Bad thinking monitoring method and device based on large language model

Abstract

The invention provides a bad thinking monitoring method and device based on a large language model, and the method is applied to a conference system, and comprises the steps: carrying out the monitoring of group discussion, and obtaining the background information of a person participating in the group discussion, and a session sequence generated by the group discussion; inputting the session sequence and the background information into a bad thinking monitoring model to obtain a thinking monitoring result output by the bad thinking monitoring model; the bad thinking monitoring model is obtained by applying sample background information of personnel participating in sample group discussion, a sample session sequence generated by the sample group discussion and bad thinking label training of the sample group discussion on the basis of a large language model, and overcomes the defect that a traditional scheme cannot monitor group discussion and cannot monitor the group discussion. According to the method and the system, the defect of bad thinking is detected, so that the bad thinking in a multi-person conference is monitored, reference and basis are provided for solving the bad thinking of a group, the comprehensiveness and accuracy of decision can be improved, and the risk of decision failure is avoided.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and device for monitoring bad thinking based on a large language model. Background Art

[0002] During the discussion, due to the lack of ability of participants to propose high-quality solutions, different interests inside and outside the meeting, and meeting time constraints, meeting decisions are often made by the highest decision-maker of the meeting or simply by the principle of minority obeying the majority, thus showing the group's unhealthy thinking such as dependent thinking, rigid thinking, and divergent thinking. In addition, the solutions proposed by this decision-making method are often based on partial information, so there is a greater risk of decision failure.

[0003] With the continuous development of artificial intelligence technology, natural language processing technology has made significant progress in text processing. However, traditional natural language processing technology is mainly concentrated in the fields of text summarization, topic extraction, machine translation, etc., and its processing capabilities for multi-person multi-round conversations in meetings are relatively limited. In addition, although traditional natural language processing technology can assist in understanding and analyzing the content of meetings to a certain extent, it cannot solve the fundamental problem in multi-person meetings, that is, it cannot effectively monitor and identify the bad thinking generated by the group during the decision-making process. Summary of the invention

[0004] The present invention provides a method and device for monitoring bad thinking based on a large language model, which is used to solve the defects of the prior art that group discussions cannot be monitored as a whole and bad thinking in group discussions cannot be detected, and realizes the monitoring of bad thinking in multi-person meetings.

[0005] The present invention provides a method for monitoring bad thinking based on a large language model, which is applied to a conference system. The method comprises:

[0006] Monitor the group discussion and obtain background information of participants in the group discussion and a conversation sequence generated by the group discussion, wherein the conversation sequence includes speeches and operations of the participants in the group discussion and their corresponding times;

[0007] Inputting the conversation sequence and the background information into an unhealthy thought monitoring model to obtain a thought monitoring result output by the unhealthy thought monitoring model;

[0008] The bad thinking monitoring model is trained on the basis of a large language model using sample background information of people participating in a sample group discussion, sample conversation sequences generated by the sample group discussion, and bad thinking labels of the sample group discussion.

[0009] According to a method for monitoring bad thoughts based on a large language model provided by the present invention, the bad thought monitoring model includes a large language model and a target language model with the same structure, and the target language model is constructed based on the large language model;

[0010] The step of inputting the conversation sequence and the background information into the bad thinking monitoring model to obtain the thinking monitoring result output by the bad thinking monitoring model includes:

[0011] Inputting the background information into the target language model in the bad thinking monitoring model to obtain the background features output by the target language model;

[0012] The background features and the conversation sequence are input into a large language model in the bad thought monitoring model to obtain a thought monitoring result output by the large language model.

[0013] According to a method for monitoring bad thoughts based on a large language model provided by the present invention, the bad thought monitoring model further comprises a feature fusion layer; the feature fusion layer connects the target language model and the large language model;

[0014] The step of inputting the background features and the conversation sequence into the large language model in the bad thought monitoring model to obtain the thought monitoring result output by the large language model includes:

[0015] Inputting the background features and the conversation sequence into the feature fusion layer to obtain fusion features output by the feature fusion layer;

[0016] The fusion feature is input into the large language model to obtain the thought monitoring result output by the large language model.

[0017] According to a method for monitoring bad thinking based on a large language model provided by the present invention, the bad thinking monitoring model is trained based on the following steps:

[0018] Based on the large language model, construct a twin model, wherein the twin model includes a first language model and a second language model with the same parameters and structure;

[0019] Inputting the sample background information into the first language model to obtain sample background features output by the first language model, inputting the sample background features and the sample conversation sequence into the second language model to obtain a predictive thinking monitoring result output by the second language model;

[0020] Based on the predicted thought monitoring result and the bad thought label, the first language model in the twin model is trained to obtain a bad thought monitoring model.

[0021] According to a method for monitoring bad thoughts based on a large language model provided by the present invention, the first language model includes a plurality of first feature conversion layers, and the second language model includes a plurality of second feature conversion layers;

[0022] The twin model further includes a plurality of initial feature fusion layers, the first language model and the second language model are connected via the plurality of initial feature fusion layers, and the first feature conversion layer, the second feature conversion layer and the initial feature fusion layer have the same number of layers;

[0023] Any initial feature fusion layer connects a first feature conversion layer of a corresponding layer in the first language model and a second feature conversion layer of an upper layer of a corresponding layer in the second language model;

[0024] Any initial feature fusion layer is used to fuse the output features of the first feature conversion layer of the corresponding layer in the first language model with the output features of the second feature conversion layer of the previous layer of the corresponding layer in the second language model, and use the fused features as the input features of the second feature conversion layer of the corresponding layer in the second language model.

[0025] According to a method for monitoring bad thoughts based on a large language model provided by the present invention, based on the predicted thought monitoring result and the bad thought label, the first language model in the twin model is trained to obtain a bad thought monitoring model, including:

[0026] Based on the predicted thought monitoring result and the bad thought label, the first language model and each initial feature fusion layer in the twin model are trained to obtain a target language model and multiple feature fusion layers;

[0027] Based on the large language model, the target language model and each feature fusion layer, a bad thinking monitoring model is constructed.

[0028] According to a method for monitoring bad thoughts based on a large language model provided by the present invention, the conversation sequence and the background information are input into the bad thought monitoring model to obtain a thought monitoring result output by the bad thought monitoring model, and then the method further includes:

[0029] When the thought monitoring result indicates that bad thoughts occur in the group discussion, obtaining bad thought warning prompts and control tools, so that the persons participating in the group discussion can control the group discussion based on the bad thought warning prompts and control tools;

[0030] The bad thinking warning prompt and regulation tool are generated by the bad thinking monitoring model based on the thinking monitoring result.

