Text thinking chain sub-chain division method and device, equipment and medium

Abstract

The invention relates to the technical field of intelligent decision making, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses a text thinking chain sub-chain division method, device, equipment and medium, and the method comprises the following steps: obtaining an input text corresponding to a target task, and carrying out text structuring processing on the input text to obtain a structured text; performing confidence coefficient analysis on the target task and the structured text to obtain a confidence coefficient lexical element sequence, and identifying a confidence coefficient score of each lexical element in the confidence coefficient lexical element sequence; calculating the confidence decrease amplitude of each lexical element according to the confidence score; identifying a target division boundary in the confidence lexical element sequence according to the confidence decrease amplitude and semantic information of the input text; and identifying a structured logic unit of the confidence lexical element sequence according to the target division boundary, and dividing a thinking chain sub-chain of the confidence lexical element sequence according to the structured logic unit. According to the method, the sub-chain division accuracy of the text thinking chain can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent decision-making, and in particular to a text thought chain sub-chain division method and device, equipment and a medium. BACKGROUND

[0002] In the field of finance and insurance, language models (LLM) have been widely used in intelligent customer service, contract clause understanding, risk assessment, claim processing, and other scenarios. To improve the reasoning ability of models in complex decision-making, contract analysis, and claim review tasks, the research community generally uses the CoT method to guide the model to reason step by step to optimize task processing results. However, the current thought chain division in this field lacks an adaptive and automated solution, making it difficult to support the efficient advancement of various sub-tasks. Therefore, in order to meet the needs of the finance and insurance industry for flexible and efficient automatic thought chain division, it is necessary to make targeted innovations in thought chain division methods to improve the accuracy of text thought chain sub-chain division.

[0003] In the field of medical health, language models are often used for medical record analysis, diagnosis assistance, and medication guidance. To improve reasoning ability, the CoT method is used. Currently, it relies on manual division of thought chain steps, which requires medical personnel to annotate based on professional knowledge. However, different disease diagnosis and treatment logics differ greatly, and manual annotation cannot fully cover all aspects. Or, fixed token number and line number rules are used for division, which often splits coherent diagnosis and treatment reasoning, ignores semantic boundaries, and results in broken sub-chain logic or information redundancy, ultimately leading to low accuracy of text thought chain sub-chain division.

[0004] In the field of financial technology business, language models are often used for credit risk control review, financial product clause interpretation, and transaction anomaly detection. To improve the reasoning ability of models in these tasks, the CoT method is often used to guide step-by-step reasoning. Currently, either manual division of thought chain steps is relied on, which requires financial risk control or product experts to operate based on professional knowledge. However, different credit types and financial product categories have different business logics, and manual adaptation is difficult to fully cover all aspects. Or, fixed token number and line number rules are used for division, which often splits coherent business reasoning, resulting in low accuracy of text thought chain sub-chain division.

[0005] The prior art processes tasks in the field of finance and insurance, etc. After natural language preprocessing of the input text, the boundary of the thinking chain is determined by artificial professional knowledge in the reasoning link, the confidence score of the token of the reasoning generated text is not identified, the confidence drop amplitude is not calculated, the boundary judgment lacks quantitative support; or the boundary is divided by fixed token number, line number and other rules without combining the semantic information of the input text and the confidence drop amplitude, which cannot capture the decision inflection point in the reasoning process, resulting in chaotic division of the structured logic unit of the reasoning generated text, redundant or missing of the sub-chain information, and finally low accuracy of the division of the text thinking chain sub-chain. SUMMARY

[0006] The present application provides a text thinking chain sub-chain division method, device, equipment and medium to solve the problem of low accuracy of text thinking chain sub-chain division.

[0007] In a first aspect, a text thinking chain sub-chain division method is provided, comprising: obtaining an input text corresponding to a target task, performing text structuring processing on the input text to obtain a structured text; performing confidence analysis on the target task and the structured text to obtain a confidence token sequence, and identifying the confidence score of each token in the confidence token sequence; calculating the confidence drop amplitude of each token according to the confidence score; identifying a target division boundary in the confidence token sequence according to the confidence drop amplitude and the semantic information of the input text; identifying a structured logic unit of the confidence token sequence according to the target division boundary, and dividing the thinking chain sub-chain of the confidence token sequence according to the structured logic unit.

[0008] In a second aspect, a text thinking chain sub-chain division device is provided, comprising: an input text structuring processing module configured to obtain an input text corresponding to a target task, perform text structuring processing on the input text to obtain a structured text; a confidence analysis module configured to perform confidence analysis on the target task and the structured text to obtain a confidence token sequence, and identify the confidence score of each token in the confidence token sequence; a confidence drop amplitude calculation module configured to calculate the confidence drop amplitude of each token according to the confidence score; a target division boundary identification module configured to identify a target division boundary in the confidence token sequence according to the confidence drop amplitude and the semantic information of the input text; The thought chain sub-chain division module is configured to identify a structured logical unit of the confidence character sequence according to the target division boundary, and divide the thought chain sub-chain of the confidence character sequence according to the structured logical unit.

[0009] In a third aspect, a computer device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the text thought chain sub-chain division method when executing the computer program.

[0010] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program, and the computer program implements the steps of the text thought chain sub-chain division method when executed by a processor.

