Method and device for processing multi-turn dialogue modifier coverage, and storage medium

Abstract

The application provides a multi-round dialogue modifier coverage processing method and device and a storage medium. The method comprises the following steps: obtaining multi-round dialogue text data input by a user according to a preset interaction interface; preprocessing the multi-round dialogue text data to obtain standard text data after processing; performing part-of-speech tagging and dependency relationship analysis on the standard text data according to a preset natural language processing tool to obtain corresponding part-of-speech tagging results and dependency relationship information; determining a plurality of candidate modifiers according to the part-of-speech tagging results, the dependency relationship information and the standard text data; obtaining a pre-constructed large language model, and performing disambiguation processing on each candidate modifier according to the large language model to obtain a plurality of target modifiers corresponding to each candidate modifier and a category to which each target modifier belongs; and determining a corresponding comprehensive coverage according to each target modifier and the category to which the target modifier belongs. The method provided by the application achieves the effect of improving the language expression richness and personalized matching degree of a multi-round dialogue system.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method, device and storage medium for processing the coverage of modifiers in multi-turn dialogues. Background Technology

[0002] In multi-turn dialogue scenarios, users increasingly demand natural, personalized, and linguistically diverse interactive experiences. However, current dialogue systems generally rely on Large Language Models (LLMs) to generate responses, but their outputs often suffer from a lack of stylistic uniformity and limited use of modifiers, resulting in a lack of depth and personalized expression in the generated content. This makes users perceive the interaction as mechanical and reduces user satisfaction. Therefore, developing a method to improve the coverage of modifiers in multi-turn dialogues is a promising direction for enhancing user and system interaction satisfaction.

[0003] In existing technologies, methods for handling modifier coverage in multi-turn dialogues mainly include: rule-based or template-based methods, which control generation through preset language templates; context modeling based on deep learning, which uses models such as recurrent neural networks (RNN) and transformers to semantically model the dialogue history; evaluation index systems that rely on text matching metrics or external feedback, such as bilingual evaluation understudy (BLEU) and recall-oriented understudy for gisting evaluation (ROUGE); and personalized recommendations and user preference modeling that analyze user history behavior through collaborative filtering or deep learning.

[0004] However, existing technologies are limited by factors such as a lack of quantitative analysis of modifier usage, insufficient matching of language style with user preferences, and a lack of dynamic optimization mechanisms, resulting in technical problems that make it difficult to meet users' needs for a natural, diverse, and personalized dialogue experience. Summary of the Invention

[0005] The method, device, and storage medium for processing the coverage of modifiers in multi-turn dialogues provided in this application are intended to improve the richness of language expression and the degree of personalized matching in multi-turn dialogue systems.

[0006] Firstly, this application provides a method for processing modifier coverage in multi-turn dialogues, including:

[0007] Based on the preset interactive interface, obtain multi-turn dialogue text data input by the user;

[0008] Preprocess the multi-turn dialogue text data to obtain the processed standard text data;

[0009] Based on the preset natural language processing tools, part-of-speech tagging and dependency analysis are performed on standard text data to obtain the corresponding part-of-speech tagging results and dependency relationship information;

[0010] Based on part-of-speech tagging results, dependency relationship information, and standard text data, multiple candidate modifiers were identified;

[0011] Obtain a pre-built large language model;

[0012] Based on the large language model, disambiguation is performed on each candidate modifier to obtain multiple target modifiers and the category to which each target modifier belongs;

[0013] Determine the overall coverage rate based on each target modifier and its category.

[0014] Secondly, this application provides a processing apparatus for multi-turn dialogue modifier coverage, comprising:

[0015] The first acquisition module is used to acquire multi-turn dialogue text data input by the user based on a preset interactive interface;

[0016] The preprocessing module is used to preprocess multi-turn dialogue text data to obtain processed standard text data;

[0017] The first processing module is used to perform part-of-speech tagging and dependency analysis on standard text data according to preset natural language processing tools, so as to obtain the corresponding part-of-speech tagging results and dependency information.

[0018] The second processing module is used to determine multiple candidate modifiers based on part-of-speech tagging results, dependency relationship information, and standard text data;

[0019] The second acquisition module is used to acquire pre-built large language models;

[0020] The third processing module is used to perform disambiguation processing on each candidate modifier based on the large language model, so as to obtain multiple target modifiers and the category to which each target modifier belongs;

[0021] The fourth processing module is used to determine the corresponding comprehensive coverage rate based on each target modifier and its category.

[0022] Thirdly, this application provides a processing device for multi-turn dialogue modifier coverage, including: a memory and a processor;

[0023] The memory stores the instructions that the computer executes;

[0024] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0025] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0026] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0027] This application provides a method, device, and storage medium for processing modifier coverage in multi-turn dialogues. It acquires multi-turn dialogue text data to provide foundational information for subsequent processing, preprocessing the data to transform it into standard text for unified analysis. Part-of-speech tagging and dependency analysis clarify the role and relationship of words in sentences, thereby identifying candidate modifiers. Using a large language model to disambiguate candidate modifiers accurately yields multiple target modifiers and their categories, avoiding misunderstandings caused by multiple meanings of a single word. Finally, based on the target modifiers and their categories, a comprehensive coverage rate is determined, measuring the overall usage of modifiers in multi-turn dialogues. These steps work together to improve the richness of language expression in multi-turn dialogue systems, making responses more tailored to user needs and enhancing personalized matching, thus achieving the technical effect of improving the richness of language expression and personalized matching in multi-turn dialogue systems. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0029] Figure 1 This application provides a schematic diagram of an application data processing system architecture.

[0030] Figure 2 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 1 ;

[0031] Figure 3 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 2 ;

[0032] Figure 4 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 3 ;

[0033] Figure 5 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 4 ;

[0034] Figure 6 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 5 ;

[0035] Figure 7 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 6 ;

[0036] Figure 8 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 7 ;

[0037] Figure 9 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 8 ;

[0038] Figure 10 A schematic diagram of the structure of the processing device for multi-turn dialogue modifier coverage provided in the embodiments of this application;

[0039] Figure 11 A schematic diagram of the structure of a processing device for multi-turn dialogue modifier coverage provided in an embodiment of this application.

[0040] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0041] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0042] Due to limitations such as a lack of quantitative analysis of modifier usage, insufficient matching of language style with user preferences, and a lack of dynamic optimization mechanisms, existing technologies have technical problems that make it difficult to meet users' needs for a natural, diverse, and personalized dialogue experience.

[0043] To address the aforementioned issues, this application provides a method, device, and storage medium for processing modifier coverage in multi-turn dialogues. It acquires multi-turn dialogue text data to provide foundational information for subsequent processing, preprocesses the data to convert it into standard text for unified analysis, and uses part-of-speech tagging and dependency analysis to clarify the role and relationship of words in sentences, thereby identifying candidate modifiers. Using a large language model to disambiguate candidate modifiers accurately yields multiple target modifiers and their categories, avoiding misunderstandings caused by multiple meanings of a single word. Finally, based on the target modifiers and their categories, a comprehensive coverage rate is determined, measuring the overall usage of modifiers in multi-turn dialogues. These steps work together to enhance the richness of language expression in multi-turn dialogue systems, making responses more tailored to user needs and improving personalized matching, thus achieving the technical effect of improving the richness of language expression and personalized matching in multi-turn dialogue systems.