[0031] The present invention also provides a device for monitoring bad thinking based on a large language model, which is applied to a conference system. The device comprises:

[0032] An acquisition unit, used to monitor the group discussion and acquire background information of participants in the group discussion and a conversation sequence generated by the group discussion, wherein the conversation sequence includes speeches and operations of participants in the group discussion and their corresponding times;

[0033] A monitoring unit, used for inputting the conversation sequence and the background information into an unhealthy thinking monitoring model to obtain a thinking monitoring result output by the unhealthy thinking monitoring model;

[0034] The bad thinking monitoring model is trained on the basis of a large language model using sample background information of people participating in a sample group discussion, sample conversation sequences generated by the sample group discussion, and bad thinking labels of the sample group discussion.

[0035] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for monitoring bad thinking based on a large language model as described above is implemented.

[0036] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for monitoring bad thinking based on a large language model.

[0037] The present invention provides a method and device for monitoring bad thinking based on a large language model, which monitors group discussions, obtains background information of persons participating in group discussions, and a conversation sequence generated by the group discussions, inputs the conversation sequence and background information into the bad thinking monitoring model, and obtains a thinking monitoring result output by the bad thinking monitoring model; the bad thinking monitoring model is obtained by training on the basis of a large language model using sample background information of persons participating in sample group discussions, sample conversation sequences generated by sample group discussions, and bad thinking labels of sample group discussions, thereby overcoming the defects of traditional solutions that cannot monitor group discussions as a whole and detect bad thinking therein, realizes the monitoring of bad thinking in multi-person meetings, and provides a reference and basis for solving the bad thinking of the group, which can greatly improve the comprehensiveness and accuracy of decision-making and avoid the risk of decision-making failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 It is a flow chart of the method for monitoring bad thinking based on a large language model provided by the present invention;

[0040] Figure 2 It is a flow chart of the group discussion monitoring and bad thinking detection process provided by the present invention;

[0041] Figure 3 It is a structural schematic diagram of the bad thinking monitoring model provided by the present invention;

[0042] Figure 4 It is a schematic diagram of the bad thinking monitoring model training provided by the present invention;

[0043] Figure 5 It is a flowchart of the anonymous discussion process provided by the present invention;

[0044] Figure 6 It is a flowchart of the nominal group discussion provided by the present invention;

[0045] Figure 7 It is a flow chart of the research adequacy check provided by the present invention;

[0046] Figure 8 It is a structural schematic diagram of an unhealthy thinking monitoring device based on a large language model provided by the present invention;

[0047] Fig. 9 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] During the discussion, due to the lack of ability of participants to propose high-quality solutions, the influence of authority, the different interests inside and outside the meeting, and the constraints of meeting time, the decision-making of the meeting is often made by the highest decision-maker of the meeting or simply by the principle of minority obeying the majority. This decision-making method easily leads to unhealthy thinking patterns such as dependent thinking, rigid thinking, and divergent thinking in the group, which in turn affects the scientificity and effectiveness of the decision. What is more serious is that this decision-making method is usually based on partial information, lacks comprehensiveness and depth, and therefore has a greater risk of decision failure.

[0050] With the continuous development of artificial intelligence technology, natural language processing (NLP) technology has made great progress. However, traditional NLP technology is mainly focused on text summarization, topic extraction, machine translation and other fields, and its processing ability for multi-person multi-round conversations in meetings is relatively limited, and it cannot effectively monitor and identify bad thinking in meetings.

[0051] Furthermore, in recent years, large models have emerged, among which large language models have demonstrated natural and powerful conversational capabilities, providing a new direction for machines to handle multi-round conversations, and also providing a feasible path for monitoring bad thinking in groups. However, although some large language models are currently able to process conversation content, in multi-person conversations, especially multi-person multi-round discussions with complex backgrounds, it is necessary to grasp the background of all parties in advance, understand the thinking state and core demands of all parties at all times, which makes monitoring bad thinking particularly difficult; that is, general large language models cannot globally control the group discussion process and complete the monitoring of bad thinking at the same time.

[0052] In this regard, the present invention provides a method for monitoring bad thinking based on a large language model, which aims to monitor group discussions using the bad thinking monitoring model constructed on the basis of the large language model, and detect bad thinking in group discussions based on the information obtained from the monitoring, thereby realizing the discovery of bad thinking in group discussions, and providing a reference and basis for the regulation after the emergence of bad thinking, which can greatly improve the comprehensiveness and accuracy of decision-making, and thus avoid the risk of decision-making failure.

[0053] Figure 1 is a flow chart of the method for monitoring bad thinking based on a large language model provided by the present invention, such as Figure 1 As shown, the method is applied to a conference system, and the method includes:

[0054] Step 110, monitoring the group discussion and obtaining background information of the participants in the group discussion and a conversation sequence generated by the group discussion, wherein the conversation sequence includes the speeches and operations of the participants in the group discussion and their corresponding times;

[0055] Step 120, inputting the conversation sequence and background information into the bad thought monitoring model to obtain the thought monitoring result output by the bad thought monitoring model;

[0056] The bad thought monitoring model is trained on the basis of a large language model using sample background information of people participating in the sample group discussion, sample conversation sequences generated by the sample group discussion, and bad thought labels from the sample group discussion.

[0057] Specifically, considering that traditional natural language processing technology cannot handle multi-person conversations well and cannot detect bad thinking in multi-person meetings, in an embodiment of the present invention, a large language model with strong understanding and interaction capabilities is introduced to construct a bad thinking monitoring model for multi-person meeting scenarios. The model is used to monitor multi-person meetings / group discussions to detect bad thinking therein, thereby realizing the discovery of bad thinking in group discussions and providing a basis for solving bad thinking.

[0058] Before that, it is first necessary to monitor the entire group discussion process to obtain information such as the speech, identity, and operations of each person participating in the group discussion, so as to achieve monitoring and control of the entire group discussion process, so that the emergence of bad thinking can be accurately detected based on the monitored information and the detection results can be obtained.

[0059] That is, when the group discussion has been started in the conference system and all parties involved in the discussion have entered the discussion, it can be confirmed that the group discussion has officially started. At this time, the conference system enters the monitoring state and starts to monitor the entire group discussion. It is worth noting that the monitoring here is not only to monitor the speeches of the participants in the group discussion, but also the time and content of the speeches, identity background, and their operations in the entire group discussion. Multi-faceted monitoring can ensure the accuracy of subsequent bad thinking detection and the comprehensiveness of group decision-making as much as possible, reducing the risk of decision-making failure.