[0011] In the scheme implemented by the text thought chain sub-chain division method, device, equipment and medium, the input text corresponding to a target task can be acquired by a client, the input text is subjected to text structuring processing to obtain a structured text, confidence analysis is performed on the target task and the structured text to obtain a confidence character sequence, a confidence score of each character in the confidence character sequence is identified, the confidence score is used to calculate the confidence drop amplitude of each character, the target division boundary in the confidence character sequence is identified according to the confidence drop amplitude and semantic information of the input text, the structured logical unit of the confidence character sequence is identified according to the target division boundary, and the thought chain sub-chain of the confidence character sequence is divided according to the structured logical unit. In the present application, the input text corresponding to the target task is first subjected to structuring processing, then confidence analysis is performed on the target task and the structured text to obtain a confidence character sequence and identify the confidence score of each character, then the confidence drop amplitude of each character is calculated according to the score, and the target division boundary is identified in combination with the semantic information of the input text, and finally the structured logical unit and the thought chain sub-chain are divided. It does not need manual annotation, solves the problems of high cost and non-scalability of manual division, divides according to the confidence inflection point, avoids semantic imbalance and information redundancy of rule division, adapts to multiple tasks of finance and insurance, can be optimized in combination with feedback, has strong explainability, and can solve the problem of low accuracy of text thought chain sub-chain division. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1is an application environment diagram of a text thought chain sub-chain division method according to an embodiment of the present application; Figure 2 is a flow diagram of a text thought chain sub-chain division method according to an embodiment of the present application; Figure 3 is Figure 2 is a specific implementation flow diagram of step S3 in the method; Figure 4 is Figure 2 is a specific implementation flow diagram of step S4 in the method; Figure 5 is a structure diagram of a text thought chain sub-chain division device according to an embodiment of the present application; Figure 6 is a structure diagram of a computer device according to an embodiment of the present application; Figure 7 is another structure diagram of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0015] The text thought chain sub-chain division method provided by the embodiments of the present application can be applied in, for example, Figure 1In an application environment of the application, a client communicates with a server through a network. The server can obtain an input text corresponding to a target task through the client, perform text structuring processing on the input text to obtain a structured text, perform confidence analysis on the target task and the structured text to obtain a confidence token sequence, identify a confidence score of each token in the confidence token sequence, calculate a confidence drop amplitude of each token according to the confidence score, identify a target division boundary in the confidence token sequence according to the confidence drop amplitude and semantic information of the input text, identify a structured logic unit of the confidence token sequence according to the target division boundary, and divide a thought chain sub-chain of the confidence token sequence according to the structured logic unit. In the application, the input text corresponding to the target task is first structured, then confidence analysis is performed on the target task and the structured text to obtain a confidence token sequence and identify a confidence score of each token, then a confidence drop amplitude of each token is calculated according to the score, a target division boundary is identified according to the input text semantic information, and finally a structured logic unit and a thought chain sub-chain are divided. The application does not require manual annotation, solves the problems of high cost and non-scalability of manual division, avoids semantic imbalance and information redundancy of rule division according to the confidence inflection point, adapts to multiple tasks in the financial insurance field, can be optimized combined with feedback, has strong explainability, and can solve the problem of low accuracy of text thought chain sub-chain division. The client can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers. The application will be described in detail below through specific embodiments.

[0016] Please refer to Figure 2 , as shown in Figure 2 A flowchart of a text thought chain sub-chain division method provided by an embodiment of the application is shown in the figure, and includes the following steps: S1, obtaining an input text corresponding to a target task, performing text structuring processing on the input text to obtain a structured text.

[0017] In an embodiment of the application, the target task is a related task in the financial insurance field, such as “explain the policy clause”, “judge whether the claim can be processed” and “calculate the refund amount”, etc. The input text is a text corresponding to these target tasks, such as a contract clause paragraph, a claim application description, etc.

[0018] In detail, the input processing module receives target tasks in the financial insurance field and input texts corresponding to the target tasks, and pre-processes the received input texts.

[0019] In the embodiment of the present application, the structured text refers to a text form that is formed after pre-processing the input text in the financial insurance field and integrates token sequence, entity information, clause structure information and other specific structures, and can adapt to the input requirements of subsequent initial inference generation module and other link models.

[0020] In the embodiment of the present application, the text structuring processing on the input text to obtain the structured text comprises: text normalization on the input text to obtain an initial processing text; labeling key entities in the initial processing text to obtain an intermediate entity text; text structure annotation on the intermediate entity text to obtain a target labeled text; associating and binding entity information and clause structure information in the target labeled text to obtain the structured text of the input text.

[0021] In detail, the initial processing text refers to a text that is formed after text normalization of the input text corresponding to the target task in the financial insurance field, and can adapt to the input requirements of the subsequent initial inference generation module. The input processing module first receives the target task and the corresponding input text in the financial insurance field, and then performs text normalization operations such as entity recognition and clause structure labeling on the received input text, thereby obtaining the initial processing text.

[0022] Specifically, the key entity refers to core information related to the target task in the financial insurance scenario, such as policy number, insured person information, claim ratio, exclusion liability range, protected disease name, medical diagnosis result, and claim application amount. The intermediate entity text refers to a text form that highlights core information formed after identifying and labeling key entities in the initial processing text. After receiving the target task and the corresponding input text in the financial insurance field, and performing tokenization preprocessing on the input text to obtain the initial processing text, the input processing module continues to identify and label the key entities in the initial processing text, thereby obtaining the intermediate entity text.

[0023] Further, the target labeled text refers to a text form that further integrates the annotation results of the text structure information in the financial insurance scenario based on the intermediate entity text, and can more accurately adapt to the input requirements of the subsequent initial inference generation module. After receiving the target task and the corresponding input text in the financial insurance field, and completing tokenization preprocessing to obtain the initial processing text, and labeling the key entities in the initial processing text to obtain the intermediate entity text, the input processing module continues to perform text structure annotation operations such as clause structure labeling on the intermediate entity text, and finally obtains the target labeled text.