[0044] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0045] Figure 1 This is a schematic diagram of an application data processing system architecture provided in an embodiment of this application. The application data processing system is a computer device. Figure 1 As shown, the above architecture includes at least one of a data acquisition device 101, a processing device 102, and a display device 103.

[0046] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the architecture of the application data processing system. In other feasible embodiments of this application, the above architecture may include more or fewer components than illustrated, or combine some components, or split some components, or arrange different components, which can be determined according to the actual application scenario and is not limited here. Figure 1 The components shown can be implemented in hardware, software, or a combination of both.

[0047] In the specific implementation process, the data acquisition device 101 may include an input / output interface or a communication interface, and the data acquisition device 101 can be connected to the processing device through the input / output interface or the communication interface.

[0048] The processing device 102 can first acquire and preprocess multi-turn dialogue text through an interactive interface, and then use natural language processing tools to perform part-of-speech tagging and dependency analysis to determine candidate modifiers. Next, a pre-built large language model is introduced to disambiguate the candidate modifiers, obtain the target modifiers and their categories, and finally determine the comprehensive coverage based on this, thereby improving the system's language expression richness and personalized matching degree.

[0049] The display device 103 can also be a touch screen or the screen of a terminal device, used to receive user commands while displaying the above-mentioned content, so as to realize interaction with the user.

[0050] It should be understood that the aforementioned processing device can be implemented by a processor reading instructions from memory and executing those instructions, or it can be implemented by a chip circuit.

[0051] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0052] Figure 2 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method for processing modifier coverage in multi-turn dialogues provided in this embodiment includes:

[0053] S201. Based on the preset interactive interface, obtain the multi-turn dialogue text data input by the user.

[0054] Based on a pre-defined interactive interface, acquire multi-turn dialogue text data generated during user interactions with these systems. This multi-turn dialogue text data includes user input and system responses.

[0055] S202. Preprocess the multi-turn dialogue text data to obtain the processed standard text data.

[0056] Preprocessing operations are performed on the acquired multi-turn dialogue text data, including removing redundant symbols and noise information, performing desensitization processing, and segmenting excessively long text data into blocks. After these operations, standard text data with uniform format and meeting the requirements of subsequent processing is obtained.

[0057] S203. Using preset natural language processing tools, perform part-of-speech tagging and dependency analysis on standard text data to obtain the corresponding part-of-speech tagging results and dependency information.

[0058] Using pre-defined natural language processing tools, part-of-speech tagging (POS) and dependency analysis are performed on each word in the standard text data. The analysis outputs the POS tagging results for each word and the dependency relationship information between words.

[0059] S204. Based on the part-of-speech tagging results, dependency relationship information, and standard text data, determine multiple candidate modifiers.

[0060] By combining the obtained part-of-speech tagging results and dependency relationship information, and referring to the context of standard text data, words that meet the characteristics of modifiers are selected from the standard text data. These selected words are multiple candidate modifiers.

[0061] S205, Obtain the pre-built large language model.

[0062] In this embodiment, the large language model is a pre-trained large language model.

[0063] Obtain a pre-built large language model, which has been trained before acquisition and has the ability to understand and discriminate text context, thus meeting the needs of subsequent processing.

[0064] S206. Based on the large language model, disambiguation is performed on each candidate modifier to obtain multiple target modifiers and the category to which each target modifier belongs.

[0065] Each candidate modifier is input into a large language model. The large language model's contextual understanding capability is used to disambiguate the candidate modifiers and determine whether they are true modifiers in a specific context. Finally, multiple target modifiers are determined, and the category to which each target modifier belongs is clarified.

[0066] S207. Determine the corresponding overall coverage rate based on each target modifier and its category.

[0067] Based on each target modifier and its category, the overall coverage rate is calculated. The overall coverage rate measures the breadth and richness of modifier usage in multi-turn dialogues, which helps to improve the richness of language expression and the degree of personalized matching in multi-turn dialogue systems.

[0068] This application provides a method for processing the coverage of modifiers in multi-turn dialogues. It acquires multi-turn dialogue text data to provide foundational information for subsequent processing, preprocessing the data to transform it into standard text for unified analysis. Part-of-speech tagging and dependency analysis clarify the role and relationship of words in sentences, thereby identifying candidate modifiers. Using a large language model to disambiguate candidate modifiers accurately yields multiple target modifiers and their categories, avoiding misunderstandings caused by multiple meanings of a single word. Finally, based on the target modifiers and their categories, a comprehensive coverage rate is determined, measuring the overall usage of modifiers in multi-turn dialogues. This series of steps works together to improve the richness of language expression in multi-turn dialogue systems, making responses more tailored to user needs and enhancing personalized matching, thus achieving the technical effect of improving the richness of language expression and personalized matching in multi-turn dialogue systems.

[0069] Figure 3 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 2 ,like Figure 3 As shown, this embodiment, based on the above embodiment, provides supplementary explanations of the specific process for determining the overall coverage rate and some subsequent processes, including:

[0070] S301. Based on each target modifier and its category, determine the category coverage, type coverage, and distribution similarity of the target modifier.

[0071] Based on each target modifier and its category, three types of indicators are calculated: Category coverage is the proportion of the subset of modifier categories that actually appear in the dialogue to the total number of predefined modifier categories; Type coverage is the proportion of the number of intersections between the set of modifier types that appear in the dialogue and the set of modifier types in the reference corpus to the total number of modifier types in the reference corpus; Distribution similarity is calculated using Jensen-Shannon Divergence (JSD). First, the probability distribution of modifier types in the dialogue and the JSD value of the ideal distribution in the reference corpus are obtained, and then the result is obtained by subtracting the JSD value from the similarity value of 1.

[0072] For example, first count the total number of predefined modifier categories, then count the number of categories that actually appear in the dialogue, and the ratio of the two is the category coverage rate; extract the modifier types in the dialogue to form a set, and find the intersection with the reference set. The ratio of the number of intersections to the total number of reference sets is the type coverage rate; calculate the JSD value of the two distributions, and then obtain the distribution similarity.

[0073] S302. Based on category coverage, type coverage, and distribution similarity, the corresponding comprehensive coverage is obtained through a weighted fusion algorithm.

[0074] The weighted fusion algorithm takes category coverage, type coverage, and distribution similarity as inputs and sets weight coefficients for each of the three indicators. The weight coefficients can be adjusted according to specific application scenarios. The comprehensive coverage is obtained by multiplying each indicator by its corresponding weight coefficient and summing the results. This indicator can uniformly measure the diversity of modifier usage.

[0075] For example, if the application scenario is customer service dialogue, assuming the category coverage weight is 0.4, the type coverage weight is 0.3, and the distribution similarity weight is 0.3, the three indicators are multiplied by their respective weights, and the sum is the comprehensive coverage, achieving a unified evaluation across multiple dimensions.

[0076] S303. Extract the corresponding semantic features based on the multi-turn dialogue text data.

