradiction problem to be detected.

[0071] In this embodiment, after the target transcript text to be detected is obtained through step S101, the target transcript text can be further identified for intent based on a pre-constructed transcript contradiction detection point set, and the key elements and correlation relationships of the case under the intent can be extracted to form the contradiction problem to be detected, so as to execute the subsequent step S103.

[0072] Among them, this application does not limit the specific content and specific construction method of the pre-constructed transcript contradiction detection point set, and can extract relevant knowledge contradictions according to actual conditions and experience values ​​for construction. An optional implementation method is that the pre-constructed transcript contradiction detection point set can include but is not limited to at least one of personnel information contradictions, personnel affiliation company contradictions, personnel confession involved amount contradictions, company information contradictions, and company affiliation company contradictions.

[0073] Specifically, an optional implementation method is that after obtaining the target transcript text to be detected, a pre-constructed transcript contradiction detection point set and the target transcript text can be further used to construct a third prompt instruction (prompt), and input it into the preset large language model LLM to obtain the intention recognition result for the target transcript text output by the large language model LLM, as well as the key elements and correlation relationships of the case under the identified intention.

[0074] Among them, the preset large language model LLM is obtained after a small sample training based on a small number of existing typical transcripts, and the trained large language model LLM is used to learn the contradiction points in the scenario of telecom fraud cases related to telecommunications networks. An optional implementation method is that the construction process of the preset large language model LLM can specifically include: first obtaining a sample transcript; wherein the sample transcript includes two transcripts with contradictions; wherein the contradictions belong to a pre-constructed transcript contradiction detection point set, and then using the sample transcript combined with the second prompt instruction to perform contradiction point recognition training on the initial large language model to obtain the preset large language model.

[0075] In this implementation, in order to build a large language model with the ability to identify contradictions in telecom network-related electronic fraud cases, a lot of preparatory work needs to be done in advance. First, at least two transcripts that have been manually confirmed to have contradictions need to be collected as sample transcript texts to form model training data. For example, taking the contradiction of personnel information as an example, two transcripts that have been manually confirmed to have the contradiction can be selected, such as the criminal interrogation transcript of the interrogated person Zhang San, in which Zhang San confessed that "my superior is Li Si"; and the criminal interrogation transcript of the interrogated person Li Si, in which Li Si confessed that "there are 3 people in my management team, they are Wang Wu, Gao Liu, and Liu Qi". Then, according to these sample transcript texts and the contradictions contained therein, combined with the second prompt instruction, the initial large language model can be trained to identify contradictions, and then a preset large language model can be generated. For example, based on the above example, the second prompt instruction that can be input into the large language model for training is: You are a transcript analysis expert, and there is a contradiction that Zhang San’s superior is Li Si. The reason is that in Zhang San’s criminal interrogation transcript, "my superior is Li Si", and in Li Si’s criminal interrogation transcript, there is no Zhang San in the confession team.

[0076] In this way, after training a preset large language model with the ability to identify contradictions in transcripts, the third prompt instruction (prompt) constructed using the pre-constructed transcript contradiction detection point set and the target transcript text can be input into the preset large language model to identify the various case business intention scenarios contained in the target transcript text, and extract the key elements and correlation relationships of the case under each intention to form the contradiction problem to be detected.

[0077] Specifically, when constructing the third prompt instruction (prompt) using the pre-constructed transcript contradiction detection point set and the target transcript text, it is necessary to first use the prompt to guide the large language model to identify the specified intent text in the target transcript text, including but not limited to personnel information intent, address information intent, position information intent, company information intent, amount involved in the case intent, etc. The specific intent content can be adjusted based on business development needs, and then use the prompt to extract the key elements and correlation relationships in various case business intent scenarios from the identified specified intent text.