[0024] Further, the entity information refers to core information related to the target task in the financial insurance scenario, such as the insured person information, the policy number, the proportion of compensation, the scope of exclusion, the guaranteed disease, the amount of claim application, the medical diagnosis result, and the like. These information are the contents marked in the target marked text by the input processing module after performing entity recognition on the input text. The clause structure information refers to the structure division information of the financial insurance related input text, such as the chapter division of the contract clause, the information module boundary of the claim application, and the like. These information are the contents marked in the target marked text by the input processing module after performing clause structure marking on the input text. After obtaining the target marked text by performing text structure marking on the intermediate entity text, the marked entity information in the target marked text is further associated with the marked clause structure information, so that each entity information is clearly attributed to the text structure position where it is located, and then the structured text of the input text which can clearly reflect the association relationship between information and adapt to the input requirement of the subsequent initial reasoning generation module is formed.

[0025] For example, in a medical scenario, the target task is to "judge whether a patient meets a certain high blood pressure chronic disease medication condition", and the input text is a patient medical record: "The patient is 65 years old, has a 5-year history of hypertension, this time the blood pressure is 150 / 95 mmHg, and the liver and kidney function is normal". Tokenization and basic information splitting are performed on the medical record to obtain an initial processing text, and key entities such as "65 years old (age)", "5-year history of hypertension (past medical history)", "150 / 95 mmHg (blood pressure value)", and "normal liver and kidney function (organ index)" are marked to obtain an intermediate entity text. The structure modules of "basic information", "past medical history", "re-examination index", and "organ function" are further marked in the medical record to obtain a target marked text, and finally the association and binding of each entity information and the corresponding structure module are obtained to obtain the structured text of the medical record.

[0026] For example, in a financial scenario, the target task is to "judge whether a major disease insurance claim application meets the compensation conditions", and the input text is a claim application description: "The policyholder Zhang recently diagnosed with thyroid cancer and received surgical treatment in a local top three hospital. The insurance type is Kangjianwuyou major disease insurance, the policy status is valid, and there is no past disease concealment or exemption situation". Tokenization and basic information splitting are performed on the application text to obtain an initial processing text, such as "policyholder Zhang", "thyroid cancer", "top three hospital", and the like token sequence. Key entities such as "Zhang (policyholder)", "thyroid cancer (diagnosed disease)", and "Kangjianwuyou major disease insurance (insurance type)" are marked to obtain an intermediate entity text. The structure modules of "policyholder information", "treatment situation", "insurance type", and "policy status" are further marked in the application text to obtain a target marked text, and finally the association and binding of each entity information and the corresponding structure module are obtained to obtain the structured text of the claim application input text.

[0027] S2, performing confidence analysis on the target task and the structured text to obtain a confidence token sequence, and identifying a confidence score of each token in the confidence token sequence.

[0028] In the embodiment of the application, the confidence token sequence refers to a sequence formed by recording the confidence score of each output token of the model at the token level in the process of receiving an input prompt and generating an answer in a thought chain manner by a basic LLM (language model, such as GPT-o1, DeepSeek-R1, etc.) when processing a target task and a corresponding structured text, which specifically represents a token sequence corresponding to each confidence one-to-one matching sequence.

[0029] In the embodiment of the application, the confidence analysis on the target task and the structured text to obtain a confidence token sequence includes: constructing an input prompt of the target task and the structured text; generating a token stream of the input prompt by using a preset basic language model; calculating a confidence score of each token in the token stream; constructing a confidence token sequence of the target task and the structured text according to the confidence score.

[0030] In detail, the input prompt is the content of the target task and the corresponding structured text in the financial insurance field after preprocessing, which is provided for the basic LLM to receive to drive it to generate a thought chain answer; the construction process receives the above-mentioned target task and structured text by the input processing module, and pre-processes the structured text by tokenization, entity recognition, clause structure marking, etc. to adapt to the subsequent model input, and then integrates the pre-processed structured text with the target task to form the input prompt required by the initial reasoning generation module, which provides a basis for the LLM to start the thought chain generation process.

[0031] Specifically, the preset basic language model refers to a basic LLM, such as GPT-o1, DeepSeek-R1, etc., which is used to receive an input prompt and drive it to generate an answer in a thought chain manner; the token stream refers to the continuous token sequence output by the basic language model in the process of generating an answer; after the input processing module pre-processes the target task and the corresponding structured text in the financial insurance field and constructs the input prompt, the initial reasoning generation module calls the preset basic language model, the basic language model receives the input prompt and starts the answer generation in a thought chain manner, and in this process, it outputs a continuous token sequence, that is, the token stream corresponding to the input prompt is formed.

[0032] Further, each word unit refers to a single token in the word unit stream output by the preset basic language model when generating an answer in the form of a thought chain after receiving an input prompt, that is, each token in the token sequence corresponding to the word unit stream; the confidence score refers to a quantitative indicator of the credibility output by the basic language model when generating each word unit; the initial reasoning generation module calls a preset basic language model, the basic language model receives the input prompt and starts answer generation in the form of a thought chain, and simultaneously records the confidence score output by the model for each word unit in the word unit stream at the token level while generating the continuous word unit stream.

[0033] Further, the confidence score output for each word unit in the word unit stream is recorded at the token level, and finally each word unit in the word unit stream is matched and associated with the corresponding confidence score to form a sequence composed of word units and corresponding confidence scores.

[0034] Exemplarily, in a medical scenario, the target task is to "judge whether a patient meets the metformin medication indication for type 2 diabetes", the structured text is the patient's medical record, the fasting blood glucose is 8.2 mmol / L, there is no abnormality in liver and kidney function, and the age is 45 years old, the input prompt is to integrate the task and the medical record text after tokenization and entity recognition preprocessing; a preset basic language model is called to receive the input prompt, a word unit stream containing "patient fasting blood glucose exceeds the standard... no medication contraindication... meets the medication indication" is generated, the confidence score of each word unit is recorded at the token level, and the word unit and the corresponding score are matched one by one to construct the confidence word unit sequence of the task and the medical record text.