[0077] For example, using the Lightweight Multi-turn Transformer (LMT) model, contextual modeling is performed on multi-turn dialogue text data, and the extracted semantic features include user intent vector, topic vector, sentiment polarity and intensity, user preference vector, and semantic consistency score. Each feature is processed and calculated through a specific layer of the model.

[0078] S304. Obtain the initial dialogue generation strategy.

[0079] In this embodiment, the initial dialogue generation strategy is constructed based on the basic generation logic of the large language model or the preset domain rules of the corresponding application scenario.

[0080] The initial dialogue generation strategy is constructed based on two methods: the basic generation logic of the large language model and the preset domain rules corresponding to the application scenario. This strategy provides an initial guiding framework for subsequent dialogue generation.

[0081] For example, if the application scenario is personalized recommendation, the initial strategy is based on the default text generation logic of LLM, combined with the rules of recommendation domain such as "vivid language and avoidance of repetition", to clarify the initial generation direction and basic requirements.

[0082] S305. Based on the comprehensive coverage and semantic features, adjust the initial dialogue generation strategy to obtain the adjusted target dialogue generation strategy.

[0083] By comparing the overall coverage with the preset threshold and combining the extracted semantic features, the initial dialogue generation strategy is adjusted by using prompt word enhancement or decoding reweighting to make the adjusted target dialogue generation strategy adaptable to the diversity of modifier usage and dialogue semantic needs.

[0084] Optionally, if the overall coverage is below the threshold, by combining user preference vectors and sentiment polarity, the model can be enhanced with prompt words to add modifiers for missing categories, or the distribution of modifiers can be adjusted by decoding and reweighting to obtain a target strategy that meets the needs.

[0085] The method for processing the coverage of modifiers in multi-turn dialogues provided in this application embodiment achieves a unified measurement of modifier diversity by comprehensively quantifying the breadth, richness, and naturalness of modifier usage, accurately capturing core dialogue information and user-related features, providing a reasonable basic framework for dialogue generation, and making the generated target dialogue generation strategy more adaptable to scenario needs and user preferences through strategy adjustment, thereby improving the quality and adaptability of dialogue generation.

[0086] Figure 4 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 3 ,like Figure 4 As shown, this embodiment, based on the above embodiments, elaborates on the process of acquiring multi-turn dialogue text data, including:

[0087] S401. Collect multi-turn dialogue logs of users interacting with the user based on the preset interactive interface.

[0088] Based on the preset interactive interface, collect multi-turn dialogue logs generated during the user's interaction with the interactive interface. The multi-turn dialogue logs contain a complete language record of the user's interaction with the system.

[0089] S402. Based on the multi-turn dialogue log, extract multiple corresponding context fragments and determine the fragment length of each context fragment.

[0090] The conversation structure based on multi-turn dialogue logs is segmented and parsed to extract multiple corresponding context fragments. Each context fragment corresponds to a continuous dialogue content. At the same time, the fragment length of each context fragment is determined by counting the number of characters or words.

[0091] Optionally, segmented parsing should follow the natural flow of the dialogue, splitting logs into nodes based on round switching, and calculating segment length based on the actual valid language content, excluding interference from redundant spaces and other meaningless characters.

[0092] S403. Compare the length of each context segment with the preset length.

[0093] The preset length is a threshold set in advance based on the processing capacity of the large language model for subsequent processing. The length of each previously determined context segment is compared with the preset length one by one to determine whether the segment length exceeds the limit.

[0094] S404. For context segments exceeding the preset length, segment them according to preset semantic units to obtain multiple segmented context segments.

[0095] For context segments that are determined to exceed the preset length after comparison, they are segmented according to preset semantic units. Semantic units are set based on units that can independently express complete meaning, such as sentences and semantic topics. After segmentation, multiple context segments that meet the length requirements are obtained.

[0096] S405. Organize each segmented context fragment and each context fragment that does not exceed the preset length to obtain the corresponding multi-turn dialogue text data.

[0097] The multiple context fragments obtained after segmentation that meet the length requirements are combined with the original context fragments that do not exceed the preset length and organized according to the time sequence and logical relationship of the dialogue, ultimately forming a multi-turn dialogue text data with a clear structure and compliant length.

[0098] The method for processing the coverage of modifiers in multi-turn dialogues provided in this application collects multi-turn dialogue logs between users and the system, accurately splits the dialogue to obtain context fragments and determines their lengths, effectively filters out excessively long fragments, reasonably segments excessively long fragments to ensure semantic integrity, and standardizes and organizes them into qualified data, providing stable, complete and compliant basic data support for subsequent text processing and model analysis.

[0099] Figure 5 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 4 ,like Figure 5 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the process for obtaining standard text data, including:

[0100] S501. Determine the corresponding session structure based on the multi-turn dialogue log.

[0101] Based on the order of user-system interactions and the logic of round connections in the multi-round dialogue logs, the corresponding conversation structure is determined. This structure clarifies the organization of the dialogue and the relationship between each part, providing a basis for subsequent segmentation processing.

[0102] S502. Based on the conversation structure, segment the multi-turn dialogue log to obtain the corresponding dialogue turn information, user input information, system response information, and timestamp information.

[0103] Based on the defined conversation structure, the multi-turn dialogue logs are segmented into complete interaction rounds, with each segment corresponding to a dialogue round. From each segment, the dialogue round information, the specific content of the user input, the specific content of the system response, and the timestamp information of the interaction round are extracted.

[0104] S503. Based on the dialogue turn information, user input information, system response information, and timestamp information, remove redundant symbols and noise information from the multi-turn dialogue text data to obtain the corresponding first text data.

[0105] In this embodiment, redundant symbols include spaces, line breaks, and paragraph marks, and noise information includes headers, footers, and system prompts.

[0106] Based on the extracted dialogue turn information, user input information, system response information, and timestamp information, the core interactive text in the multi-turn dialogue log is focused on, and redundant symbols such as spaces, line breaks, and paragraph marks are removed. At the same time, irrelevant noise information such as headers, footers, and system prompts is removed, and the first text data is obtained after processing.

[0107] S504. Based on the first text data, determine the corresponding sensitive information and perform desensitization processing on the sensitive information to obtain the corresponding standard text data.

[0108] The first text data is thoroughly screened to identify sensitive information such as user identity information. Security processing methods such as replacement and masking can be used to de-identify this sensitive information to prevent leakage. The resulting text is the standard text data.

[0109] The method for processing the coverage of modifiers in multi-turn dialogues provided in this application provides a clear basis for segmented processing by clarifying the conversation structure of multi-turn dialogue logs, accurately extracting key interaction information to ensure data integrity, removing redundant symbols and noise information to purify data content, and desensitizing sensitive information to ensure data security. Finally, it obtains standardized, secure, and effective standard text data, providing high-quality data support for subsequent modifier recognition and semantic analysis.

[0110] Figure 6 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 5 ,like Figure 6 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the specific process for determining the category coverage, type coverage, and distribution similarity of the target modifier, including:

[0111] S601. Obtain the pre-constructed set of modifier categories.

[0112] In this embodiment, the set of modifier categories is constructed based on linguistic rules or domain knowledge.