[0060] Here, by monitoring the group discussion, the conference system can obtain the conversation sequence generated during the group discussion, and the conversation sequence here includes the speeches, operations and corresponding times of the participants in the group discussion; that is, by monitoring the group discussion, the speeches, speech times, operations and operation times of the participants in the group discussion can be directly obtained; however, since this part of the information obtained by direct monitoring is very scattered, it is difficult for the model to monitor the thinking state of the personnel from the scattered information and detect bad thinking in the group. Therefore, in the embodiment of the present invention, it is necessary to further process the above-mentioned directly monitored information to integrate the scattered information, encode the scattered information as special characters, and write it into the speech, so as to obtain a multi-person speech record containing complete information, that is, the conversation sequence of the group discussion.

[0061] It is worth noting that the background information of the participants in the group discussion may be obtained by monitoring the conference system during the group discussion. For example, the identity information, educational information, personality characteristics, communication skills and other background information of the corresponding person or other persons may be obtained from the speeches of the participants in the group discussion. The hands-on ability, past experience, identity information, interests and hobbies of the corresponding person or other persons may also be obtained from the operations of the participants in the group discussion. The background information may also be actively reported / filled in by the participants in the group discussion at the beginning of the group discussion or before the beginning of the group discussion. The conference system may also actively collect the background information at the beginning of the group discussion or before the beginning of the group discussion. The embodiments of the present invention do not make specific limitations on this.

[0062] After obtaining the background information of the participants in the group discussion and the conversation sequence generated by the group discussion, in an embodiment of the present invention, this information can be pushed to the bad thinking monitoring model so that the model can detect bad thinking based on the received information, thereby obtaining the detection result of bad thinking in the group discussion, that is, the thinking monitoring result.

[0063] It should be noted that during the entire group discussion process, after the group discussion begins, the bad thinking monitoring model will log in to the system as an assistant and enter the group discussion to monitor the entire group discussion, detect bad thinking in the entire group discussion, control the progress of the group discussion, and make comprehensive and accurate decisions.

[0064] Figure 2 : is a flow chart of the group discussion monitoring and bad thinking detection process provided by the present invention, such as Figure 2 As shown, when the conference system monitors the group discussion, it will determine whether the discussion termination conditions are met after each round of discussion. If so, the group discussion will be terminated. Otherwise, the acquired background information and conversation sequence will be pushed to the bad thinking monitoring model so that it can detect bad thinking based on the received information, thereby obtaining the thinking monitoring results.

[0065] Correspondingly, the bad thinking monitoring model will continue to receive background information and conversation sequences pushed by the conference system during group discussions, and can conduct bad thinking detection on group discussions based on these two, so as to detect whether bad thinking such as dependent thinking, rigid thinking, divergent thinking, etc. appears in the group discussion process, thereby obtaining the detection results of bad thinking and outputting them, so as to obtain the thinking monitoring results output by the model.

[0066] Here, the thought monitoring result is an indication of whether bad thoughts appear in the group discussion after detection. It can be a parameter, command, warning, prompt, etc. indicating that bad thoughts appear in the group discussion, or it can be an operation, prompt, instruction, etc. indicating that bad thoughts do not appear in the group discussion. The embodiment of the present invention does not make specific limitations on this.

[0067] In addition, it should be noted that when it is detected that bad thinking occurs in group discussions, that is, when the thinking monitoring results reflect that bad thinking occurs in group discussions, in order to ensure the smooth progress of the discussion and to reach a comprehensive and accurate decision, in the embodiment of the present invention, the group discussion can be intervened to return it to a normal thinking mode and avoid falling into the dilemma of bad thinking, which leads to limitations in the final decision and a greater risk of decision failure. For example, an alarm can be generated to remind the people currently participating in the group discussion that they should pay attention to avoid bad thinking in the discussion; further, a control tool can be provided so that the people participating in the group discussion can access the control tool, and use it to get rid of the dilemma of bad thinking and re-discuss to get a feasible decision.

[0068] Among them, before using the bad thinking monitoring model to monitor bad thinking in group discussions, in an embodiment of the present invention, it is also necessary to apply the sample background information of the participants in the sample group discussion, the sample conversation sequence generated by the sample group discussion, and the bad thinking label training of the sample group discussion to pre-train the bad thinking monitoring model.

[0069] Here, the training process of the bad thinking monitoring model includes: out of consideration for the powerful understanding and interaction capabilities of the large language model, in the embodiment of the present invention, a large language model is selected to process the conversation in the multi-person conference scenario, so as to use the powerful ability of the large language model to analyze and understand the content of the conversation, thereby identifying bad thinking in the group discussion. That is, a large language model is selected to construct the initial model of the training process, and then, based on the pre-collected sample background information of the participants in the sample group discussion, the sample conversation sequence generated by the sample group discussion, and the bad thinking labels of the sample group discussion, the initial model is trained to obtain a trained bad thinking monitoring model.

[0070] The bad thinking monitoring method based on a large language model provided by the present invention monitors group discussions, obtains background information of persons participating in group discussions, and a conversation sequence generated by the group discussions, inputs the conversation sequence and the background information into the bad thinking monitoring model, and obtains a thinking monitoring result output by the bad thinking monitoring model; the bad thinking monitoring model is obtained by training on the basis of a large language model using sample background information of persons participating in sample group discussions, sample conversation sequences generated by sample group discussions, and bad thinking labels of sample group discussions, thus overcoming the defects of traditional schemes that cannot monitor group discussions as a whole and monitor bad thinking therein, thus realizing the monitoring of bad thinking in multi-person meetings, and providing a reference and basis for solving the bad thinking of the group, thus greatly improving the comprehensiveness and accuracy of decision-making, and avoiding the risk of decision-making failure.

[0071] Based on the above embodiment, the unhealthy thinking monitoring model includes a large language model and a target language model with the same structure, and the target language model is constructed based on the large language model; step 120 includes:

[0072] Inputting the background information into the target language model in the bad thinking monitoring model to obtain the background features output by the target language model;

[0073] The background features and conversation sequences are input into the large language model in the bad thought monitoring model to obtain the thought monitoring results output by the large language model.

[0074] Taking into account that large language models often have the problem of attention dilution when processing long texts, resulting in poor processing effects, based on this, in an embodiment of the present invention, when constructing an unhealthy thinking monitoring model based on a large language model for detecting unhealthy thinking in a group, a twin structure can be adopted to utilize the symmetrical twin structure to alleviate the problem of attention dilution caused by the long input text when using a single large language model, thereby improving processing efficiency and processing effects, thereby achieving an improvement in the precision and accuracy of unhealthy thinking detection.