[0035] Exemplarily, in a financial scenario, the target task is to "judge whether a user's car insurance claim application is acceptable", the structured text is the user's claim application description: "rain caused the engine of the vehicle to be flooded, the vehicle damage insurance was purchased, and there is no record of exemption from liability", the input prompt is to integrate the task and the application text after tokenization and entity recognition preprocessing; a preset basic language model is called to receive the input prompt, a word unit stream containing "user purchased vehicle damage insurance... rain caused the engine to be flooded... no exemption from liability... claim application is acceptable" is generated, the confidence score of each word unit is recorded at the token level, and the word unit and the corresponding score are matched one by one to construct the confidence word unit sequence of the claim judgment task and the application text.

[0036] In the embodiment of the application, the confidence score refers to a quantitative indicator of the credibility output by the basic language model when generating each word unit, including softmax probability, logits maximum value, etc.

[0037] In detail, the input processing module receives a target task in the financial insurance field and a corresponding structured text, and constructs an input prompt after preprocessing such as tokenization, entity recognition, and clause structure marking. The initial reasoning generation module calls a preset basic LLM to receive the input prompt and generates a word token stream in the form of a thinking chain, while recording the confidence score of each word token at the token level to form a confidence word token sequence.

[0038] S3, calculating the confidence drop amplitude of each word token based on the confidence score.

[0039] In the embodiment of the application, the confidence drop amplitude is an index for reflecting the confidence drop of each word token position in the automatic thinking chain division method based on confidence drop, and its calculation method includes two types: one is to subtract the confidence score of the previous word token from the confidence score of the current word token, and the formula is , wherein is the confidence of the current word token, is the confidence of the previous word token; and the other is to calculate the average value of the confidence difference of each adjacent word token in a local sliding window under the premise of setting the window size , and the formula is This index is a key basis for determining the step boundary in the subsequent recognition and reasoning process.

[0040] In the embodiment of the application, as shown in Figure 3 , the confidence drop amplitude of each word token is calculated based on the confidence score, and includes: S31, identifying the reference confidence value of each word token; S32, calculating the difference between the confidence score and the reference confidence value to obtain a confidence drop value; S33, proportionally converting the confidence drop value to obtain the confidence drop amplitude of each word token.

[0041] In detail, the reference confidence value is a reference benchmark for judging whether the confidence drop is significant, which is determined based on the confidence curve obtained after processing by the confidence trajectory sampling module when the adaptive threshold method is used to identify the potential step boundary, and specifically represented by the mean value and the standard deviation of the confidence curve; the reference confidence value of each word token is identified by recording the confidence score of each word token output by the initial reasoning generation module at the token level, and then performing smoothing processing and difference calculation on the confidence curve by the confidence trajectory sampling module to obtain the confidence drop amplitude of each word token position, and then identifying the mean value and the standard deviation of the above confidence curve by the decision point identification module through the adaptive threshold method.As a baseline confidence value, breakpoint conditions are then set based on this baseline, thereby completing the baseline value identification and application process on which the judgment of whether each lexical corresponds to the boundary of the potential step is based.

[0042] Specifically, the confidence level decrease value is the magnitude of the confidence level decrease. Confidence decrease is a core indicator used to measure the change in confidence at each lexical position and to provide a basis for identifying the "decision inflection point" in inference. Its calculation is based on the confidence scores recorded by the initial inference generation module at each token level during the LLM token generation process. The process of calculating the difference between the confidence score and the baseline confidence value to obtain the confidence decrease value involves the confidence trajectory sampling module smoothing and differentiating the aforementioned confidence curve. The confidence decrease value is obtained in two ways: first, by directly calculating the difference between the current lexical confidence and the previous lexical confidence; and second, by calculating the value within a set window size. The average of the confidence differences between adjacent words within the local sliding window, while the baseline confidence value corresponds to the mean of the confidence curve under the adaptive threshold method in the decision point recognition module. and standard deviation .

[0043] Furthermore, after the initial inference generation module records the confidence score of each lexical generated by LLM at the token level, the confidence trajectory sampling module first smooths and differs the generated confidence curve, and then determines the sliding window size for local sliding window scenarios. Calculate the value of the sliding window from the inside. arrive The difference in confidence scores between adjacent terms at each position Then the sum of these differences is divided by the sliding window size. The confidence level decrease of the corresponding word can be obtained by converting the ratio.

[0044] For example, in a medical scenario, the input text describes a patient's symptoms as "recurrent cough with chest pain for one week and night sweats" along with the task of "preliminary diagnosis of tuberculosis." After the input processing module preprocesses the text, the initial inference generation module uses LLM to generate diagnostic inference and records the confidence score for each term. The confidence trajectory sampling module smooths the curve and calculates the difference. The decision point identification module uses the mean of the confidence curve... and standard deviation As the baseline confidence score for each word, the difference between the confidence score of each word and the baseline is used to calculate the confidence decrease value. Then, the confidence decrease value is calculated according to the set sliding window size. The confidence decrease values ​​are averaged and proportionally converted to obtain the confidence decrease range for each word element. For example, the decrease range at the word element "needs to be combined with sputum test results" is significant and can be used as the step boundary for diagnostic reasoning.

[0045] For example, in the auto insurance claims review process within a financial context, the input includes a claim description of "rear-end collision resulting in frontal damage to the vehicle, reported within 24 hours of the accident" and the task of "determining whether the claim can be approved." After the input processing module preprocesses the text, the initial inference generation module uses LLM to generate review inference and records the confidence score of each term. The decision point identification module uses the mean of the confidence curve... and standard deviation As the baseline confidence score for each word, the difference between the confidence score of each word and the baseline is used to calculate the confidence decrease value. Then, the confidence decrease value is calculated according to the set sliding window size. The confidence decrease values ​​are averaged and proportionally converted to obtain the confidence decrease rate for each word element. For example, the decrease rate is significant at the word element "whether it meets the time limit for reporting a case", which can be used as the step boundary for review and reasoning.