[0113] Obtain a pre-constructed set of modifier categories. This set is constructed based on linguistic rules or specific domain knowledge and covers common modifier categories such as degree adverbs, sentiment modifiers, and mood modalities, providing a unified reference standard for subsequent category coverage calculations.

[0114] S602. Determine the number of categories of different categories of modifiers in the target modifier, and calculate the ratio of the number of categories to the total number of categories in the modifier category set to obtain the category coverage rate.

[0115] The number of different categories in the target modifier is counted, that is, the total number of modifier categories actually appearing in the current dialogue. The number of these categories is divided by the total number of categories in the modifier category set. The resulting ratio is the category coverage rate, which measures the breadth of modifier category coverage. The specific formula for category coverage rate is as follows:

[0116]

[0117] in, Indicates category coverage. This represents a subset of modifier categories that actually appear in the current multi-turn dialogue text. This represents the total number of categories, i.e., the size of the reference set.

[0118] For example, assuming the total number of categories in the modifier category set is 8, and there are actually 4 different categories in the target modifier, the category coverage rate is calculated by dividing 4 by 8, which intuitively reflects the category coverage.

[0119] S603. Construct a reference set of corresponding modifier categories based on a pre-set large-scale dialogue dataset or domain standard dictionary.

[0120] Based on a pre-set large-scale dialogue dataset or domain standard dictionary, the specific types of various modifiers are extracted and organized into a modifier category reference set, which serves as a benchmark for measuring the richness of modifier types in the current dialogue.

[0121] S604. Construct a set of current dialogue modifier categories based on each target modifier.

[0122] Each target modifier is treated as an independent modifier type. All target modifiers are aggregated, and duplicate types are removed to form a set of modifier categories for the current dialogue. This set fully represents the specific types of modifiers actually used in the current dialogue.

[0123] For example, if the target modifiers include "very", "extremely", and "pleasant", then the current dialogue modifier category set is {"very", "extremely", and "pleasant"}, accurately recording the modifier type composition of the current dialogue.

[0124] S605. Calculate the intersection of the current dialogue modifier category set and the modifier category reference set, and calculate the ratio of the number of types in the intersection to the total number of types in the modifier category reference set to obtain the type coverage.

[0125] Find the modifier types that are the same in the current dialogue modifier category set and the modifier category reference set, count the number of these same types, and calculate the ratio of this number to the total number of types in the modifier category reference set. The result is the type coverage rate. The formula for calculating the type coverage rate is as follows:

[0126]

[0127] in, U represents type coverage. D U represents the set of modifier types appearing in the current dialogue; ref This represents the set of modifier types in the reference corpus; This indicates the number of modifier types shared by the dialogue and the reference set.

[0128] For example, assuming the total number of types in the modifier category reference set is 150 and the number of types in the intersection of the two sets is 30, then dividing 30 by 150 yields a type coverage rate of 0.2, which quantifies the richness of the modifier vocabulary.

[0129] S606. Based on the current set of dialog modifier categories, determine the frequency of occurrence of each type in the current set of dialog modifier categories, so as to determine the first probability distribution of the current dialog modifier type.

[0130] For each modifier type in the current dialogue modifier category set, count the number of times it appears in the multi-turn dialogue text data. Based on the ratio of the number of times each type appears to the total number of times all types appear, determine the frequency of each type, and thus form the first probability distribution of the current dialogue modifier type.

[0131] For example, suppose that “very” appears 8 times and “happy” appears 4 times in the current set, with a total occurrence count of 20. Then the frequency of “very” is 0.4 and that of “happy” is 0.2. The frequencies of all types are determined accordingly to form the first probability distribution.

[0132] S607. Obtain a pre-built reference corpus, and based on the reference corpus, obtain the second probability distribution of modifier types in the reference corpus that are set based on user preferences or domain standards.

[0133] Obtain a pre-built reference corpus containing a large amount of high-quality dialogue data. Based on user preferences formed from past interactions or domain-recognized standards for modifier usage, calculate the occurrence probability of each modifier type in the reference corpus and construct a second probability distribution for the modifier type.

[0134] S608. Based on the first probability distribution and the second probability distribution, determine the corresponding symmetric divergence, and take the difference between the first probability distribution and the symmetric divergence as the distribution similarity.

[0135] Based on the first and second probability distributions, the symmetric divergence between them is calculated. This symmetric divergence is the Jensen-Shannon divergence, with a value range between 0 and 1. The difference between 1 and this symmetric divergence is taken as the distribution similarity; the closer the value is to 1, the more natural the distribution. The formula for calculating the distribution similarity is:

[0136]

[0137]

[0138]

[0139] in, Indicates distribution similarity; This represents the first probability distribution of modifier types in the current dialogue, i.e., the frequency of each type. This represents the second probability distribution of modifier types in the reference corpus.

[0140] The method for processing the coverage of modifiers in multi-turn dialogues provided in this application provides a unified reference standard for modifier categories, quantifies the coverage breadth of modifier categories, constructs a comprehensive benchmark for comparing modifier types, clarifies the composition of modifier types in the current dialogue, measures the richness of modifier vocabulary, determines the frequency distribution of current modifier types, provides an ideal modifier distribution benchmark, quantifies the difference in naturalness between the current and ideal distributions, provides precise sub-indicators for calculating the overall coverage, and supports the optimization of subsequent dialogue generation strategies.

[0141] Figure 7 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 6 ,like Figure 7 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the semantic feature extraction process, including:

[0142] S701. Obtain a preset lightweight multi-turn dialogue converter model.

[0143] In this embodiment, the lightweight multi-turn dialogue converter model is based on the TransformerEncoder and integrates a linear attention mechanism, a context gating module, and a modifier awareness mechanism.

[0144] Obtain a pre-defined lightweight multi-turn dialogue converter model. This model is based on a converter encoder and integrates a linear attention mechanism, a context gating module, and a modifier awareness function, which can adapt to the semantic modeling requirements of multi-turn dialogue scenarios.

[0145] S702. Concatenate standard text data into a sequence format containing round embeddings according to the dialogue rounds.

[0146] According to the order of dialogue rounds, the user input information and system response information of each round in the standard text data are concatenated to form a sequence format. This sequence includes turn embedding, as well as word embedding and positional encoding, to monitor the contextual association of the dialogue.

[0147] S703. Input the sequence format into the lightweight multi-turn dialogue converter model for context modeling to generate the hidden state of each text tag.

[0148] The concatenated sequence format containing turn embeddings is input into a lightweight multi-turn dialogue converter model. This model processes the sequence through an embedding layer, and then performs context modeling through a linear attention encoder to capture the semantic associations and turn information of words in the sequence, and finally generates the hidden state corresponding to each text tag.

[0149] S704. Extract the corresponding semantic features based on the hidden state.

[0150] In this embodiment, the semantic features include user intent vector, topic vector, sentiment polarity and user preference vector. The user intent vector is taken from the hidden state of the sequence start marker. The topic vector is obtained by weighted average pooling of the hidden states by the context gating module. The user preference vector is generated based on the user's historical dialogue data through a preset multilayer perceptron (MLP) training.