[0078] Example: Based on the above example, an example of constructing the third prompt instruction (prompt) using the pre-constructed transcript contradiction detection point set and the target transcript text is: "You are a transcript analysis expert. Please learn the following requirements and examples first, and think carefully before answering the questions:--Please identify the speaker's intention and extract the following elements of the question;--The speaker's intention can only be within the following range, and the text cannot be changed: personnel information, address information, position information, company information, and the amount involved in the case;--If the intended text is inconsistent with the above range, directly return to "other";--The personnel information intention element can be: Name, ID type, ID number, etc.; Example 1: Please query Zhang San’s information, the return format is: Intent: Personnel information intent, name: Wang Hua;... ", when the target transcript text to be detected is input to the large language model, the corresponding result that can be output according to the above prompt is: "Intent: Personnel information intent, name: Zhang San; ID type: ID card; ID card number is **;"; If the intent is not within the range, the corresponding result that can be output by the above prompt is: "Others", the intent and related elements set in this example can be shown in Table 1 below, including some intents and key elements and associations under the intent:

[0079]

[0080] Table 1

[0081] Furthermore, after identifying the various case business intention scenarios contained in the target transcript text and extracting the key case elements and correlation relationships under each intention (as shown in Table 1), the identified correlation relationships can be further processed using existing or future data processing methods to form contradiction problems to be detected, so as to execute the subsequent step S103.

[0082] Among them, an optional implementation method is to add "whether there is a contradiction" after the identified association relationship to form a contradiction problem to be detected, as shown in Table 2 below, and the present application does not limit the number of contradiction problems to be detected.

[0083] Relationship Contradictory issues Zhang San's superior is Li Si Is there a contradiction that Zhang San's superior is Li Si? Zhang San's leader is Liu Ba Is there a contradiction that Zhang San's leader is Liu Ba? Liu Ba's position is the company's legal representative Is there a contradiction in Liu Ba's position as the company's legal person? Liu Ba's position is department manager Is there a contradiction in Liu Ba's position being a department manager? …… ……

[0084] Table 2

[0085] S103: construct a first prompt instruction prompt using the contradiction problem to be detected, input it into a preset large language model, obtain a reply result for the contradiction problem to be detected output by the large language model, and determine the contradiction detection result corresponding to the target transcript text according to the reply result.

[0086] In this embodiment, after forming the contradictory problem to be detected (for example, as shown in Table 2) through step S102, the contradictory problem to be detected can be further used to construct a first prompt instruction prompt, and it can be input into the preset large language model LLM to construct an automated large language model thinking chain (Auto-CoT (Chain-of-Thought)), so as to use the large language model LLM to output the contradiction judgment result and reasoning logic process for the contradictory problem to be detected, and then determine the final contradiction detection result corresponding to the target transcript text according to the reply result, so that the large language model thinking chain can be constructed, and based on the formed contradictory problem to be detected, the contradictions of the target transcript text can be more accurately identified.

[0087] Specifically, in an optional implementation manner, the specific implementation process of "using the contradiction problem to be detected to construct the first prompt instruction prompt" in step S103 may include the following steps AC:

[0088] Step A: Sample the problem set where the contradictory problem to be detected is located to obtain the sampled problem.

[0089] In this implementation, after obtaining all the contradictory problems to be detected (for example, as shown in Table 2), one contradictory problem to be detected can be selected as the target contradictory problem, and further, the text semantic similarity algorithm can be used to perform collective clustering on all the contradictory problems to be detected in the problem set where the contradictory problem to be detected is located (the number of clusters is not limited, for example, the number of clusters can be set to 3), or, the remaining contradictory problems to be detected in the problem set where the contradictory problem to be detected is located except the target contradictory problem can be collectively clustered, and then one problem closest to the center point is selected from each clustered class, thereby completing problem sampling and obtaining the sampled problem for executing the subsequent step B.

[0090] For example: take the target contradiction problem as "Is there a contradiction if Zhang San's superior is Li Si" in Table 2 as an example, assuming that the problem set where the contradiction problem to be detected is located includes the 4 problems shown in Table 2, then the semantic similarity clustering of the remaining contradiction problems to be detected except the target contradiction problem in the problem set (that is, except "Is there a contradiction if Zhang San's superior is Li Si") can be clustered into 3 categories (that is, each contradiction problem to be detected is one category), and then these 3 questions are sampling problems, that is, "Is there a contradiction if Zhang San's leader is Liu Ba", "Is there a contradiction if Liu Ba's position is the legal person of the company", and "Is there a contradiction if Liu Ba's position is the department manager" are sampled problems.