[0046] S4. Identify the target partition boundary in the confidence word sequence based on the confidence decrease magnitude and the semantic information of the input text.

[0047] In this embodiment of the invention, the semantic information refers to the information obtained after preprocessing the input text in the financial insurance field, including the text structure after tokenization, entity information obtained by entity recognition, and clause framework information obtained by clause structure marking, etc.; the target partition boundary refers to the boundary of the thought chain steps that can reflect the "decision inflection point" in the thought chain reasoning process, which is finally determined after semantic verification based on the potential step boundary with significant confidence decrease in the confidence level lexical sequence and the boundary of the thought chain steps that reflects the "decision inflection point" in the thought chain reasoning process.

[0048] In this embodiment of the invention, reference is made to Figure 4 As shown, the step of identifying the target partition boundary in the confidence word sequence based on the confidence decrease magnitude and the semantic information of the input text includes: S41. Perform threshold filtering on the confidence level decrease magnitude to obtain the target decrease position of the confidence level word sequence; S42. Using the semantic information of the input text, perform target semantic verification on the target descent position to obtain the candidate partition boundary of the confidence word sequence; S43. Fine-tune the position of the candidate partition boundary according to the semantic structure of the input text to obtain the target partition boundary of the confidence word sequence.

[0049] In detail, the target descent position refers to the word position in the confidence word sequence where the confidence decrease meets the preset threshold screening condition and can be used as a potential boundary for subsequent thought chain steps. There are three ways to obtain this position by threshold screening of the confidence decrease: using a fixed threshold method; if the calculated confidence decrease of a certain word position is less than... ( If the threshold is fixed, then the position of that morpheme is the target descent position. Using an adaptive thresholding method, the average descent rate of the confidence curve is first calculated. and standard deviation Then determine the position of a certain word element. Is it less than ( (where the adjustment coefficient is used). If the condition is met, the word position is the target descent position. Using the Top-K method, the word positions that rank among the top K in the entire confidence word sequence in terms of confidence decrease are directly selected as the target descent positions. The target descent position can be determined from the confidence word sequence by using any of the above threshold filtering methods.

[0050] Specifically, the candidate partition boundary refers to the boundary that falls on a semantically reasonable cutting point after adjusting the target descent position using the semantic information of the input text, and can serve as a preliminary basis for dividing the steps of the thought chain. First, the semantic information obtained in the input text preprocessing stage is obtained, and then this semantic information is applied to the target descent position that has been determined through threshold screening. The position is then finely adjusted forward or backward so that the adjusted position conforms to semantic logic. The adjusted position is the candidate partition boundary of the confidence word sequence.

[0051] Furthermore, semantic structure refers to the structural information related to the semantics of the input text obtained after preprocessing in the financial and insurance field. Specifically, it includes the basic text structure formed by tokenization, the entity association structure obtained by entity recognition, and the clause framework structure obtained by clause structure tagging. The semantic structure information generated in the input text preprocessing stage is called, and the candidate partition boundary is judged based on this information to determine whether it falls on a semantically reasonable cutting point. If the position of the candidate partition boundary does not conform to the semantic logic, its position is adjusted forward or backward. The boundary obtained after adjustment that meets the semantic structure requirements is the target partition boundary of the confidence word sequence.

[0052] S5. Identify the structured logical units of the confidence word sequence based on the target boundary, and divide the thought chain sub-chain of the confidence word sequence based on the structured logical units.

[0053] In this embodiment of the invention, the structured logic unit refers to a number of sub-segments formed by the step segmentation and reorganization module after the target segmentation boundary is determined by the decision point identification module in the automatic thinking chain segmentation method based on confidence decrease. Each sub-segment corresponds to a sub-logic unit in the thinking chain reasoning process and can be used by the downstream reasoning module to generate a sub-Prompt or perform further processing.

[0054] In this embodiment of the invention, the structured logic unit for identifying the confidence lexical sequence based on the target partition boundary includes: The confidence word sequence is segmented using the target partition boundary to obtain word sub-segments; The semantic coherence of the lexical sub-segment is checked to obtain the semantic lexical sub-segment, and the internal logical intent of the semantic lexical sub-segment is analyzed to obtain the intent identifier lexical sub-segment. Identify the target intent of the intent identifier lexical segment, and perform structured encapsulation processing on the intent identifier lexical segment according to the target intent to obtain the structured logical unit of the confidence lexical sequence.

[0055] In detail, a word segment refers to several segments formed after the original confidence word sequence is segmented according to the target partition boundary. Each segment corresponds to a sub-logical unit in the thought chain reasoning and can be used by the downstream reasoning module to generate a sub-Prompt or for further processing. The input processing module receives the financial insurance task and corresponding text and performs preprocessing. Then, the initial reasoning generation module uses basic LLM to generate the answer and records the confidence score of each token output according to the token level to form a confidence word sequence. Subsequently, the confidence trajectory sampling module smooths the confidence curve and calculates the confidence decrease magnitude / slope at each token position. Then, the decision point identification module identifies the positions where the confidence decreases significantly as the target partition boundary using the fixed threshold method, adaptive threshold method, or Top-K method. Finally, the step segmentation and recombination module segments the confidence word sequence according to the identified target partition boundary to obtain word segments.

