A method and system for multi-modal dialogue sequence discovery in complex message flows

Abstract

This invention discloses a method and system for discovering multimodal dialogue sequences in complex message streams, belonging to the fields of natural language processing and artificial intelligence. To address the problems of high cost and poor generalization in existing technologies, this invention achieves its goal through the following approach: First, association rule matching is performed based on the publication attributes of each message in the message stream to construct message blocks with potential dialogue relationships within the message sequence; then, semantic disambiguation and modality alignment are performed on the multimodal information in the message blocks to generate structured descriptive information; finally, a comprehensive confidence assessment is conducted based on the structured descriptive information and message association features to determine the target dialogue sequence. This invention can accurately reconstruct dialogue logic in complex social scenarios, supports mixed modalities of text, images, and files, and requires no model fine-tuning, exhibiting strong cross-scenario applicability.

Description

Technical Field

[0001] This invention discloses a method and system for discovering multimodal dialogue sequences in complex message streams, belonging to the fields of natural language processing and artificial intelligence technology. Background Technology

[0002] A dialogue sequence refers to a logically structured and sequential set of dialogues formed by users through interactions on online forums and social media platforms. Dialogue sequences originate from the communication process of users "initiating a topic - responding to others - extending the topic" and the platform's interactive mechanisms such as "replying, forwarding, and commenting." In online forums, they often present a tree-like structure with the main topic as the root and multi-level comments as branches, while in social media, they may form a network structure with interwoven multi-party dialogues. Their characteristics include temporal sequence, semantic relevance, and the potential existence of implicit response relationships. Dialogue sequence discovery is a technical task in the field of natural language processing. Its core objective is to identify explicit or implicit response relationships between utterances from unstructured or semi-structured raw dialogue corpora through steps such as discourse segmentation, clue extraction, associative reasoning, and structure reconstruction, thereby discovering message combinations with dialogue relationships. The dialogue sequence discovery task is a prerequisite for downstream applications such as dialogue summarization, sentiment analysis, topic tracking, and intelligent customer service intent recognition, possessing significant theoretical research value and broad industrial application significance.

[0003] Existing dialogue sequence discovery tasks typically employ deep learning methods. For example, the paper "Who did they respond to? Conversation structure modeling using masked hierarchical transformer" proposes a neural network model based on a masked hierarchical transformer. This model can explicitly utilize the contextual information of the dialogue and achieves performance superior to baseline methods. "A deep sequential model for discourse parsing on multi-party dialogues" proposes a deep sequence model for multi-party dialogue sequences, defining different types of response relationships and predicting response relationships between dialogues through a joint alternation approach, suitable for multi-party, multi-turn dialogue scenarios. Chinese invention patent CN 119578536 A proposes a response relationship discovery method based on query window modeling. By optimizing dataset construction and neural network model training methods, it alleviates the problem of model inference results being biased towards negative samples due to the abundance of short texts and the scarcity of positive samples.

[0004] Existing technologies for solving the problem of dialogue sequence discovery typically employ deep learning-based methods. These methods require training specialized neural network models to learn the semantic representations of message text and the meanings constituting the dialogue relationships. This includes steps such as dataset construction, dialogue modeling, manual annotation, network structure design, and training.

[0005] First, building deep learning models requires a significant amount of manual work. Dialogue across different platforms differs significantly in data carriers, content formats, and user relationships. For example, online forums primarily use posts, with content consisting of long articles, rational analysis, and clashes of opinions, and weaker user social relationships. In contrast, social media uses dynamic posts and images, with content primarily consisting of fragmented messages and colloquial expressions, and stronger user social relationships. Therefore, it is necessary to construct datasets containing the dialogue characteristics of each platform. This requires manual data collection, cleaning, and labeling, consuming substantial human resources.

[0006] Secondly, the trained model has weak generalization ability. The training data and network structure of the model are customized and designed for the dialogue features of a specific platform. Therefore, when using this model to perform dialogue discovery tasks on other platforms, the features learned by the model cannot be adapted to the dialogue rules of other platforms due to the large difference in feature distribution between the training data and the target scene, resulting in weak versatility and generalization ability of the model.