[0091] Step B: Use the preset large language model to analyze the sampling problem and obtain the analysis results of the reasoning process.

[0092] In this implementation, after obtaining the sampling problem through step A, the preset large language model can be further used as a tool to generate the reasoning process, and the fourth prompt instruction can be constructed using the sampling problem, and the preset large language model is input for analysis to obtain the reasoning process analysis results for the sampling problem output by the large language model, including but not limited to the reply result and the reasoning process. Among them, this application does not limit the content and specific form of the reasoning process, which may include but not be limited to the question and answer pair (QApair) text about the sampling problem, and does not limit the steps of the reasoning process, such as it can be set to any value not exceeding 5 steps.

[0093] For example: Taking the question "Is there a contradiction in Zhang San's leader being Liu Ba?" as an example, the input and corresponding output of the fourth prompt instruction constructed using this sampling question can be: "Based on the content of the transcript you have mastered, please analyze whether there is a contradiction in Zhang San's leader being Liu Ba, and give the reasoning process. At the same time, the reasoning process is required to be no more than 5 steps: - The return format is: contradiction, reasoning logic: 1. Zhang San confessed in the criminal interrogation transcript that his leader was Liu Wu; 2. Liu Ba confessed in the criminal interrogation transcript that his company position was employee;". From this example, we can see that the QA pair generated by the large language model in the reasoning process is: "Q: Is there a contradiction in Zhang San's leader being Liu Ba? A: There is a contradiction in Zhang San's leader being Liu Ba. In Zhang San's criminal interrogation record, he confessed that his leader was Liu Wu, which means that Liu Wu played a leadership role in Zhang San's company or organization. In Liu Ba's criminal interrogation record, he confessed that his position in the company was an employee, which is inconsistent with the first piece of information because an employee cannot be a leader at the same time. The two identities are mutually exclusive in a conventional organizational structure. Based on the existing information, we can determine that there is a contradiction in Liu Wu's identity description."

[0094] Step C: construct a first prompt using the sampling problem, the analysis results of the reasoning process and the contradiction problem to be detected.

[0095] In this implementation, after obtaining the sampling problem through step A and the reasoning process analysis result through step B, the sampling problem, the reasoning process analysis result and the contradiction problem to be detected (i.e., the target contradiction problem mentioned in step A) can be further used to constitute the above-mentioned standard prompt added to the context of in-context learning (In-Context-Learning) to obtain an improved prompt as the first prompt instruction prompt. The prompt can require the preset large language model to display the step-by-step reasoning steps in the output while outputting the answer to the question, so as to improve the accuracy of the output results of the large language model, and the output of the intermediate steps also makes it easier for users to understand the thinking process of the large language model, thereby improving the interpretability of the reasoning of the large language model.

[0096] Example: Based on the above example, still taking the target contradiction problem that the large language model needs to answer as "Is there a contradiction in that Zhang San's superior is Li Si" in Table 2 as an example, an example of the first prompt instruction (prompt) constructed using the sampling problem, the reasoning process analysis results and the target contradiction problem can be: "You are a transcript analysis expert. When answering the question "Is there a contradiction in that Zhang San's leader is Liu Ba", give the result of "Contradiction, reasoning logic: 1. Zhang San's criminal interrogation record stated that his leader was Liu Wu; 2. Liu Ba's criminal interrogation record stated that his company position was employee;" .... Now please refer to the above question and reasoning, and answer "Is there a contradiction in that Zhang San's superior is Li Si", requiring the reasoning logic in the output result to give the analysis basis of the original transcript text." According to the prompt, an example of the output result of the large language model can be: "Contradiction, reasoning logic: 1. Zhang San confessed in the criminal interrogation record that his leader was Liu Wu; 2. Li Si confessed in the criminal interrogation record that his company position was employee;" and at the same time, the associated text fragments of Zhang San's criminal record associated with "Zhang San's criminal interrogation record confessed that his leader was Liu Wu" and the associated text fragments of Li Si's criminal record associated with "Li Si's criminal interrogation record confessed that his company position was employee" are given.