[0151] Based on the hidden states of each text tag generated by the model, the corresponding semantic features are extracted. The user intent vector is taken from the hidden state of the sequence start tag (768 dimensions). The topic vector is obtained by weighting and filtering out noise words through the context gating module, and then average pooling the weighted hidden states. The sentiment polarity is obtained by outputting the category probability through a specific layer of the model task. The user preference vector is generated and dynamically updated based on the user's historical dialogue data through a preset multilayer perceptron.

[0152] Optionally, user intent vector It can be obtained through the following formula:

[0153]

[0154] Among them, LMT output This represents the output of the last layer of the linear attention encoder in the LMT model. It is the hidden state vector of the sequence start marker; this vector has a dimension of 768 and is directly used as a distributed representation of the intent.

[0155] Topic Vectors The calculation formula is:

[0156]

[0157] Among them, h i It is the hidden state vector of the i-th valid word (non-special label) in the sequence (from the last layer of LMT); N is the number of valid words; g i It is the context gating weight, calculated by the gating module.

[0158]

[0159] Where c is the historical context summary vector, It is the sigmoid function, W g These are trainable parameters.

[0160] Emotional polarity It is obtained through the following formula:

[0161]

[0162] Among them, W p and W v These are the trainable weights of the task-specific layer at the top of LMT. The sentiment polarity output class probability (e.g., positive / negative) has an L2 norm scalar strength.

[0163] User preference vector It can be obtained through the following formula:

[0164]

[0165] Among them, H history It is an LMT embedding sequence of the user's historical dialogues, and MLP is a multilayer perceptron. This vector is dynamically updated to reflect real-time preferences.

[0166] The method for processing the coverage of modifiers in multi-turn dialogues provided in this application provides an efficient modeling model that adapts to multi-turn dialogues, ensuring that the sequence format fully preserves the dialogue context and turn association. It generates accurate text tag hidden states through context modeling and comprehensively extracts semantic features that reflect the core information, theme, emotion and user preferences of the dialogue, providing high-quality semantic data support for subsequent dialogue generation strategy adjustments.

[0167] Figure 8 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 7 ,like Figure 8 As shown, this embodiment, based on the above embodiment, provides supplementary explanations of the subsequent process for obtaining the corresponding comprehensive coverage rate, including:

[0168] S801. Based on the comprehensive coverage and semantic features, construct a corresponding joint indicator system.

[0169] Based on comprehensive coverage and semantic features, and integrating relevant indicators such as semantic consistency score and user preference matching degree, a corresponding joint indicator system is constructed. This system can comprehensively and multidimensionally evaluate the quality and adaptability of dialogue generation, and provide a unified evaluation standard for subsequent strategy adjustment and analysis.

[0170] S802. Vectorize the comprehensive coverage, semantic features, and joint indicator system to obtain the corresponding vectorized results.

[0171] Vector encoding is used to transform the numerical value of comprehensive coverage, the vector representation of each semantic feature (user intent vector, topic vector, etc.), and the quantitative results of each indicator in the joint indicator system into vectorized results in a unified format, ensuring that the data can be compatible with subsequent storage and retrieval systems.

[0172] S803. Associate the vectorized results with the dialogue turn information and timestamp information, and store them in the pre-built semantic database and vector retrieval library to obtain a structured dataset.

[0173] In this embodiment, the table structure of the semantic database and vector retrieval library includes fields such as session number, round number, modifier category distribution, coverage value, semantic vector, user preference matching degree, and timestamp, which facilitates subsequent fast querying and data analysis.

[0174] The vectorized results are then associated and bound with the corresponding dialogue turn information and timestamp information to ensure that each piece of data can be traced back to a specific dialogue scenario. The associated data is then stored in a pre-built semantic database and vector retrieval library, ultimately forming a structured dataset with a standardized structure and clear associations.

[0175] S804. Obtain the indicator values ​​of each item in the joint indicator system and compare the indicator values ​​with the preset threshold values.

[0176] From the constructed joint indicator system, specific quantitative values ​​of each indicator are extracted. These values ​​reflect the performance of dialogue generation in terms of modifier usage, semantic expression, and user adaptation. Then, each indicator value is compared with a pre-set threshold value to determine whether the indicator meets the standard.

[0177] S805. If the indicator value is lower than the preset threshold, the target dialogue generation strategy is adjusted a second time by means of prompt word enhancement or decoding reweighting to obtain the adjusted target dialogue generation strategy.

[0178] In one possible implementation, after obtaining the adjusted target dialogue generation strategy, the method further includes: recording new dialogue interaction data generated after the second adjustment based on the adjusted target dialogue generation strategy, wherein the new dialogue interaction data includes adjusted semantic features, modifier usage information and user feedback information; and updating the structured dataset based on the new dialogue interaction data so that the updated structured dataset can be used to adjust the parameters of the lightweight multi-turn dialogue converter model and the large language model.

[0179] If one or more indicators are below the preset threshold, it indicates that the dialogue generation is insufficient in the corresponding dimension. The target dialogue generation strategy is then adjusted by using prompt word enhancement or decoding reweighting. After adjustment, new dialogue interaction data is recorded, including the adjusted semantic features, modifier usage information and user feedback information. This data is then used to update the structured dataset, and the parameters of the lightweight multi-turn dialogue converter model and the large language model are adjusted accordingly.

[0180] S806. Based on the structured dataset and the adjusted target dialogue generation strategy, generate a corresponding modifier coverage analysis report and semantic optimization suggestions.

[0181] Based on the updated structured dataset and the adjusted target dialogue generation strategy, the system automatically generates a corresponding modifier coverage analysis report and semantic optimization suggestions. The report and suggestions can be exported to portable document (PDF), spreadsheet software (Excel), and other formats.

[0182] The method for processing the coverage of modifiers in multi-turn dialogues provided in this application provides a comprehensive evaluation basis by constructing a unified joint indicator system, achieving standardized and compatible data storage and retrieval through vectorized processing, standardizing data association and management through structured datasets, accurately identifying shortcomings in dialogue generation through threshold comparison, continuously optimizing strategies and model performance through secondary adjustments and data updates, and providing practical guidance for business applications through analysis reports and optimization suggestions, thereby improving the overall quality, adaptability, and business usability of dialogue generation.

[0183] Figure 9 A flowchart illustrating the method for processing modifier coverage in multi-turn dialogues provided in this application embodiment. Figure 8 ,like Figure 9 As shown, this embodiment elaborates on the specific construction process of the joint indicator system based on the above embodiments, including:

[0184] S901. Based on the standard text data, determine the corresponding multi-turn dialogue context fragments.

[0185] Based on standard text data, and according to the turn-by-turn logic and semantic coherence of the dialogue, continuous and context-related dialogue content is extracted and identified as corresponding multi-turn dialogue context fragments. Each fragment must fully retain the user input and system response information of the relevant turn.

[0186] S902. Input the multi-turn dialogue context fragments into the lightweight multi-turn dialogue converter model to obtain the corresponding semantic feature-related hidden states.