[0075] In view of this, in the embodiment of the present invention, the structure of the bad thinking monitoring model used for bad thinking detection can be set to a twin structure, that is, the bad thinking monitoring model includes two models with the same structure, namely the large language model and the target language model. In order to ensure the consistency of the two structures, the target language model in the embodiment of the present invention can be constructed on the basis of the large language model.

[0076] In this case, in the process of inputting the conversation sequence and background information into the bad thinking monitoring model so that the bad thinking monitoring model can monitor bad thoughts accordingly to obtain the thinking monitoring results output by the bad thinking monitoring model, the conversation sequence and background information can be respectively input into two different models in the twin model, so that different models can be used to analyze and process the conversation sequence and background information respectively. This can not only make full use of the powerful understanding and analysis capabilities of the large language model, but also alleviate the problem of attention dilution when using a single large language model to process long texts, thereby obtaining accurate detection results.

[0077] Specifically, here the background information can be input into the target language model in the bad thinking monitoring model, so that the target language model analyzes and processes it to extract the background features corresponding to the background information. Then, the background features and the conversation sequence can be input into the large language model in the bad thinking monitoring model, so that the large language model can perform bad thinking detection on group discussions based on words, so as to detect whether bad thinking such as dependent thinking, rigid thinking, divergent thinking, etc. appears in the group discussion, thereby obtaining the thinking monitoring results.

[0078] Based on the above embodiment, the bad thinking monitoring model further includes a feature fusion layer; the feature fusion layer connects the target language model and the large language model;

[0079] Input the background features and conversation sequence into the large language model in the bad thought monitoring model to obtain the thought monitoring results output by the large language model, including:

[0080] The background features and conversation sequences are input into the feature fusion layer to obtain the fusion features output by the feature fusion layer; the fusion features are input into the large language model to obtain the thought monitoring results output by the large language model.

[0081] Specifically, since the large language model and the target language model in the bad thinking monitoring model are independent of each other, in an embodiment of the present invention, a feature fusion layer can be set between the two to connect the large language model and the target language model.

[0082] Figure 3 Schematic diagram of the structure of the bad thinking monitoring model provided by the present invention. Figure 3 As shown, in addition to the large language model and the target language model, the bad thinking monitoring model also includes a feature fusion layer connecting the two. Here, the feature fusion layer can be constructed on the basis of MLP (Multilayer Perceptron).

[0083] Among them, the feature fusion layer is used to fuse the output features of the target language model and the conversation sequence, and input the fused features into the input features of the large language model, so that the large language model can detect bad thinking in group discussions, thereby obtaining the corresponding thinking monitoring results of the group discussions.

[0084] Here, it should be noted that the large language model includes multiple feature transformation layers (transformer layers); and since the large language model and the target language model have the same structure, the target language model also includes multiple feature transformation layers. In addition, the parameters of the feature transformation layer in the target language model are different from those of the feature transformation layer in the large language model.

[0085] Correspondingly, there are multiple feature fusion layers in the bad thinking monitoring model, that is, the number of feature fusion layers is the same as the number of feature conversion layers in the large language model and the target language model. In addition, each feature fusion layer is connected to the feature conversion layer of the current layer in the target language model and the feature conversion layer of the previous layer in the large language model.

[0086] Specifically, when the background features and conversation sequences output by the target language model are input into the large language model so that the large language model can detect bad thinking based on them, the background features and conversation sequences can be first input into the feature fusion layer to perform feature fusion through the feature fusion layer to obtain fusion features output by the feature fusion layer. Then, the fusion features can be input into the large language model so that the large language model can detect bad thinking in group discussions based on them, and obtain the thinking monitoring results output by the large language model.

[0087] Here, specifically, the background information and the conversation sequence are segmented through the tokenizer, and then the segmented background information is input into the first feature conversion layer in the target language model to obtain the output feature of the first feature conversion layer; then, the output feature and the segmented conversation sequence are input into the first feature fusion layer, so that the feature fusion layer fuses the two to obtain the fused feature; then, the fused feature can be input into the first feature conversion layer in the large language model to process the fused feature and output it accordingly; then, the output feature can be input into the second feature fusion layer together with the input feature of the second feature conversion layer in the target language model to obtain the fused feature after the second fusion, and the fused feature is input into the second feature conversion layer in the large language model, and this is repeated until the output feature of the last feature conversion layer in the large language model is obtained, and the final thinking detection result can be obtained after passing through the detokenizer.

[0088] The detokenizer corresponds to the tokenizer model. The tokenizer is used to segment the original text into a series of tokens, while the detokenizer is responsible for recombining these tokens into meaningful sentences or paragraphs. Specifically in the embodiment of the present invention, the tokenizer is used to segment the conversation sequence and background information, and the detokenizer is used to receive the output of the model and convert it into a specific form.

[0089] In the embodiment of the present invention, a feature fusion layer constructed on the basis of MLP acts as a mask for the output features of two models with the same structure when performing feature fusion, so that the large language model for processing the conversation can fully utilize the output features of each feature conversion layer in the target language model; and the hot-pluggable architecture can shield the MLP structure and bypass the basic large model at any time, making the model more widely used.

[0090] Based on the above embodiment, the bad thinking monitoring model is trained based on the following steps:

[0091] Based on the large language model, a twin model is constructed, which includes a first language model and a second language model with the same parameters and structure;

[0092] Inputting sample background information into a first language model to obtain sample background features output by the first language model, inputting the sample background features and the sample conversation sequence into a second language model to obtain a predictive thinking monitoring result output by the second language model;

[0093] Based on the predicted thinking monitoring results and bad thinking labels, the first language model in the twin model is trained to obtain the bad thinking monitoring model.

[0094] Specifically, the training process of the bad thinking monitoring model includes the following steps:

[0095] In order to avoid the problem of attention dilution when a single large language model processes long texts, in the embodiment of the present invention, the bad thinking monitoring model is set as a twin structure, so that the two symmetrical models in the twin structure can be used to process the background information and the conversation sequence respectively, thereby improving the processing efficiency and processing effect. Therefore, when training the model, it is also necessary to pre-build two initial models of the twin structure, that is, a twin model can be built based on a large language model, and the twin model contains two initial models with the same structure and parameters, namely the first language model and the second language model.