[0056] Specifically, semantic lexical segments refer to lexical segments whose boundaries fall on semantically reasonable cutting points and are semantically logically coherent after semantic coherence verification of the initially segmented lexical segments. Their formation relies on fine-tuning of the boundaries of lexical segments based on information such as clauses, entities, and sentence boundaries. Intent-identifying lexical segments refer to semantic lexical segments whose internal logical intent is analyzed to clarify the specific reasoning intent carried by each segment and to mark the corresponding intent. Based on the lexical segments obtained in the segmentation and recombination module, the semantic coherence of the lexical segments is verified using the clause information, entity information, and sentence boundary information in the financial insurance scenario, combined with the optional semantic verification function in the decision point identification module. This verifies the semantic coherence of the lexical segments using clause information, entity information, and sentence boundary information in the financial insurance scenario. Segments with possible semantic breaks or unreasonable boundaries are fine-tuned to meet the semantic coherence requirements, thus obtaining semantic lexical segments. Then, based on the processing logic of each sub-logic unit focusing on specific reasoning tasks in the downstream reasoning module, the internal content of the semantic lexical segments is analyzed to identify the specific reasoning target or functional positioning corresponding to each segment and add intent identifiers, thus obtaining intent-identifying lexical segments.

[0057] Furthermore, the target intent refers to the core purpose of the specific reasoning sub-task in the financial insurance scenario, which is already labeled in the intent identifier word segment. These intents directly correspond to the specific reasoning goals that each sub-logical unit needs to complete in the reasoning process of the thought chain. After obtaining the intent identifier word segment, the target intent it carries is first identified based on the existing intent labeling of the segment. Then, based on the target intent combined with the characteristics of the financial insurance task and the requirements of the downstream reasoning module for focused reasoning, the intent identifier word segment is structurally encapsulated. The segment content is organized into a unit form with clear reasoning direction and information focus according to its target intent, and finally a structured logical unit of confidence word sequence is formed. This unit can be directly used by the downstream reasoning module to call LLM or special sub-reasoning model for inference or supplementation within the sub-chain. At the same time, it also provides a structured foundation for the visualization and feedback module to present the logical context of the reasoning steps.

[0058] In this embodiment of the invention, the sub-chain of the thought chain refers to several segments that are divided into by the original token stream generated by the basic LLM according to the significant position of confidence decrease calculated by the confidence trajectory sampling module in the automatic thought chain segmentation method based on confidence decrease. Each sub-segment is a unit with independent sub-logic, which can be adapted to the step-by-step reasoning needs of tasks such as contract clause parsing and claims application review in the financial and insurance fields, and can generate sub-prompts or perform further inferences for downstream reasoning modules.

[0059] In this embodiment of the invention, the step of dividing the thought chain sub-chain of the confidence word sequence according to the structured logic unit includes: Candidate breakpoints of the confidence word sequence are identified using a preset breakpoint determination strategy; The candidate breakpoints are fine-tuned based on the semantic boundary information of the structured logic units to obtain the target breakpoints of the confidence word sequence; The confidence word sequence is divided according to the target breakpoint to obtain the thought chain sub-chain.

[0060] In detail, the preset breakpoint determination strategy is a method used to identify positions where confidence decreases significantly, specifically including the fixed threshold method, the adaptive threshold method, and the Top-K method. Candidate breakpoints refer to positions where confidence decreases significantly after the confidence trajectory sampling module smooths the token-level confidence scores recorded by the initial inference generation module and calculates the confidence decrease magnitude / slope d_t for each token position. These positions, identified by the preset breakpoint determination strategy, represent potential boundaries of thought chain steps. After the input processing module receives and preprocesses the financial insurance task and corresponding text, the initial inference generation module generates answers using basic LLM and records the confidence score for each token. The confidence trajectory sampling module processes the confidence curve and calculates d_t. Then, the preset breakpoint determination strategy is used to judge d_t, identifying candidate breakpoints with significant confidence decreases in the confidence term sequence. Furthermore, clause, entity, and sentence boundary information can be used to fine-tune the candidate breakpoints forward or backward as needed, ensuring they fall at semantically reasonable cutting points, thus completing the identification of candidate breakpoints.

[0061] Specifically, semantic boundary information refers to the clause information, entity information, and sentence-level boundary information obtained during the text preprocessing process corresponding to financial insurance tasks. This information can identify semantically reasonable cutting positions in the text. The target breakpoint refers to the boundary of the thought chain steps of the confidence word sequence that is finally determined after the candidate breakpoints are adjusted by semantic boundary information and fall on a semantically reasonable cutting point. It is the basis for subsequent thought chain segmentation. After identifying candidate breakpoints through a preset breakpoint judgment strategy, the semantic verification function in the decision point recognition module is called. Combined with the semantic boundary information obtained above, the position of the candidate breakpoints is fine-tuned forward or backward so that the fine-tuned breakpoints fall on semantically reasonable cutting points such as the end of the sentence or the edge of the clause, thereby obtaining the target breakpoint of the confidence word sequence.

[0062] Furthermore, after the input processing module receives and preprocesses the financial insurance task and corresponding text, the initial inference generation module generates answers using basic LLM and records the token sequence and corresponding confidence level, the confidence trajectory sampling module smooths the confidence curve and calculates the confidence decrease rate / slope at each token position, and the decision point identification module identifies candidate breakpoints through a preset breakpoint judgment strategy and fine-tunes the target breakpoint by combining semantic boundary information, the step segmentation and recombination module segments the original token stream generated by the initial inference according to the determined target breakpoint position. The segmented sub-segments are the sub-chains of the thought chain. Each sub-segment can be used as a sub-logical unit, and for each sub-segment, a sub-Prompt can be generated for the downstream inference module or further processed, thereby completing the process of dividing the confidence word sequence according to the target breakpoint to obtain the sub-chain of the thought chain.