[0007] Finally, most existing technologies only support processing text-based data. Since semantic information is needed to analyze the dialogue relationships between sentences, this semantic information is naturally embedded within the content of text data. In contrast, non-textual modal data such as images and documents require additional feature extraction and transformation operations to extract their corresponding semantic information. Therefore, due to considerations of controlling labor costs, existing technologies typically focus only on processing text-based single-modal data. Summary of the Invention

[0008] The purpose of this invention is to address the technical problems of high manpower costs, weak cross-scenario generalization ability, and insufficient multimodal data adaptability in existing dialogue discovery technologies. It proposes a method and system for discovering multimodal dialogue sequences in complex message streams, which can accurately extract cross-scenario and multimodal dialogue chains through the collaboration of rules and large language models, thereby reducing model development and maintenance costs and improving the comprehensiveness and accuracy of dialogue recognition.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] A method for discovering multimodal dialogue sequences in complex message streams includes the following steps: Based on the publication attributes of each message in the message stream to be processed, association rules are matched to construct message blocks with potential dialogue relationships in the message sequence; Semantic disambiguation and modality alignment are performed on the multimodal information in the message block to generate structured description information corresponding to the message block; A comprehensive confidence assessment is performed based on the structured description information and the correlation features between messages, and the target dialogue sequence is determined from the message block according to the assessment results.

[0011] Furthermore, association rule matching is performed based on the publication attributes of each message in the message stream to be processed, and message blocks with potential dialogue relationships are constructed in the message sequence, including: Determine the functional attributes of the target message from the target publisher in the message stream to be processed within the message sequence; Based on the functional attributes, a bidirectional expansion operation is performed on the target message to combine messages that conform to preset association rules with the target message to obtain a preliminary message block.

[0012] Furthermore, a bidirectional expansion operation is performed on the target message, including: If the function attribute is to open a dialogue, then the target message is expanded backward to retrieve the publisher of subsequent messages in the message sequence and the time interval between them and the target message; If the function attribute is a reply dialogue, then the target message is expanded forward to retrieve the publisher of the previous message in the message sequence and the time interval between the previous message and the target message.

[0013] Furthermore, semantic disambiguation and modality alignment are performed on the multimodal information in the message block to generate structured description information corresponding to the message block, including: Functional symbol replacement and link content summary processing are performed on the text information in the message block to achieve semantic disambiguation at the text dimension; Visual description transformation and content structure extraction are performed on the non-text information in the message block to achieve modal alignment of non-text information to text content.

[0014] Furthermore, visual description conversion and content structure extraction are performed on the non-text information in the message block, including: The various file types in the message block are uniformly converted into image sequences; The system identifies the visual categories of the image sequence and the original image information in the message block, and converts the image content into text description information based on a preset narrative template.

[0015] Furthermore, a comprehensive confidence assessment is performed based on the structured description information and the correlation features between messages, including: Analyze the logical relationships and content correspondence features among the messages in the structured description information to determine the semantic confidence level; Extract the publication time distribution features and publisher number features of messages in the message block, and calculate the rule confidence.

[0016] Furthermore, the logical relationships and content correspondence features between messages in the structured description information are analyzed to determine semantic confidence, including: Identify semantic progression or transition relationships between messages and extract semantic features used to characterize logical connections; Match the same object or topic mentioned in different messages and extract related features to characterize content relevance.

[0017] Further, the publication time distribution characteristics and publisher number characteristics of messages in the message block are extracted, and the rule confidence is calculated, including: Calculate the average time interval between messages within the message block, and determine the time penalty score based on the average time interval; The total number of publishers in the message block is counted, and the number of publishers is penalized based on the deviation of the total number of publishers from the preset number benchmark. The rule confidence score is obtained by subtracting the time penalty score and the number of people penalty score from the preset initial confidence score.

[0018] Further, based on the evaluation results, the target dialogue sequence is determined from the message block, including: The semantic confidence and the rule confidence are normalized to obtain the corresponding score components; The overall confidence level is obtained by weighted linear combination of each score component, and message blocks with an overall confidence level exceeding the preset confidence threshold are retained as the target dialogue sequence.

[0019] A multimodal dialogue sequence discovery system in complex message streams, comprising: The message chunking module is used to perform association rule matching based on the publication attributes of each message in the message stream to be processed, and to construct message chunks with potential dialogue relationships in the message sequence; The semantic alignment module is used to perform semantic disambiguation and modality alignment processing on the multimodal information in the message block, and generate structured description information corresponding to the message block; The dialogue sequence confirmation module is used to perform a comprehensive confidence assessment based on the structured description information and the correlation features between messages, and to determine the target dialogue sequence from the message block according to the assessment results.