[0096] Specifically, here, a large language model can be used as a benchmark to build another large language model with the same structure and parameters, and these two are called the first language model and the second language model respectively. For example, the parameters of the large language model can be copied, and a language model can be built according to its architecture as the first language model, and the original large language model can be used as the second language model, so that two models with the same structure and parameters are obtained, that is, two initial models. Further, the two can be constructed to form a twin model.

[0097] Next, the sample background information of the participants in the sample group discussion can be input into the first language model so that the first language model can analyze and process it to obtain the sample background features output by the first language model, and the sample background features and the sample conversation sequence generated by the sample group discussion can be input into the second language model so that the second language model can monitor bad thinking based on it, thereby obtaining the predicted thinking monitoring results output by the second language model.

[0098] Here, the predicted thinking monitoring result is an indication of whether bad thinking occurs in the sample group discussion after detection. It may be a parameter, command, warning, prompt, etc. indicating that bad thinking occurs in the sample group discussion, or it may be an operation, prompt, instruction, etc. indicating that bad thinking does not occur in the sample group discussion. The embodiment of the present invention does not make specific limitations on this.

[0099] The sample background information and the sample conversation sequence may be determined in the following manner:

[0100] In order to realize the monitoring of bad thinking based on large language models, so that the bad thinking monitoring model has the ability to detect bad thinking in group discussions, the model must be trained in advance so that it can learn the mapping relationship between sample data and labels during the training process, so that it can directly output the detection results corresponding to the input information based on this mapping relationship in the subsequent application process. Before training the model, training data must be collected in advance, that is, the background information and conversation sequences corresponding to multiple group discussions are collected to obtain sample background information and sample conversation sequences of sample group discussions. For example, in the database of the conference system, desensitized data of multiple group discussions can be obtained. The desensitized data of each group discussion includes the background, speech, various operations of each participant on the meeting or system during the discussion, as well as the speech time and operation time of each participant in the group discussion.

[0101] Furthermore, after obtaining the desensitized data of multiple group discussions, it is necessary to process the desensitized data to form training sample data and labels. That is, the desensitized data can be processed first to convert it into text form and obtain the speech of each participant. For example, for the audio file of the group discussion, the ASR (Automatic Speech Recognition) technology can be used to convert it into text. On this basis, the operation of the personnel can be combined, and the speaker annotation technology can be used to complete the speaker separation and annotation to obtain the speech text of each participant; then, it can be cleaned to write the operation and its corresponding time as special characters into the speech context to obtain the speech record of the participants, that is, the sample conversation sequence. Correspondingly, based on the desensitized data, the personnel id (Identity document)-background dictionary, that is, the sample background information, can be obtained.

[0102] After obtaining the predictive thinking monitoring results output by the twin model, model training can be performed based on the predictive thinking monitoring results to obtain a trained bad thinking monitoring model.

[0103] Figure 4 Schematic diagram of the bad thinking monitoring model training provided by the present invention, such as Figure 4 As shown, the parameters of the first language model in the twin model can be iterated according to the predicted thinking monitoring results output by the second language model and the bad thinking labels, so that the features output by the first language model after parameter adjustment and the results of bad thinking detection obtained by cooperating with the features output by the second language model are as consistent as possible with the bad thinking labels, and finally a trained bad thinking monitoring model can be obtained.

[0104] Here, it should be noted that the training of the twin model adopts the LoRA fine-tuning method. During the fine-tuning process, the parameters of the second language model are frozen and do not participate in the training. The first language model is fine-tuned using LoRA technology. During fine-tuning, the sample background information of a specific person and the corresponding speech records are used as input. The speech records of the next person of the specific person in the sample conversation sequence (including the Assistant's speech, that is, containing bad thinking labels) are used as supervision signals to fine-tune the first language model. Finally, the trained first language model can be obtained, on this basis, a bad thinking monitoring model can be constructed.

[0105] Among them, the bad thinking labels can be obtained by manual labeling by thinking experts after obtaining the sample conversation sequence, or by automatic labeling using a general large model. Specifically, when labeling, Assistant speech can be inserted into the sample conversation sequence where forced substitution, disconnection, language communication barriers, groupthink, blind application of experience, failure to distinguish between assumptions and facts, and excessive thinking appear, indicating that bad thinking occurs there, thereby obtaining the bad thinking labels corresponding to the sample group discussion.

[0106] Based on the above embodiment, the first language model includes a plurality of first feature conversion layers, and the second language model includes a plurality of second feature conversion layers;

[0107] The twin model also includes multiple initial feature fusion layers, the first language model and the second language model are connected through multiple initial feature fusion layers, and the number of layers of the first feature conversion layer, the second feature conversion layer and the initial feature fusion layer is the same;

[0108] Any initial feature fusion layer connects the first feature conversion layer of the corresponding layer in the first language model and the second feature conversion layer of the previous layer of the corresponding layer in the second language model;

[0109] The initial feature fusion layer is used to fuse the output features of the first feature conversion layer of the corresponding layer in the first language model with the output features of the second feature conversion layer of the previous layer of the corresponding layer in the second language model, and use the fused features as the input features of the second feature conversion layer of the corresponding layer in the second language model.

[0110] Specifically, see Figure 4 It can be seen that in the embodiment of the present invention, the twin model also includes multiple initial feature fusion layers, and the initial feature fusion layer here is constructed on the basis of MLP. The initial feature fusion layer is used to connect the first language model and the second language model in the twin model. Among them, the first language model contains multiple first feature conversion layers (transformer layer), and the second language model contains multiple second feature conversion layers (transformer layer); and the number of layers of the first feature conversion layer, the second feature conversion layer and the initial feature fusion layer in the twin model are the same.

[0111] Among them, when any initial feature fusion layer is connected to the first language model and the second language model, it connects the first feature conversion layer of the corresponding layer in the first language model and the second feature conversion layer of the previous layer of the corresponding layer in the second language model; in short, the second layer / the second initial feature fusion layer connects the second layer / the second first feature conversion layer in the first language model in the twin model and the first layer / the first second feature conversion layer in the second language model; the third layer / the third initial feature fusion layer connects the third layer / the third first feature conversion layer in the first language model in the twin model and the second layer / the second second feature conversion layer in the second language model. The first layer / the first initial feature fusion layer connects the first layer / the first first feature conversion layer in the first language model in the twin model and the tokenizer of the sample conversation sequence.