[0063] As can be seen, the above scheme first performs structured processing on the input text corresponding to the target task, then conducts confidence analysis on the target task and the structured text to obtain a confidence word sequence and identify the confidence score of each word. Next, it calculates the decrease in confidence of each word based on the score, and combines the semantic information of the input text to identify the target partition boundary, finally dividing the structured logical units and thought chain sub-chains. It eliminates the need for manual annotation, solving the problems of high cost and lack of scalability associated with manual partitioning; partitioning based on confidence inflection points avoids semantic imbalance and information redundancy in rule-based partitioning, adapts to multiple tasks in finance and insurance, can be optimized based on feedback, has strong interpretability, and can solve the problem of low accuracy in text thought chain sub-chain partitioning.

[0064] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0065] In one embodiment, a text thinking chain sub-chain segmentation device 100 is provided, which corresponds one-to-one with the text thinking chain sub-chain segmentation method in the above embodiments. For example... Figure 5 As shown, the text thinking chain sub-chain segmentation device 100 includes an input text structuring processing module 101, a confidence analysis module 102, a confidence decrease calculation module 103, a target segmentation boundary recognition module 104, and a thinking chain sub-chain segmentation module 105. Detailed descriptions of each functional module are as follows: The input text structuring processing module 101 is used to obtain the input text corresponding to the target task, perform text structuring processing on the input text, and obtain structured text. The confidence analysis module 102 is used to perform confidence analysis on the target task and the structured text to obtain a confidence word sequence and identify the confidence score of each word in the confidence word sequence. The confidence decrease calculation module 103 is used to calculate the confidence decrease of each word based on the confidence score; The target segmentation boundary recognition module 104 is used to identify the target segmentation boundary in the confidence word sequence based on the confidence decrease magnitude and the semantic information of the input text; The sub-chain segmentation module 105 is used to identify the structured logical units of the confidence word sequence according to the target segmentation boundary, and to segment the sub-chains of the confidence word sequence according to the structured logical units.

[0066] In one embodiment, the text structuring processing module 101, when performing text structuring processing on the input text to obtain structured text, is used to: The input text is normalized to obtain the initial processed text; The key entities in the initial processed text are marked to obtain the intermediate entity text; Text structure annotation is performed on the intermediate entity text to obtain the target marked text; The entity information in the target marked text is associated and bound with the clause structure information to obtain the structured text of the input text.

[0067] In one embodiment, the confidence analysis module 102, when performing confidence analysis on the target task and the structured text to obtain a confidence term sequence, is used to: Construct input prompts for the target task and the structured text; The word stream of the input prompt is generated using a preset basic language model; Calculate the confidence score for each word in the word stream; Construct a confidence term sequence for the target task and the structured text based on the confidence score.

[0068] In one embodiment, the confidence decrease calculation module 103, when performing the calculation of the confidence decrease of each word based on the confidence score, is used to: Identify the baseline confidence value for each term; The confidence score is compared with the baseline confidence value to calculate the confidence decrease value. The confidence decrease value is proportionally converted to obtain the confidence decrease rate of each word.

[0069] In one embodiment, the target segmentation boundary recognition module 104, when performing the task of recognizing the target segmentation boundary in the confidence word sequence based on the confidence decrease magnitude and the semantic information of the input text, is configured to: The target decrease position of the confidence word sequence is obtained by thresholding the decrease in confidence level. Using the semantic information of the input text, the target descent position is subjected to target semantic verification to obtain the candidate partition boundary of the confidence word sequence; The candidate partition boundaries are fine-tuned based on the semantic structure of the input text to obtain the target partition boundaries of the confidence word sequence.

[0070] In one embodiment, the thought chain sub-chain partitioning module 105, when executing a structured logic unit that identifies the confidence lexical sequence based on the target partitioning boundary, is used to: The confidence word sequence is segmented using the target partition boundary to obtain word sub-segments; The semantic coherence of the lexical sub-segment is checked to obtain the semantic lexical sub-segment, and the internal logical intent of the semantic lexical sub-segment is analyzed to obtain the intent identifier lexical sub-segment. Identify the target intent of the intent identifier lexical segment, and perform structured encapsulation processing on the intent identifier lexical segment according to the target intent to obtain the structured logical unit of the confidence lexical sequence.

[0071] In one embodiment, the thought chain sub-chain partitioning module 105, when performing the partitioning of the thought chain sub-chain according to the structured logic unit into the confidence word sequence, is further configured to: Candidate breakpoints of the confidence word sequence are identified using a preset breakpoint determination strategy; The candidate breakpoints are fine-tuned based on the semantic boundary information of the structured logic units to obtain the target breakpoints of the confidence word sequence; The confidence word sequence is divided according to the target breakpoint to obtain the thought chain sub-chain.

[0072] This invention provides a device for dividing textual thought chain subchains. First, the input text corresponding to the target task is structured. Then, confidence analysis is performed on the target task and the structured text to obtain a confidence word sequence and identify the confidence score of each word. Next, the confidence decrease rate of each word is calculated based on the score. Combined with the semantic information of the input text, the target division boundary is identified, and finally, structured logical units and thought chain subchains are divided. It eliminates the need for manual annotation, solving the problems of high cost and lack of scalability associated with manual division. Division based on confidence inflection points avoids semantic imbalance and information redundancy in rule-based division, adapts to multiple tasks in finance and insurance, and can be optimized based on feedback. It has strong interpretability and can solve the problem of low accuracy in textual thought chain subchain division.

[0073] Specific limitations regarding the sub-chain division mechanism for text-based thinking chains can be found in the above description of the sub-chain division method, and will not be repeated here. Each module in the aforementioned sub-chain division mechanism can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0074] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements server-side functions or steps of a text-based thought chain sub-chain partitioning method.

[0075] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 7As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of a text-based thought chain sub-chain partitioning method.