[0187] The determined multi-turn dialogue context fragments are input into a lightweight multi-turn dialogue converter model. The model processes the data through an embedding layer, a linear attention encoder, and a context gating module to generate hidden states of each text tag related to semantic features.

[0188] For example, the lightweight multi-turn dialogue converter model first performs word embedding, position embedding and turn embedding on the multi-turn dialogue context fragments, and then calculates through an 8-layer linear attention encoder to output a 768-dimensional hidden state vector corresponding to each word.

[0189] S903. Perform vector aggregation on the semantic feature-related hidden states to obtain the average vector of the dialogue context.

[0190] Vector aggregation is performed on the semantic feature-related hidden states, excluding hidden states with special tags, and the total number of hidden states of effective words is counted. The sum of the hidden state vectors of all effective words is then divided by the number of effective words to obtain the average vector of the dialogue context.

[0191] S904. Obtain the system response text for the current round and input the system response text for the current round into the lightweight multi-turn dialogue converter model to obtain the corresponding current response vector.

[0192] Extract the system response text corresponding to the current turn from standard text data, ensuring that the text is complete and free of redundant information. Input the system response text into a lightweight multi-turn dialogue converter model, and obtain the corresponding current response vector after model encoding processing.

[0193] S905. Calculate the first cosine similarity between the average vector of the dialogue context and the current response vector, and use the first cosine similarity as the semantic consistency score.

[0194] The cosine similarity formula is used to calculate the first cosine similarity between the average vector of the dialogue context and the current response vector. The value of the first cosine similarity is directly used as the semantic consistency score to measure the degree of semantic relevance between the current response and the context.

[0195] Optionally, the range of the first cosine similarity is [-1, 1]. The closer the value is to 1, the closer the current response is to the logical connection of the dialogue context, and the more effectively it can identify semantic gaps or deviations from the topic.

[0196] S906. Based on the user preference vector in the semantic features, the system response text of the current round is encoded through a lightweight multi-turn dialogue converter model to obtain the semantic embedding vector of the system response.

[0197] Using the user preference vector in the semantic features as a reference, the system response text of the current round is input into a lightweight multi-turn dialogue converter model. Through the task-specific layer encoding processing of this model, the response text is mapped into a high-dimensional semantic embedding vector of the system response.

[0198] S907. Calculate the second cosine similarity between the user preference vector and the semantic embedding vector of the system response, and use the second cosine similarity as the user preference matching degree.

[0199] The second cosine similarity is calculated using the cosine similarity formula between the user preference vector in the semantic features and the semantic embedding vector of the system response. This similarity value is then determined as the user preference matching degree, quantifying the degree of fit between the response and the user's preferences.

[0200] Optionally, the range of the second cosine similarity is [-1, 1]. The closer the value is to 1, the more the language style and expression habits of the system's response match the user's past preferences, and the more accurately it reflects the personalized matching effect.

[0201] S908. The comprehensive coverage rate, semantic consistency score and user preference matching degree are weighted and integrated to construct a joint indicator system.

[0202] We set weight coefficients for the three indicators of comprehensive coverage, semantic consistency score and user preference matching degree to adapt to specific application scenarios. By multiplying the value of each indicator with the corresponding weight coefficient and summing them, we complete the weighted integration and finally build a comprehensive joint indicator system.

[0203] The method for processing the coverage of modifiers in multi-turn dialogues provided in this application provides basic data for semantic analysis by splitting out semantically related dialogue context fragments, supporting subsequent vector calculations through hidden states generated by the model, obtaining a stable average context vector through vector aggregation, ensuring the feasibility of similarity calculation through encoded response vectors, accurately quantifying semantic consistency and user preference matching degree through cosine similarity, and forming a multi-dimensional joint indicator system through weighted integration, providing a comprehensive and accurate evaluation basis for optimizing dialogue generation strategies.

[0204] Figure 10 This is a schematic diagram of the structure of a multi-turn dialogue modifier coverage processing device provided in an embodiment of this application. The device in this embodiment can be in the form of software and / or hardware. For example... Figure 10 As shown in the embodiment of this application, the processing device 1000 for multi-turn dialogue modifier coverage includes: a first acquisition module 1001, a preprocessing module 1002, a first processing module 1003, a second processing module 1004, a second acquisition module 1005, a third processing module 1006, and a fourth processing module 1007.

[0205] The first acquisition module 1001 is used to acquire multi-turn dialogue text data input by the user based on a preset interactive interface;

[0206] The preprocessing module 1002 is used to preprocess the multi-turn dialogue text data to obtain the processed standard text data.

[0207] The first processing module 1003 is used to perform part-of-speech tagging and dependency analysis on standard text data according to a preset natural language processing tool, so as to obtain the corresponding part-of-speech tagging results and dependency information.

[0208] The second processing module 1004 is used to determine multiple candidate modifiers based on part-of-speech tagging results, dependency relationship information, and standard text data;

[0209] The second acquisition module 1005 is used to acquire a pre-built large language model;

[0210] The third processing module 1006 is used to perform disambiguation processing on each candidate modifier according to the large language model, so as to obtain multiple target modifiers and the category to which each target modifier belongs;

[0211] The fourth processing module 1007 is used to determine the corresponding comprehensive coverage rate based on each target modifier and its category.

[0212] In one possible implementation, the fourth processing module 1007 is further configured to:

[0213] Based on each target modifier and its category, determine the category coverage, type coverage, and distribution similarity of the target modifier;

[0214] Based on category coverage, type coverage, and distribution similarity, a weighted fusion algorithm is used to obtain the corresponding comprehensive coverage.

[0215] In one possible implementation, the fourth processing module 1007 is further configured to:

[0216] Extract corresponding semantic features from multi-turn dialogue text data;

[0217] Obtain the initial dialogue generation strategy, which is constructed based on the basic generation logic of the large language model or the preset domain rules of the corresponding application scenario;

[0218] Based on the comprehensive coverage and semantic features, the initial dialogue generation strategy is adjusted to obtain the adjusted target dialogue generation strategy.

[0219] In one possible implementation, the first acquisition module 1001 is further configured to:

[0220] Based on the preset interactive interface, collect multi-turn dialogue logs during user interaction;

[0221] Based on the multi-turn dialogue logs, extract multiple corresponding context fragments and determine the fragment length of each context fragment;

[0222] Compare the length of each context segment with the preset length;

[0223] For context segments exceeding a preset length, they are segmented according to preset semantic units to obtain multiple segmented context segments;

[0224] Each segmented context fragment and each context fragment that does not exceed a preset length are organized to obtain the corresponding multi-turn dialogue text data.

[0225] In one possible implementation, the preprocessing module 1002 is further configured to:

[0226] Determine the corresponding session structure based on the multi-turn dialogue logs;

[0227] Based on the conversation structure, the multi-turn dialogue logs are segmented to obtain the corresponding dialogue turn information, user input information, system response information, and timestamp information;

[0228] Based on the dialogue turn information, user input information, system response information, and timestamp information, redundant symbols and noise information are removed from the multi-turn dialogue text data to obtain the corresponding first text data; among them, redundant symbols include spaces, line breaks, and paragraph marks, and noise information includes headers and footers and system prompts;

[0229] Based on the first text data, the corresponding sensitive information is identified and de-identified to obtain the corresponding standard text data.