[0112] Furthermore, any initial feature fusion layer fuses the output features of the first feature conversion layer of the corresponding layer in the first language model and the output features of the second feature conversion layer of the previous layer of the corresponding layer in the second language model to obtain the fused features of the corresponding layer, and uses the fused features as the input features of the second feature conversion layer of the corresponding layer in the second language model. That is, the initial feature fusion layer of the current layer is used to fuse the output features of the first feature conversion layer of the current layer and the output features of the second feature conversion layer of the previous layer.

[0113] Based on the above embodiment, based on the predicted thinking monitoring results and the bad thinking labels, the first language model in the twin model is trained to obtain the bad thinking monitoring model, including:

[0114] Based on the predicted thought monitoring results and bad thought labels, the first language model and each initial feature fusion layer in the twin model are trained to obtain the target language model and multiple feature fusion layers;

[0115] Based on the large language model, the target language model and various feature fusion layers, a bad thinking monitoring model is constructed.

[0116] Specifically, the process of training the first language model in the twin model according to the predicted thinking monitoring results and the bad thinking labels to obtain the bad thinking monitoring model may specifically include:

[0117] Firstly, the twin model is trained using the predictive thinking monitoring results and bad thinking labels output by the twin model. Specifically, the parameters of the second language model are frozen during training, and the first language model and the initial feature fusion layer are fine-tuned using LoRA technology. That is, according to the predictive thinking detection results and bad thinking labels, the first language model and each initial feature fusion layer in the twin model are trained to adjust the parameters of the first language model and the initial feature fusion layer, so that the predictive thinking monitoring results output by the twin model after parameter adjustment for the sample background information and sample conversation sequence discussed by the input sample group can be as close as possible to the bad thinking labels discussed by the sample group, thereby obtaining the trained model and fusion layer, that is, the target language model and each feature fusion layer.

[0118] Subsequently, a bad thinking monitoring model can be constructed based on the target language model and each feature fusion layer. Specifically, the final bad thinking monitoring model can be constructed based on the target language model and each feature fusion layer, combined with a large language model. Here, since the structure of the initial model corresponding to the target language model is the same as that of the first language model, and the first language model is a large language model, the bad thinking monitoring model finally constructed includes a symmetrical large language model and a target language model, as well as multiple feature fusion layers connecting the two.

[0119] Furthermore, after obtaining the bad thinking monitoring model, in the embodiment of the present invention, model evaluation can also be performed to test the performance of the trained bad thinking monitoring model. In the embodiment of the present invention, two evaluation methods are proposed: one is that the model predicts the speech record of the next person according to the speech record of the current person, and inputs the prediction result and the test set groundtruth into the large language model respectively, and takes the output of the n / 2 hidden layers after the large language model to calculate the similarity; the second is that the model predicts the speech record of the next person according to the previous part of the speech record, and adds the prediction result as a historical record to iterate continuously until the end, and inputs the conversation sequence generated by the model and the conversation sequence and background information annotated by the test set into the large model to evaluate the model performance.

[0120] Based on the above embodiment, the conversation sequence and background information are input into the bad thinking monitoring model to obtain the thinking monitoring result output by the bad thinking monitoring model, and then the following steps are further included:

[0121] When the thinking monitoring results indicate that bad thinking occurs in the group discussion, bad thinking warning prompts and control tools are obtained so that the people participating in the group discussion can control the group discussion based on the bad thinking warning prompts and control tools; the bad thinking warning prompts and control tools are generated by the bad thinking monitoring model based on the thinking monitoring results.

[0122] Specifically, in step 120, after the conversation sequence and background information are input into the bad thinking monitoring model and the thinking monitoring results output by the bad thinking monitoring model are obtained, if it is detected that bad thinking occurs in the group discussion, that is, the thinking monitoring results indicate that bad thinking occurs in the group discussion, then in order to ensure that the discussion can reach a comprehensive and accurate decision, the group discussion can be intervened, that is, bad thinking alarm prompts and control tools can be obtained, so that the people participating in the group discussion can know that bad thinking has occurred in the discussion through the bad thinking alarm prompts, and can make timely adjustments, and the group discussion can be regulated through the control tools, so as to get rid of the dilemma of bad thinking, return to normal thinking mode, and re-discuss to obtain comprehensive, accurate, and reliable decisions.

[0123] Among them, the bad thinking warning prompts and control tools can be generated by the bad thinking monitoring model when bad thinking is detected in the discussion, and can be output together with the thinking monitoring results reflecting the occurrence of bad thinking in the group discussion.

[0124] Specifically, in the embodiments of the present invention, the model may provide a variety of control tools, such as anonymous discussions, nominal groups, group discussions, discussion adequacy checks, etc.

[0125] Among them, anonymous discussion tools are: In order to avoid the potential impact of the identity of people outside the discussion on the discussion, the conference system provides anonymous discussion function. Figure 5 It is a flowchart of the anonymous discussion process provided by the present invention, such as Figure 5 As shown in the figure, when the participants in the group discussion actively initiate an anonymous discussion or the model detects bad thinking, the conference system will initiate an anonymous invitation in the name of the system. Each participant can choose to enter the anonymous state. When all participants enter the anonymous state, the non-anonymous status will no longer be displayed. Otherwise, the participants who enter the anonymous state will be displayed in the form of incremental personnel to ensure the effectiveness of anonymity. Among them, the participants are indexed by the unique identifier of the person and displayed as a random id in the group discussion. For audio and video conferences, virtual digital human technology is used to replace the image and voice of the original participants.

[0126] Nominal group tool: In order to avoid misjudgment of the silent opinions, the conference system provides a nominal group function. Figure 6 It is a flowchart of the nominal group discussion provided by the present invention, such as Figure 6 As shown in the figure, when the participants in the group discussion actively initiate a nominal group discussion or the model detects bad thinking, the conference system will initiate a nominal group invitation in the name of the system, and each participant will be asked to express his or her own views and reasons on the discussion topic. In order to avoid individual opinions taking up too much time, entering the nominal group mode will limit the upper limit of each person's speaking time.

[0127] The group discussion tool is: In order to reduce the pressure of individuals opposing group opinions, the conference system provides a group discussion function. The discussion group inherits multiple functions of the entire conference system, including the anonymous discussion function, and can be ended at any time and returned to the original group discussion venue. When any person participating in the group discussion feels the pressure caused by potential opposition to group opinions, he can initiate a discussion group creation invitation with the system role to conduct a group discussion. In addition, by default, the person enters the discussion group anonymously, while keeping the original group discussion online, ensuring that the pressure of the original group discussion will not be transmitted to the current discussion group through any information.