[0076] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain the input text corresponding to the target task, and perform text structuring processing on the input text to obtain structured text; Confidence analysis is performed on the target task and the structured text to obtain a confidence word sequence, and the confidence score of each word in the confidence word sequence is identified. The confidence level decrease for each term is calculated based on the confidence score. Identify the target partition boundary in the confidence word sequence based on the confidence decrease magnitude and the semantic information of the input text; Based on the target boundary, identify the structured logical units of the confidence word sequence, and divide the thought chain sub-chain of the confidence word sequence according to the structured logical units.

[0077] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain the input text corresponding to the target task, and perform text structuring processing on the input text to obtain structured text; Confidence analysis is performed on the target task and the structured text to obtain a confidence word sequence, and the confidence score of each word in the confidence word sequence is identified. The confidence level decrease for each term is calculated based on the confidence score. Identify the target partition boundary in the confidence word sequence based on the confidence decrease magnitude and the semantic information of the input text; Based on the target boundary, identify the structured logical units of the confidence word sequence, and divide the thought chain sub-chain of the confidence word sequence according to the structured logical units.

[0078] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0079] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0080] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0081] It should be noted that if any software tools or components not belonging to our company appear in the embodiments of this application, they are merely for illustrative purposes and do not represent actual use.

[0082] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method of dividing a text thought chain sub-chain, characterized by, The method comprises the following steps: obtaining an input text corresponding to a target task, performing text structuring processing on the input text to obtain a structured text; performing confidence analysis on the target task and the structured text to obtain a confidence token sequence, and identifying a confidence score of each token in the confidence token sequence; calculating a confidence drop amplitude of each token according to the confidence score; identifying a target division boundary in the confidence token sequence according to the confidence drop amplitude and semantic information of the input text; identifying a structured logical unit of the confidence token sequence according to the target division boundary, and dividing a thinking chain sub-chain of the confidence token sequence according to the structured logical unit.

2. The method of claim 1, wherein the dividing of the text thought chain sub-chain is characterized by, The method of performing text structuring processing on the input text to obtain a structured text comprises the following steps: performing text normalization on the input text to obtain an initial processing text; labeling key entities in the initial processing text to obtain an intermediate entity text; performing text structure annotation on the intermediate entity text to obtain a target labeled text; associating and binding entity information and clause structure information in the target labeled text to obtain a structured text of the input text.

3. The method of claim 1, wherein the dividing of the text thought chain sub-chain is characterized by, The method of performing confidence analysis on the target task and the structured text to obtain a confidence token sequence comprises the following steps: constructing an input prompt of the target task and the structured text; generating a token stream of the input prompt by using a preset basic language model; calculating a confidence score of each token in the token stream; constructing a confidence token sequence of the target task and the structured text according to the confidence score.

4. The method of claim 1, wherein the dividing of the text thought chain sub-chain is characterized by, The method of calculating a confidence drop amplitude of each token according to the confidence score comprises the following steps: identifying a reference confidence value of each token; calculating a confidence drop value by difference between the confidence score and the reference confidence value; performing proportional conversion on the confidence drop value to obtain the confidence drop amplitude of each token.

5. The method of claim 1, wherein the dividing of the text thought chain sub-chain is characterized by, The method of identifying a target division boundary in the confidence token sequence according to the confidence drop amplitude and semantic information of the input text comprises the following steps: performing threshold screening on the confidence drop amplitude to obtain a target drop position of the confidence token sequence; performing target semantic verification on the target drop position by using semantic information of the input text to obtain a candidate division boundary of the confidence token sequence; performing position fine-tuning on the candidate division boundary according to a semantic structure of the input text to obtain a target division boundary of the confidence token sequence.

6. The method of claim 1, wherein the dividing of the text thought chain sub-chain is characterized by, The method of identifying a structured logical unit of the confidence token sequence according to the target division boundary comprises the following steps: segmenting the confidence token sequence by using the target division boundary to obtain a token subsegment; performing semantic coherence verification on the token subsegment to obtain a semantic token subsegment, and performing internal logical intent analysis on the semantic token subsegment to obtain an intent identification token subsegment; The target intent of the intent identification word element sub-segment is identified, and the intent identification word element sub-segment is processed by structural encapsulation according to the target intent, to obtain a structured logic unit of the confidence word element sequence.

7. The method of claim 1, wherein the dividing of the text thought chain sub-chain is characterized by, The method for dividing the thought chain sub-chain of the confidence word element sequence according to the structured logic unit comprises: A candidate breakpoint of the confidence word element sequence is identified by using a preset breakpoint judgment strategy; The position of the candidate breakpoint is fine-tuned according to semantic boundary information of the structured logic unit, to obtain a target breakpoint of the confidence word element sequence; The confidence word element sequence is divided according to the target breakpoint, to obtain a thought chain sub-chain.

8. A device for dividing sub-chains of a textual thought chain, characterized in that, Comprise: An input text structural processing module is configured to obtain an input text corresponding to a target task, and perform text structural processing on the input text to obtain a structured text; A confidence analysis module is configured to perform confidence analysis on the target task and the structured text to obtain a confidence word element sequence, and identify a confidence score of each word element in the confidence word element sequence; A confidence drop amplitude calculation module is configured to calculate a confidence drop amplitude of each word element according to the confidence score; A target division boundary identification module is configured to identify a target division boundary in the confidence word element sequence according to the confidence drop amplitude and semantic information of the input text; A thought chain sub-chain division module is configured to identify a structured logic unit of the confidence word element sequence according to the target division boundary, and divide the confidence word element sequence into a thought chain sub-chain according to the structured logic unit.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method for dividing the thought chain sub-chain of the text according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the method for dividing the thought chain sub-chain of the text according to any one of claims 1 to 7.

Images