[0020] The present invention has achieved the following beneficial effects.

[0021] 1. This invention employs a combination of general large-scale model application without fine-tuning and rule-based algorithm-based automated initial screening, avoiding dependence on labeling data from specific platforms and customized network training. This fundamentally breaks the closed loop of manual data processing and significantly reduces the overall human resource costs.

[0022] 2. This invention utilizes the general cognitive ability of a large language model combined with rule design that is weakly dependent on specific scenarios. Through a complementary mechanism of rule-based candidate selection and model semantic verification, it achieves "one-time deployment, multi-scenario adaptation," effectively solving the problems of poor transfer performance and weak generalization ability of existing technologies across different social platforms.

[0023] 3. This invention innovatively designs a visual description conversion operator and a file operation operator, which unifies and normalizes images, text, and office suite files into structured text representations, achieving deep fusion of cross-modal semantics, avoiding omissions in dialogue recognition due to ignoring non-textual information, and greatly improving the comprehensiveness of the solution.

[0024] 4. This invention breaks down the complex dialogue discovery task into fine-grained subtasks such as message block construction, modality alignment, semantic judgment, and comprehensive scoring. It leverages the rigor of rule-based algorithms and the flexibility of large models to effectively alleviate the illusion problem when large models process complex long sequences and improve the reliability of recognition results.

[0025] 5. This invention introduces a linear weighted evaluation mechanism of semantic confidence and rule confidence, which quantifies and scores message combinations from two dimensions: the logic of dialogue content and the distribution of spatiotemporal attributes, ensuring the high accuracy and effectiveness of the final output dialogue sequence.

[0026] 6. This invention aims to integrate the rigor of rule-based algorithms with the flexibility of general-purpose large language models, to structurally decompose the dialogue discovery task, and to select rule-based algorithms or general-purpose large language models according to the characteristics of each sub-task, ultimately designing a dialogue discovery method with scenario universality and support for multimodal data processing. Attached Figure Description

[0027] Figure 1 This is a flowchart of a multimodal dialogue sequence discovery method in a complex message stream, as illustrated in this embodiment. Figure 2 This is a block diagram of a multimodal dialogue sequence discovery system in a complex message stream, as shown in the embodiment. Detailed Implementation

[0028] To make the various technical features, advantages, or effects of the present invention more apparent and understandable, detailed descriptions are provided below through embodiments.

[0029] This invention discloses a method for discovering multimodal dialogue sequences in complex message streams, such as... Figure 1 As shown, the specific steps include:

[0030] Step S1: Perform association rule matching based on the publication attributes of each message in the message stream to be processed, and construct message blocks with potential dialogue relationships in the message sequence.

[0031] Specifically, all messages are reordered chronologically and stored as a sequential list, denoted as . .

[0032] In an optional embodiment of the present invention, step S1 may include: Step S11: Determine the functional attributes of the target message of the target publisher in the message sequence of the message stream to be processed.

[0033] In an optional embodiment of the present invention, a publisher is selected as denoted as Retrieve and record the publisher in the message. All messages, including those from the publisher in the sequence list All text messages, image messages, and attachment messages sent in the middle.

[0034] Specifically, select a publisher The sent message is denoted as Based on the message content, the message's function is determined to be either initiating or replying to a conversation. (The superscript is not included in this translation.) Indicates the sender of the message; subscript Indicates the message location identifier.

[0035] Step S12: Based on the functional attributes, perform a bidirectional expansion operation on the target message, combining messages that conform to preset association rules with the target message to obtain a preliminary message block.

[0036] In an optional embodiment of the present invention, step S12 may include: Step S121: If the function attribute is to open a dialogue, then expand the target message backward to retrieve the publisher of the subsequent message in the message sequence and the time interval between the target message and the subsequent message.

[0037] Specifically, if the message If the function attribute is to start a conversation, then it will be based on messages. Perform a backward expansion operation on the header message. The operation process is as follows: in the sequential list... View messages The subsequent news will be analyzed based on the release time and the publisher of the subsequent news.