[0128] In order to avoid reaching a consensus quickly without in-depth discussion of the issue, a discussion adequacy check can be conducted. Figure 7 It is a flow chart of the research adequacy check provided by the present invention, such as Figure 7 As shown, the subject tool is first called to check the relevance of the participants' speeches to the topic and the completeness of the arguments provided. Irrelevant arguments are removed from the argument set. When the completeness of the argument set is sufficient to support the root argument, the current discussion is considered sufficient and the discussion is allowed to end or change the topic. Otherwise, the subject tool will obtain permission to prohibit the discussion from ending or continuing.

[0129] In an embodiment of the present invention, when the thinking monitoring results indicate the occurrence of unhealthy thinking in a group discussion, unhealthy thinking warning prompts and control tools are obtained to control the group discussion. For the first time, the expert's control over unhealthy thinking after it occurs is applied to online group discussions / conferences in the form of an assistant through a large language model, providing a specific and feasible path for solving unhealthy thinking in online group discussions / conferences.

[0130] The following is a description of the unhealthy thinking monitoring device based on a large language model provided by the present invention. The unhealthy thinking monitoring device based on a large language model described below and the unhealthy thinking monitoring method based on a large language model described above can be referenced to each other.

[0131] Figure 8 Schematic diagram of the structure of the bad thinking monitoring device based on the large language model provided by the present invention. Figure 8 As shown, the device is applied to a conference system, including:

[0132] The acquisition unit 810 is used to monitor the group discussion and acquire the background information of the participants in the group discussion and the conversation sequence generated by the group discussion, wherein the conversation sequence includes the speeches and operations of the participants in the group discussion and their corresponding times;

[0133] A monitoring unit 820, configured to input the conversation sequence and the background information into an unhealthy thinking monitoring model, and obtain a thinking monitoring result output by the unhealthy thinking monitoring model;

[0134] The bad thinking monitoring model is trained on the basis of a large language model using sample background information of people participating in a sample group discussion, sample conversation sequences generated by the sample group discussion, and bad thinking labels of the sample group discussion.

[0135] The bad thinking monitoring device based on a large language model provided by the present invention monitors group discussions, obtains background information of persons participating in group discussions, and a conversation sequence generated by the group discussions, inputs the conversation sequence and the background information into the bad thinking monitoring model, and obtains a thinking monitoring result output by the bad thinking monitoring model; the bad thinking monitoring model is obtained by training on the basis of a large language model using sample background information of persons participating in sample group discussions, sample conversation sequences generated by sample group discussions, and bad thinking labels of sample group discussions, thus overcoming the defects of traditional solutions that cannot monitor group discussions as a whole and detect bad thinking therein, thus realizing the monitoring of bad thinking in multi-person meetings, and providing a reference and basis for solving the bad thinking of the group, thus greatly improving the comprehensiveness and accuracy of decision-making and avoiding the risk of decision-making failure.

[0136] Based on the above embodiment, the bad thinking monitoring model includes a large language model and a target language model with the same structure, and the target language model is constructed based on the large language model; the monitoring unit 820 is used to:

[0137] Inputting the background information into the target language model in the bad thinking monitoring model to obtain the background features output by the target language model;

[0138] The background features and the conversation sequence are input into a large language model in the bad thought monitoring model to obtain a thought monitoring result output by the large language model.

[0139] Based on the above embodiment, the bad thinking monitoring model further includes a feature fusion layer; the feature fusion layer connects the target language model and the large language model;

[0140] The monitoring unit 820 is used for:

[0141] Inputting the background features and the conversation sequence into the feature fusion layer to obtain fusion features output by the feature fusion layer;

[0142] The fusion feature is input into the large language model to obtain the thought monitoring result output by the large language model.

[0143] Based on the above embodiment, the device further includes a training unit, which is used to:

[0144] Based on the large language model, construct a twin model, wherein the twin model includes a first language model and a second language model with the same parameters and structure;

[0145] Inputting the sample background information into the first language model to obtain sample background features output by the first language model, inputting the sample background features and the sample conversation sequence into the second language model to obtain a predictive thinking monitoring result output by the second language model;

[0146] Based on the predicted thought monitoring result and the bad thought label, the first language model in the twin model is trained to obtain a bad thought monitoring model.

[0147] Based on the above embodiment, the first language model includes a plurality of first feature conversion layers, and the second language model includes a plurality of second feature conversion layers;

[0148] The twin model further includes a plurality of initial feature fusion layers, the first language model and the second language model are connected via the plurality of initial feature fusion layers, and the first feature conversion layer, the second feature conversion layer and the initial feature fusion layer have the same number of layers;

[0149] Any initial feature fusion layer connects a first feature conversion layer of a corresponding layer in the first language model and a second feature conversion layer of an upper layer of a corresponding layer in the second language model;

[0150] Any initial feature fusion layer is used to fuse the output features of the first feature conversion layer of the corresponding layer in the first language model with the output features of the second feature conversion layer of the previous layer of the corresponding layer in the second language model, and use the fused features as the input features of the second feature conversion layer of the corresponding layer in the second language model.

[0151] Based on the above embodiment, the training unit is used for:

[0152] Based on the predicted thought monitoring result and the bad thought label, the first language model and each initial feature fusion layer in the twin model are trained to obtain a target language model and multiple feature fusion layers;

[0153] Based on the large language model, the target language model and each feature fusion layer, a bad thinking monitoring model is constructed.

[0154] Based on the above embodiment, the device further includes a control unit, which is used to:

[0155] When the thought monitoring result indicates that bad thoughts occur in the group discussion, obtaining bad thought warning prompts and control tools, so that the persons participating in the group discussion can control the group discussion based on the bad thought warning prompts and control tools;

[0156] The bad thinking warning prompt and regulation tool are generated by the bad thinking monitoring model based on the thinking monitoring result.

[0157] Fig. 9 An example of a physical structure diagram of an electronic device is shown in FIG. Fig. 9 As shown, the electronic device may include: a processor 910, a communication interface 920, a memory 930 and a communication bus 940, wherein the processor 910, the communication interface 920 and the memory 930 communicate with each other through the communication bus 940. The processor 910 may call the logic instructions in the memory 930 to execute the bad thinking monitoring method based on the large language model, the method comprising: monitoring the group discussion, and obtaining the background information of the participants in the group discussion, and the conversation sequence generated by the group discussion, wherein the conversation sequence includes the speeches, operations and corresponding times of the participants in the group discussion; inputting the conversation sequence and the background information into the bad thinking monitoring model to obtain the thinking monitoring result output by the bad thinking monitoring model; the bad thinking monitoring model is obtained by applying the sample background information of the participants in the sample group discussion, the sample conversation sequence generated by the sample group discussion, and the bad thinking label training of the sample group discussion on the basis of the large language model.