[0230] In one possible implementation, the fourth processing module 1007 is further configured to:

[0231] Obtain a pre-built set of modifier categories, which is constructed based on linguistic rules or domain knowledge;

[0232] Determine the number of categories of different categories of modifiers in the target modifiers, and calculate the ratio of the number of categories to the total number of categories in the modifier category set to obtain the category coverage rate;

[0233] Construct a reference set of corresponding modifier categories based on a pre-set large-scale dialogue dataset or domain standard dictionary;

[0234] Construct a set of current dialogue modifier categories based on each target modifier;

[0235] Calculate the intersection of the current dialogue modifier category set and the modifier category reference set, and calculate the ratio of the number of types in the intersection to the total number of types in the modifier category reference set to obtain the type coverage rate;

[0236] Based on the current set of dialog modifier categories, determine the frequency of occurrence of each type in the current set of dialog modifier categories, so as to determine the first probability distribution of the current dialog modifier type;

[0237] Obtain a pre-built reference corpus, and based on the reference corpus, obtain the second probability distribution of modifier types in the reference corpus that are set based on user preferences or domain standards;

[0238] Based on the first probability distribution and the second probability distribution, the corresponding symmetric divergence is determined, and the difference between the first probability distribution and the symmetric divergence is taken as the distribution similarity.

[0239] In one possible implementation, the fourth processing module 1007 is further configured to:

[0240] Obtain a preset lightweight multi-turn dialogue converter model; the lightweight multi-turn dialogue converter model is based on a converter encoder and integrates a linear attention mechanism, a context gating module, and modifier awareness construction;

[0241] Standard text data is concatenated into a sequence format containing round embeddings based on the dialogue rounds;

[0242] The sequence format is input into a lightweight multi-turn dialogue converter model for context modeling to generate the hidden state of each text tag;

[0243] Based on the hidden state, the corresponding semantic features are extracted. The semantic features include user intent vector, topic vector, sentiment polarity and user preference vector. The user intent vector is taken from the hidden state of the sequence start marker. The topic vector is obtained by weighted average pooling of the hidden state by the context gating module. The user preference vector is generated by training a preset multilayer perceptron based on the user's historical dialogue data.

[0244] In one possible implementation, the fourth processing module 1007 is further configured to:

[0245] Based on the comprehensive coverage and semantic features, a corresponding joint indicator system is constructed;

[0246] The comprehensive coverage, semantic features, and joint indicator system are stored in a structured manner to obtain the corresponding structured dataset;

[0247] The target dialogue generation strategy is adjusted based on the joint indicator system to obtain the adjusted target dialogue generation strategy.

[0248] Based on the structured dataset and the adjusted target dialogue generation strategy, a corresponding modifier coverage analysis report and semantic optimization suggestions are generated.

[0249] In one possible implementation, the fourth processing module 1007 is further configured to:

[0250] Based on standard text data, determine the corresponding multi-turn dialogue context fragments;

[0251] Multi-turn dialogue context fragments are input into a lightweight multi-turn dialogue converter model to obtain the corresponding semantic feature-related hidden states;

[0252] Vector aggregation is performed on the semantically relevant hidden states to obtain the average vector of the dialogue context;

[0253] Obtain the system response text for the current round and input it into the lightweight multi-turn dialogue converter model to obtain the corresponding current response vector;

[0254] Calculate the first cosine similarity between the average vector of the dialogue context and the current response vector, and use the first cosine similarity as the semantic consistency score;

[0255] Based on the user preference vector in the semantic features, the system response text of the current round is encoded through a lightweight multi-turn dialogue converter model to obtain the semantic embedding vector of the system response;

[0256] Calculate the second cosine similarity between the user preference vector and the semantic embedding vector of the system response, and use the second cosine similarity as the user preference matching degree;

[0257] We will weight and integrate comprehensive coverage, semantic consistency score and user preference matching degree to construct a joint indicator system.

[0258] In one possible implementation, the fourth processing module 1007 is further configured to:

[0259] The comprehensive coverage, semantic features, and joint indicator system are vectorized to obtain the corresponding vectorized results;

[0260] The vectorized results are associated with dialogue turn information and timestamp information and stored in a pre-built semantic database and vector retrieval library to obtain a structured dataset.

[0261] In one possible implementation, the fourth processing module 1007 is further configured to:

[0262] Obtain the indicator values ​​of each item in the joint indicator system and compare the indicator values ​​with the preset threshold values;

[0263] If the indicator value is lower than the preset threshold, the target dialogue generation strategy will be adjusted a second time by enhancing prompt words or re-weighting decoding to obtain the adjusted target dialogue generation strategy.

[0264] In one possible implementation, the fourth processing module 1007 is further configured to:

[0265] Based on the adjusted target dialogue generation strategy, record the new dialogue interaction data generated after the second adjustment. The new dialogue interaction data includes the adjusted semantic features, modifier usage information and user feedback information.

[0266] The structured dataset is updated based on the new dialogue interaction data so that the updated structured dataset can be used to adjust the parameters of the lightweight multi-turn dialogue converter model and the large language model.

[0267] The multi-turn dialogue modifier coverage processing device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0268] Figure 11 A schematic diagram of the structure of a processing device for multi-turn dialogue modifier coverage provided in an embodiment of this application. For example... Figure 11 As shown, the multi-turn dialogue modifier coverage processing device 1100 provided in this embodiment includes at least one processor 1101 and a memory 1102. Optionally, the device 1100 also includes a communication component 1103. The processor 1101, memory 1102, and communication component 1103 are connected via a bus.

[0269] In a specific implementation, at least one processor 1101 executes computer execution instructions stored in memory 1102, causing at least one processor 1101 to perform the above-described method.

[0270] The specific implementation process of processor 1101 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0271] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0272] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0273] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0274] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0275] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0276] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0277] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0278] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0279] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0280] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0281] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0282] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0283] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for processing modifier coverage in multi-turn dialogues, characterized in that, include: Based on the preset interactive interface, obtain multi-turn dialogue text data input by the user; The multi-turn dialogue text data is preprocessed to obtain processed standard text data; Based on the preset natural language processing tools, the standard text data is subjected to part-of-speech tagging and dependency analysis to obtain the corresponding part-of-speech tagging results and dependency information; Based on the part-of-speech tagging results, the dependency relationship information, and the standard text data, multiple candidate modifiers are determined; Obtain a pre-built large language model; Based on the large language model, each candidate modifier is disambiguated to obtain multiple target modifiers and the category to which each target modifier belongs; The corresponding comprehensive coverage rate is determined based on each target modifier and its category.

2. The method according to claim 1, characterized in that, The step of determining the corresponding comprehensive coverage rate based on each candidate modifier and its category includes: Based on each target modifier and its category, determine the category coverage, type coverage, and distribution similarity of the target modifier; Based on the category coverage, the type coverage, and the distribution similarity, the corresponding comprehensive coverage is obtained through a weighted fusion algorithm.