[0038] In an optional embodiment of the present invention, the comprehensive judgment process is as follows: First, check the publisher of the subsequent message; if the publisher of the subsequent message is the publisher... I, on the other hand, analyze the situation; specifically, this step considers that users might break a complete sentence into multiple short sentences when sending a message. If multiple consecutive messages are sent by the sender... All published messages are included in the scope of analysis. If the publisher of subsequent messages is not the same as the publisher... I will then check the publication time of subsequent messages. If the subsequent message is the same as the initial message... If the transmission interval is less than a preset time threshold, the analysis is passed; otherwise, it is not passed. The preset time threshold can be 30 minutes.

[0039] Specifically, information is added to the message through analysis. Then, it is stored as a message block.

[0040] Step S122: If the function attribute is a reply dialogue, then the target message is expanded forward to retrieve the publisher of the previous message in the message sequence and the time interval between the previous message and the target message.

[0041] Specifically, if the message If the function attribute is to reply to a conversation, then it will be in the form of a message. Perform a forward expansion operation on the tail message of the table. The operation process is as follows: in the sequential list... View messages The preceding messages are analyzed based on their publication time and publisher, and this analysis process is consistent with the comprehensive analysis process in step S121. Messages that pass the analysis are then added to the message list. Previously, it was stored as a message block.

[0042] In an optional embodiment of the present invention, the dialogue is time-sensitive, and the above-described judgment logic ensures that the message interval between initiating the dialogue and receiving a reply is within a reasonable range. Steps S11 to S122 are repeated until all messages from the publisher have been processed.

[0043] Step S2: Perform semantic disambiguation and modality alignment processing on the multimodal information in the message block to generate structured description information corresponding to the message block.

[0044] Specifically, semantic disambiguation and modal alignment are performed on each text message, image message, and file message in the message block based on a multimodal large language model.

[0045] In an optional embodiment of the present invention, step S2 may include: Step S21: Perform functional symbol replacement and link content summary processing on the text information in the message block to achieve semantic disambiguation at the text dimension.

[0046] Specifically, for text messages, social media functional symbols appearing in the content are replaced with the same semantics. For text messages that reference external links, a large language model is used to summarize and simplify the content of the external links, and then concatenate them with the original content.

[0047] In an optional embodiment of the present invention, since hashtags and keywords are often used as topic tags in social media, hashtags are replaced with topics.

[0048] Step S22: Perform visual description transformation and content structure extraction on the non-text information in the message block to achieve modal alignment of non-text information to text content.

[0049] Specifically, a multimodal large language model is used to perform visual narration of non-textual information. By designing customized narrative templates to describe non-textual content, structured information is formed, thereby accurately and completely converting non-textual information into textual content.

[0050] In an optional embodiment of the present invention, step S22 may include: Step S221: Convert the multi-type file information in the message block into an image sequence.

[0051] Specifically, customized operations are performed based on file type. For PDF files, a script-based conversion library is used to convert the file into an image; for office suite files including documents, presentations, and spreadsheets, they are first converted to PDF files using a conversion tool, and then processed according to the PDF file processing method; the processing approach for other file types is the same.

[0052] Step S222: Identify the visual categories of the image sequence and the original image information in the message block, and convert the image content into text description information based on a preset narrative template.

[0053] Specifically, the images converted from files are processed according to a visual description conversion process to obtain text content. The visual category to which the image belongs is identified, and the corresponding prompt words are invoked.

[0054] In an optional embodiment of the present invention, if an image is identified as a screenshot of a document, a prompt word for processing document images is invoked to guide the large language model to focus on extracting structured information such as the document's title and main content. If the image is identified as a landscape image, a prompt word for processing non-text images is invoked to guide the large language model to describe the image's environmental information and people information.

[0055] Specifically, each message in the original message block is processed according to the text dimension, visual dimension, and file conversion dimension to achieve semantic disambiguation and modal alignment of the entire message block.

[0056] Step S3: Based on the structured description information and the correlation features between messages, a comprehensive confidence assessment is performed, and the target dialogue sequence is determined from the message block according to the assessment results.

[0057] Specifically, the message blocks are scored for reliability and effectiveness based on rules and a large language model, thereby obtaining a dialogue combination that has a dialogue relationship and contains effective information.