[0097] It should be noted that before executing the three steps of problem sampling, sampling problem analysis, and context learning paradigm prompt mentioned in the above steps AC, an automated large language model thinking chain is constructed to output the analysis results for each contradiction problem to be detected (i.e., the target contradiction problem), each contradiction problem to be detected can also be fine-tuned in a targeted manner, and the specific fine-tuning method and content are not limited. In addition, it should be noted that the problem sampling can choose the similarity sampling method, and the number of problem clusters depends on the number of previous texts required by the context learning paradigm prompt. At the same time, although the present embodiment mentions in the above example that a tracing method can be adopted to output the text fragment of the reasoning record to solve the hallucination problem, in actual application, from the perspective of case business, the "everything should be available" principle may be proposed, that is, the full amount of text fragments are required to be output, so there may still be probabilistic hallucinations, so the contradiction detection results obtained by this application are usually used as reference detection results for staff, rather than as the final qualitative detection results.

[0098] In summary, the present embodiment provides a transcript contradiction detection method, which first obtains a target transcript text to be detected; then, based on a pre-constructed transcript contradiction detection point set, performs intent recognition on the target transcript text, and extracts key case elements and association relationships under the intent to form a contradiction problem to be detected; then, a first prompt instruction prompt is constructed using the contradiction problem to be detected, and is input into a preset large language model to obtain a reply result for the contradiction problem to be detected output by the large language model, and the contradiction detection result corresponding to the target transcript text is determined according to the reply result.

[0099] It can be seen that since this application utilizes a preset large language model to detect contradictions in the target transcript text based on a constructed set of transcript contradiction detection points, it not only breaks through the strong dependence on the knowledge graph construction effect and solves the limitations of the detection scenario, but also improves the efficiency and accuracy of transcript contradiction detection, thereby improving the detection effect.

[0100] Second embodiment

[0101] This embodiment will introduce a device for detecting contradictions in transcripts. For related content, please refer to the above method embodiment.

[0102] See also Figure 2 , is a schematic diagram of the composition of a transcript contradiction detection device provided in this embodiment, the device 200 includes:

[0103] The first acquisition unit 201 is used to acquire the target transcript text to be detected;

[0104] The identification unit 202 is used to perform intent recognition on the target transcript text based on the pre-constructed transcript contradiction detection point set, and extract the key elements and association relationships of the case under the intent to form the contradiction problem to be detected;

[0105] The input unit 203 is used to construct a first prompt instruction prompt using the contradiction problem to be detected, input it into a preset large language model, obtain a response result for the contradiction problem to be detected output by the large language model, and determine the contradiction detection result corresponding to the target transcript text according to the response result.

[0106] In an implementation of this embodiment, the device further includes:

[0107] A second acquisition unit is used to acquire a sample transcript text; the sample transcript text includes two transcript texts with conflicting points; the conflicting points belong to a pre-constructed transcript conflict detection point set;

[0108] The training unit is used to use the sample transcript text in combination with the second prompt instruction to perform conflict point recognition training on the initial large language model to obtain the preset large language model.

[0109] In an implementation of this embodiment, the identification unit 202 is specifically configured to:

[0110] A third prompt instruction is constructed using a pre-constructed set of transcript contradiction detection points and the target transcript text, and the preset large language model is input to obtain the intention recognition result for the target transcript text output by the large language model, as well as the key elements and correlation relationships of the case under the identified intention.

[0111] In an implementation of this embodiment, the input unit 203 includes:

[0112] A sampling subunit, used for sampling the problem set where the conflicting problem to be detected is located, to obtain a sampled problem;

[0113] An analysis subunit, used to analyze the sampling problem using the preset large language model to obtain an analysis result of the reasoning process;

[0114] The construction subunit is used to construct a first prompt instruction prompt by using the sampling problem, the reasoning process analysis result and the contradictory problem to be detected.

[0115] In one implementation of this embodiment, the analysis subunit is specifically used for:

[0116] The fourth prompt instruction is constructed using the sampled question, and is input into the preset large language model for analysis to obtain the answer result and reasoning process for the sampled question output by the large language model.