[0158] In addition, the logic instructions in the above-mentioned memory 930 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0159] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the bad thinking monitoring method based on a large language model provided by the above-mentioned methods, and the method includes: monitoring group discussions and obtaining background information of participants in the group discussions, and a conversation sequence generated by the group discussions, wherein the conversation sequence includes speeches, operations and their corresponding times of the participants in the group discussions; inputting the conversation sequence and the background information into a bad thinking monitoring model to obtain a thinking monitoring result output by the bad thinking monitoring model; the bad thinking monitoring model is obtained on the basis of a large language model, using sample background information of participants in a sample group discussion, a sample conversation sequence generated by the sample group discussion, and bad thinking labels of the sample group discussion.

[0160] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the bad thinking monitoring method based on a large language model provided by the above-mentioned methods, the method comprising: monitoring group discussions, and obtaining background information of persons participating in the group discussions, and a conversation sequence generated by the group discussions, wherein the conversation sequence includes speeches, operations and their corresponding times of the persons participating in the group discussions; inputting the conversation sequence and the background information into a bad thinking monitoring model to obtain a thinking monitoring result output by the bad thinking monitoring model; the bad thinking monitoring model is obtained by training on the basis of a large language model using sample background information of persons participating in the sample group discussions, a sample conversation sequence generated by the sample group discussions, and bad thinking labels of the sample group discussions.

[0161] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0162] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring bad thinking based on a large language model, characterized in that: Applied to a conference system, the method comprises: Monitor the group discussion and obtain background information of participants in the group discussion and a conversation sequence generated by the group discussion, wherein the conversation sequence includes speeches and operations of the participants in the group discussion and their corresponding times; Inputting the conversation sequence and the background information into an unhealthy thought monitoring model to obtain a thought monitoring result output by the unhealthy thought monitoring model; The bad thinking monitoring model is trained on the basis of a large language model using sample background information of people participating in a sample group discussion, sample conversation sequences generated by the sample group discussion, and bad thinking labels of the sample group discussion.

2. The method for monitoring bad thinking based on a large language model according to claim 1, characterized in that: The bad thinking monitoring model includes a large language model and a target language model with the same structure, and the target language model is constructed based on the large language model; The step of inputting the conversation sequence and the background information into the bad thinking monitoring model to obtain the thinking monitoring result output by the bad thinking monitoring model includes: Inputting the background information into the target language model in the bad thinking monitoring model to obtain the background features output by the target language model; The background features and the conversation sequence are input into a large language model in the bad thought monitoring model to obtain a thought monitoring result output by the large language model.

3. The method for monitoring bad thinking based on a large language model according to claim 2, characterized in that: The bad thinking monitoring model also includes a feature fusion layer; The feature fusion layer connects the target language model and the large language model; The step of inputting the background features and the conversation sequence into the large language model in the bad thought monitoring model to obtain the thought monitoring result output by the large language model includes: Inputting the background features and the conversation sequence into the feature fusion layer to obtain fusion features output by the feature fusion layer; The fusion feature is input into the large language model to obtain the thought monitoring result output by the large language model.

4. The method for monitoring bad thinking based on a large language model according to any one of claims 1 to 3, characterized in that: The bad thinking monitoring model is trained based on the following steps: Based on the large language model, construct a twin model, wherein the twin model includes a first language model and a second language model with the same parameters and structure; Inputting the sample background information into the first language model to obtain sample background features output by the first language model, inputting the sample background features and the sample conversation sequence into the second language model to obtain a predictive thinking monitoring result output by the second language model; Based on the predicted thought monitoring result and the bad thought label, the first language model in the twin model is trained to obtain a bad thought monitoring model.

5. The method for monitoring bad thinking based on a large language model according to claim 4, characterized in that: The first language model includes a plurality of first feature conversion layers, and the second language model includes a plurality of second feature conversion layers; The twin model further includes a plurality of initial feature fusion layers, the first language model and the second language model are connected via the plurality of initial feature fusion layers, and the first feature conversion layer, the second feature conversion layer and the initial feature fusion layer have the same number of layers; Any initial feature fusion layer connects a first feature conversion layer of a corresponding layer in the first language model and a second feature conversion layer of an upper layer of a corresponding layer in the second language model; Any initial feature fusion layer is used to fuse the output features of the first feature conversion layer of the corresponding layer in the first language model with the output features of the second feature conversion layer of the previous layer of the corresponding layer in the second language model, and use the fused features as the input features of the second feature conversion layer of the corresponding layer in the second language model.

6. The method for monitoring bad thinking based on a large language model according to claim 5, characterized in that: The step of training the first language model in the twin model based on the predicted thinking monitoring result and the bad thinking label to obtain the bad thinking monitoring model includes: Based on the predicted thought monitoring result and the bad thought label, the first language model and each initial feature fusion layer in the twin model are trained to obtain a target language model and multiple feature fusion layers; Based on the large language model, the target language model and each feature fusion layer, a bad thinking monitoring model is constructed.

7. The method for monitoring bad thinking based on a large language model according to any one of claims 1 to 3, characterized in that: The step of inputting the conversation sequence and the background information into the bad thinking monitoring model to obtain the thinking monitoring result output by the bad thinking monitoring model further includes: When the thought monitoring result indicates that bad thoughts occur in the group discussion, obtaining bad thought warning prompts and control tools, so that the persons participating in the group discussion can control the group discussion based on the bad thought warning prompts and control tools; The bad thinking warning prompt and regulation tool are generated by the bad thinking monitoring model based on the thinking monitoring result.

8. A device for monitoring bad thoughts based on a large language model, characterized in that: Applied to a conference system, the device comprises: An acquisition unit, used to monitor the group discussion and acquire background information of participants in the group discussion and a conversation sequence generated by the group discussion, wherein the conversation sequence includes speeches and operations of participants in the group discussion and their corresponding times; A monitoring unit, used for inputting the conversation sequence and the background information into an unhealthy thinking monitoring model to obtain a thinking monitoring result output by the unhealthy thinking monitoring model; The bad thinking monitoring model is trained on the basis of a large language model using sample background information of people participating in a sample group discussion, sample conversation sequences generated by the sample group discussion, and bad thinking labels of the sample group discussion.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for monitoring bad thoughts based on a large language model as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for monitoring bad thoughts based on a large language model as described in any one of claims 1 to 7 is implemented.

Images