[0058] In an optional embodiment of the present invention, step S3 may include: Step S31: Analyze the logical relationships and content correspondence features between messages in the structured description information to determine the semantic confidence level.

[0059] Specifically, based on the content of messages in a message block, a large inference language model is used to analyze the logical connections and content correspondences between messages, determine whether they constitute a dialogue relationship, and provide semantic confidence scores. Semantic confidence can be directly provided by the large language model using prompt words, for example, with a score range of 1 to 10, where a higher score indicates higher confidence.

[0060] In an optional embodiment of the present invention, step S31 may include: Step S311: Identify the semantic progression or transition relationship between messages and extract semantic features used to characterize logical correlation.

[0061] Specifically, logical association refers to the deductive relationship between messages, such as progression or transition.

[0062] In an optional embodiment of the present invention, an example of logical association is that publisher A sends a message: "It's very windy outside today, and the temperature has dropped." Publisher B replies with a message: "Then I need to wear thicker clothes when I go out, otherwise I'll easily catch a cold."

[0063] Step S312: Match the same object or topic mentioned in different messages and extract the associated features used to characterize the content responsiveness.

[0064] Specifically, content echoing refers to different messages mentioning the same content.

[0065] In an optional embodiment of the present invention, an example of content echoing is that publisher A sends a message: I went to the park today and saw many cherry blossoms in full bloom, they were so beautiful; publisher B replies with a message: The cherry blossoms are indeed very beautiful, pink and white, and when the wind blows, it looks like snow.

[0066] Step S32: Extract the publication time distribution features and publisher number features of messages in the message block, and calculate the rule confidence.

[0067] Specifically, based on the principle that the shorter the time interval between message releases and the fewer the number of publishers in a message block, the greater the likelihood of forming a dialogue relationship, a rule confidence score is given. .

[0068] In an optional embodiment of the present invention, step S32 may include: Step S321: Calculate the average time interval between messages within the message block, and determine the time penalty score based on the average time interval.

[0069] Step S322: Count the total number of publishers in the message block, and determine the number penalty score based on the deviation of the total number of publishers from the preset number benchmark.

[0070] Step S323: Based on the preset initial confidence value, subtract the time penalty score and the number of people penalty score to obtain the rule confidence.

[0071] Specifically, rule confidence The calculation formula is as follows: when hour, ; In other cases, .

[0072] in, This is the time penalty coefficient; The average time interval between messages; Penalty coefficient based on the number of people; For the number of people. In actual implementation, it can be set to... It is 0.5. It is 10.

[0073] In an optional embodiment of the present invention, when the number of people When the score is less than 2 (i.e., equal to 1), it means that all the messages in the message block were sent by a single publisher, and it is determined that the message does not belong to a dialogue. The rule confidence score is recorded as 0. When the calculated rule confidence score is less than 0, it is also recorded as 0.

[0074] Step S33: Normalize the semantic confidence and the rule confidence to obtain the corresponding score components.

[0075] Specifically, semantic confidence is considered jointly. With rule confidence Calculate the overall confidence score. First, perform normalization, as follows: in, Represents the normalized semantic components; This represents the normalized rule component; the values ​​10 and 100 represent the preset maximum values ​​for semantic confidence and rule confidence, respectively.

[0076] Step S34: Perform a weighted linear combination of each score component to obtain the overall confidence level, and retain message blocks whose overall confidence level exceeds the preset confidence level threshold as the target dialogue sequence.

[0077] Specifically, weights are assigned and linear combinations are performed to obtain the overall confidence level. The calculation formula is as follows: in, It is the semantic confidence weight; This refers to the rule confidence weight. The relevant parameters satisfy: , ,and .

[0078] In an optional embodiment of the present invention, the overall confidence level exceeding a confidence threshold is retained. The message block. Confidence threshold. The value range is 0 to 1, and it can be set to 0.8 in practice.

[0079] This invention also provides a multimodal dialogue sequence discovery system in complex message streams, such as... Figure 2 As shown, it includes: The message chunking module is used to perform association rule matching based on the publication attributes of each message in the message stream to be processed, and to construct message chunks with potential dialogue relationships in the message sequence; The semantic alignment module is used to perform semantic disambiguation and modality alignment processing on the multimodal information in the message block, and generate structured description information corresponding to the message block; The dialogue sequence confirmation module is used to perform a comprehensive confidence assessment based on the structured description information and the correlation features between messages, and to determine the target dialogue sequence from the message block according to the assessment results.