[0117] In one implementation of this embodiment, the reasoning process includes a question-answer pair text about the sampled question.

[0118] In one implementation of this embodiment, the pre-constructed set of transcript contradiction detection points includes at least one of personnel information contradictions, personnel affiliation company contradictions, personnel confession of the amount involved in the case contradictions, company information contradictions, and company affiliation company contradictions.

[0119] Furthermore, the embodiment of the present application also provides a transcript contradiction detection device, including: a processor, a memory, and a system bus;

[0120] The processor and the memory are connected via the system bus;

[0121] The memory is used to store one or more programs, and the one or more programs include instructions. When the instructions are executed by the processor, the processor executes any one of the implementation methods of the above-mentioned transcript contradiction detection method.

[0122] Furthermore, an embodiment of the present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any one of the implementation methods of the above-mentioned transcript contradiction detection method.

[0123] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any one of the implementation methods of the above-mentioned transcript contradiction detection method.

[0124] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0125] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0126] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0127] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting contradictions in transcripts, characterized in that: include: Obtain the target transcript text to be detected; Based on the pre-constructed transcript contradiction detection point set, the target transcript text is subjected to intent recognition, and the key elements and correlation relationships of the case under the intent are extracted to form the contradiction problem to be detected; A first prompt instruction prompt is constructed using the contradiction problem to be detected, and is input into a preset large language model to obtain a response result for the contradiction problem to be detected output by the large language model, and a contradiction detection result corresponding to the target transcript text is determined according to the response result.

2. The method according to claim 1, characterized in that: The method further comprises: Acquire sample transcripts; the sample transcripts include two transcripts with conflicting points; the conflicting points belong to a pre-constructed transcript conflict detection point set; The sample transcript text is combined with the second prompt instruction to perform conflict point recognition training on the initial large language model to obtain the preset large language model.

3. The method according to claim 1, characterized in that The method of performing intent recognition on the target transcript text based on the pre-constructed transcript contradiction detection point set and extracting the key elements and correlation relationships of the case under the intent includes: A third prompt instruction is constructed using a pre-constructed set of transcript contradiction detection points and the target transcript text, and the preset large language model is input to obtain the intention recognition result for the target transcript text output by the large language model, as well as the key elements and correlation relationships of the case under the identified intention.

4. The method according to claim 1, characterized in that: The step of constructing a first prompt instruction prompt by using the contradiction problem to be detected includes: Sampling the problem set where the conflicting problem to be detected is located to obtain sampled problems; Analyze the sampling problem using the preset large language model to obtain an analysis result of the reasoning process; A first prompt instruction prompt is constructed using the sampling problem, the reasoning process analysis result and the contradictory problem to be detected.

5. The method according to claim 4, characterized in that The method of analyzing the sampling problem by using the preset large language model to obtain the reasoning process analysis result includes: The fourth prompt instruction is constructed using the sampled question, and is input into the preset large language model for analysis to obtain the answer result and reasoning process for the sampled question output by the large language model.

6. The method according to claim 5, characterized in that The reasoning process includes question-answer pair texts about the sampled question.

7. The method according to any one of claims 1 to 6, characterized in that: The pre-constructed transcript contradiction detection point set includes at least one of personnel information contradiction, personnel affiliation company contradiction, personnel confession involved amount contradiction, company information contradiction, company affiliation company contradiction.

8. A device for detecting contradictions in written records, characterized in that: include: A first acquisition unit, used to acquire a target transcript text to be detected; An identification unit, configured to perform intent recognition on the target transcript text based on a pre-constructed transcript contradiction detection point set, and extract key case elements and association relationships under the intent to form a contradiction problem to be detected; The input unit is used to construct a first prompt instruction prompt using the contradiction problem to be detected, input it into a preset large language model, obtain a response result for the contradiction problem to be detected output by the large language model, and determine the contradiction detection result corresponding to the target transcript text according to the response result.

9. A device for detecting contradictions in written records, characterized in that: include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the method according to any one of claims 1 to 7.

Images