3. The method according to claim 2, characterized in that, After determining the corresponding comprehensive coverage rate based on each target modifier and its category, the method further includes: Based on the multi-turn dialogue text data, extract the corresponding semantic features; Obtain an initial dialogue generation strategy, wherein the initial dialogue generation strategy is constructed based on the basic generation logic of a large language model or the preset domain rules of the corresponding application scenario; Based on the comprehensive coverage and the semantic features, the initial dialogue generation strategy is adjusted to obtain the adjusted target dialogue generation strategy.

4. The method according to claim 3, characterized in that, The step of acquiring multi-turn dialogue text data input by the user based on a preset interactive interface includes: Based on the preset interactive interface, collect multi-turn dialogue logs during user interaction; Based on the multi-turn dialogue log, extract multiple corresponding context fragments and determine the fragment length of each context fragment; The length of each context segment is compared with a preset length; For the context fragment exceeding the preset length, it is segmented according to preset semantic units to obtain multiple segmented context fragments; Each segmented context fragment and each context fragment that does not exceed the preset length are organized to obtain the corresponding multi-turn dialogue text data.

5. The method according to claim 4, characterized in that, The preprocessing of the multi-turn dialogue text data to obtain processed standard text data includes: Based on the multi-turn dialogue logs, determine the corresponding session structure; Based on the conversation structure, the multi-turn dialogue log is segmented to obtain the corresponding dialogue turn information, user input information, system response information, and timestamp information; Based on the dialogue turn information, the user input information, the system response information, and the timestamp information, redundant symbols and noise information are removed from the multi-turn dialogue text data to obtain the corresponding first text data; wherein, the redundant symbols include spaces, line breaks, and paragraph marks, and the noise information includes headers and footers, and system prompts; Based on the first text data, the corresponding sensitive information is determined, and the sensitive information is de-identified to obtain the corresponding standard text data.

6. The method according to claim 5, characterized in that, The step of determining the category coverage, type coverage, and distribution similarity of the target modifiers based on each target modifier and its category includes: Obtain a pre-constructed set of modifier categories, wherein the set of modifier categories is constructed based on linguistic rules or domain knowledge; The number of categories of different categories of modifiers in the target modifier is determined, and the ratio of the number of categories to the total number of categories in the modifier category set is calculated to obtain the category coverage rate; Construct a reference set of corresponding modifier categories based on a pre-set large-scale dialogue dataset or domain standard dictionary; Construct a current dialogue modifier category set based on each of the target modifiers; Calculate the intersection of the current dialogue modifier category set and the modifier category reference set, and calculate the ratio of the number of types in the intersection to the total number of types in the modifier category reference set to obtain the type coverage rate; Based on the current dialogue modifier category set, determine the frequency of occurrence of each type in the current dialogue modifier category set, so as to determine the first probability distribution of the current dialogue modifier type; Obtain a pre-built reference corpus, and based on the reference corpus, obtain a second probability distribution of modifier types in the reference corpus that are set based on user preferences or domain standards; Based on the first probability distribution and the second probability distribution, the corresponding symmetric divergence is determined, and the difference between the first probability distribution and the symmetric divergence is taken as the distribution similarity.

7. The method according to claim 6, characterized in that, The step of extracting corresponding semantic features based on the multi-turn dialogue text data includes: Obtain a preset lightweight multi-turn dialogue converter model; wherein, the lightweight multi-turn dialogue converter model is based on a converter encoder and integrates a linear attention mechanism, a context gating module, and modifier awareness construction; The standard text data is concatenated into a sequence format containing round embeddings according to the dialogue rounds; The sequence format is input into the lightweight multi-turn dialogue converter model for context modeling to generate the hidden state of each text tag; Based on the hidden state, the corresponding semantic features are extracted; wherein, the semantic features include user intent vector, topic vector, sentiment polarity and user preference vector, the user intent vector is taken from the hidden state of the sequence start marker, the topic vector is obtained by weighted average pooling of the hidden state by the context gating module, and the user preference vector is generated by training a preset multilayer perceptron based on the user's historical dialogue data.

8. The method according to any one of claims 3-7, characterized in that, After obtaining the corresponding comprehensive coverage rate through a weighted fusion algorithm based on the category coverage rate, the type coverage rate, and the distribution similarity, the method further includes: Based on the comprehensive coverage rate and the semantic features, a corresponding joint indicator system is constructed; The comprehensive coverage, semantic features, and joint indicator system are stored in a structured manner to obtain the corresponding structured dataset; The target dialogue generation strategy is adjusted according to the joint indicator system to obtain the adjusted target dialogue generation strategy. Based on the structured dataset and the adjusted target dialogue generation strategy, a corresponding modifier coverage analysis report and semantic optimization suggestions are generated.

9. The method according to claim 8, characterized in that, The step of constructing a corresponding joint indicator system based on the comprehensive coverage rate and the semantic features includes: Based on the standard text data, determine the corresponding multi-turn dialogue context fragments; The multi-turn dialogue context fragments are input into the lightweight multi-turn dialogue converter model to obtain the corresponding semantic feature-related hidden states; The semantic feature-related hidden states are aggregated to obtain the average vector of the dialogue context. Obtain the system response text for the current round and input the system response text for the current round into the lightweight multi-turn dialogue converter model to obtain the corresponding current response vector; Calculate the first cosine similarity between the average vector of the dialogue context and the current response vector, and use the first cosine similarity as the semantic consistency score; Based on the user preference vector in the semantic features, the system response text of the current round is encoded by the lightweight multi-turn dialogue converter model to obtain the semantic embedding vector of the system response; Calculate the second cosine similarity between the user preference vector and the semantic embedding vector of the system response, and use the second cosine similarity as the user preference matching degree; The comprehensive coverage rate, the semantic consistency score, and the user preference matching degree are weighted and integrated to construct the joint indicator system.

10. The method according to claim 9, characterized in that, The step of structurally storing the comprehensive coverage, the semantic features, and the joint indicator system to obtain the corresponding structured dataset includes: The comprehensive coverage rate, the semantic features, and the joint indicator system are vectorized to obtain the corresponding vectorized results; The vectorized results are associated with the dialogue turn information and the timestamp information, and stored in a pre-built semantic database and vector retrieval library to obtain the structured dataset.

11. The method according to claim 10, characterized in that, The step of adjusting the target dialogue generation strategy according to the joint indicator system to obtain the adjusted target dialogue generation strategy includes: Obtain the indicator values ​​of each item in the joint indicator system, and compare the indicator values ​​with preset threshold values; If the indicator value is lower than the preset threshold, the target dialogue generation strategy is adjusted a second time by means of prompt word enhancement or decoding reweighting to obtain the adjusted target dialogue generation strategy.

12. The method according to claim 11, characterized in that, After obtaining the adjusted target dialogue generation strategy, the method further includes: Based on the adjusted target dialogue generation strategy, record the new dialogue interaction data generated after the second adjustment. The new dialogue interaction data includes the adjusted semantic features, modifier usage information, and user feedback information. The structured dataset is updated based on the new dialogue interaction data so that the updated structured dataset can be used to adjust the parameters of the lightweight multi-turn dialogue converter model and the large language model.

13. A device for solving operations research problems, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-12.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-12.

Images