[0080] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the protection scope of the present invention, which is defined by the claims.

Claims

1. A method for multi-modal dialogue sequence discovery in complex message flow, characterized in that, Includes the following steps: Based on the publication attributes of each message in the message stream to be processed, association rules are matched to construct message blocks with potential dialogue relationships in the message sequence; Semantic disambiguation and modality alignment are performed on the multimodal information in the message block to generate structured description information corresponding to the message block; A comprehensive confidence assessment is performed based on the structured description information and the correlation features between messages, and the target dialogue sequence is determined from the message block according to the assessment results.

2. The method as described in claim 1, characterized in that, Based on the publication attributes of each message in the message stream to be processed, association rules are matched to construct message blocks with potential dialogue relationships in the message sequence, including: Determine the functional attributes of the target message from the target publisher in the message stream to be processed within the message sequence; Based on the functional attributes, a bidirectional expansion operation is performed on the target message to combine messages that conform to preset association rules with the target message to obtain a preliminary message block.

3. The method as described in claim 2, characterized in that, Performing a bidirectional expansion operation on the target message includes: If the function attribute is to open a dialogue, then the target message is expanded backward to retrieve the publisher of subsequent messages in the message sequence and the time interval between them and the target message; If the function attribute is a reply dialogue, then the target message is expanded forward to retrieve the publisher of the previous message in the message sequence and the time interval between the previous message and the target message.

4. The method as described in claim 1, characterized in that, Semantic disambiguation and modality alignment are performed on the multimodal information in the message block to generate structured description information corresponding to the message block, including: Functional symbol replacement and link content summary processing are performed on the text information in the message block to achieve semantic disambiguation at the text dimension; Visual description transformation and content structure extraction are performed on the non-text information in the message block to achieve modal alignment of non-text information to text content.

5. The method as described in claim 4, characterized in that, Visual description transformation and content structure extraction are performed on the non-text information in the message block, including: The various file types in the message block are uniformly converted into image sequences; The system identifies the visual categories of the image sequence and the original image information in the message block, and converts the image content into text description information based on a preset narrative template.

6. The method as described in claim 1, characterized in that, A comprehensive confidence assessment is performed based on the structured description information and the correlation features between messages, including: Analyze the logical relationships and content correspondence features among the messages in the structured description information to determine the semantic confidence level; Extract the publication time distribution features and publisher number features of messages in the message block, and calculate the rule confidence.

7. The method as described in claim 6, characterized in that, Analyze the logical relationships and content correspondence features between messages in the structured description information to determine semantic confidence, including: Identify semantic progression or transition relationships between messages and extract semantic features used to characterize logical connections; Match the same object or topic mentioned in different messages and extract related features to characterize content relevance.

8. The method as described in claim 6, characterized in that, Extract the publication time distribution characteristics and publisher number characteristics of messages in the message block, and calculate the rule confidence, including: Calculate the average time interval between messages within the message block, and determine the time penalty score based on the average time interval; The total number of publishers in the message block is counted, and the number of publishers is penalized based on the deviation of the total number of publishers from the preset number benchmark. The rule confidence score is obtained by subtracting the time penalty score and the number of people penalty score from the preset initial confidence score.

9. The method as described in claim 1, characterized in that, Based on the evaluation results, the target dialogue sequence is determined from the message block, including: The semantic confidence and the rule confidence are normalized to obtain the corresponding score components; The overall confidence level is obtained by weighted linear combination of each score component, and message blocks with an overall confidence level exceeding the preset confidence threshold are retained as the target dialogue sequence.

10. A multimodal dialogue sequence discovery system in a complex message stream, characterized in that, include: The message chunking module is used to perform association rule matching based on the publication attributes of each message in the message stream to be processed, and to construct message chunks with potential dialogue relationships in the message sequence; The semantic alignment module is used to perform semantic disambiguation and modality alignment processing on the multimodal information in the message block, and generate structured description information corresponding to the message block; The dialogue sequence confirmation module is used to perform a comprehensive confidence assessment based on the structured description information and the correlation features between messages, and to determine the target dialogue sequence from the message block according to the assessment results.

Images