Dialogue interaction method and system based on long historical dialogue semantic understanding

Abstract

The invention provides a dialogue interaction method and system based on long historical dialogue semantic understanding, and belongs to the technical field of natural language process.The dialogue interaction method comprises the steps that firstly, a historical dialogue information set containing multiple rounds of user consultation and customer service response of an e-commerce platform target user is obtained, and then deep semantic mining is conducted on the historical dialogue information set; dialogue intention features and dialogue theme association features are obtained, a historical dialogue semantic understanding result containing user core demand induction and a dialogue theme development knowledge path is generated according to the two features, an adaptive dialogue interaction strategy is generated according to the historical dialogue semantic understanding result, and the adaptive dialogue interaction strategy comprises response emphasis, guide direction and rhythm control information. And finally, executing the strategy to realize dialogue interaction, outputting a response statement and receiving new input of the user, thereby improving customer service quality and user experience of the e-commerce platform.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to a dialogue interaction method and system based on long historical dialogue semantic understanding. BACKGROUND

[0002] In the customer service scenario of an e-commerce platform, dialogue interaction is an important bridge connecting users and the platform. With the continuous expansion of e-commerce business and the increasing diversification of user needs, users often involve multiple rounds of complex communication when having a dialogue with customer service, covering product information inquiry, order processing request, after-sales problem feedback, and other aspects.

[0003] However, the existing dialogue interaction methods have obvious deficiencies in processing user historical dialogue. On the one hand, most methods only stay at the simple understanding of dialogue surface information, and cannot deeply mine the user's potential demand direction, making it difficult to accurately grasp the user's real intention in different rounds of dialogue. On the other hand, there is a lack of effective analysis and processing of the theme continuation and conversion relationship between different rounds of dialogue, and the development context of the dialogue theme cannot be clearly sorted out. This leads to the fact that customer service can only give simple replies to the current problem when responding to users, and cannot provide comprehensive, coherent and user-potential-demand-compliant services, resulting in poor user experience, and also affecting the service quality and operational efficiency of the e-commerce platform. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, a dialogue interaction method based on long historical dialogue semantic understanding is provided, which comprises: obtaining a set of historical dialogue information of a target user in an e-commerce platform, the set of historical dialogue information comprising multiple rounds of user inquiry sentences and corresponding customer service response sentences, the user inquiry sentences involving product information inquiry, order processing request, after-sales problem feedback, and the customer service response sentences comprising answer content and guidance information for the user inquiry sentences; performing deep semantic mining processing on the set of historical dialogue information to obtain dialogue intention features and dialogue theme association features of the target user, the dialogue intention features embodying the potential demand direction of the target user in each round of dialogue, and the dialogue theme association features reflecting the theme continuation and conversion relationship between different rounds of dialogue; generating a historical dialogue semantic understanding result according to the dialogue intention features and the dialogue theme association features, the historical dialogue semantic understanding result comprising user core demand induction information and dialogue theme development knowledge path; generating an adapted dialogue interaction strategy according to the historical dialogue semantic understanding result, the dialogue interaction strategy comprising response content emphasis information, guidance direction and rhythm control mode information; The dialogue interaction strategy is executed to realize the dialogue interaction operation with the target user, output a response sentence conforming to the dialogue interaction strategy, and receive a new input sentence of the target user.

[0005] In still another aspect, the application further provides a dialogue interaction system based on long historical dialogue semantic understanding, comprising a processor and a machine readable storage medium, the machine readable storage medium is connected with the processor, the machine readable storage medium is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the machine readable storage medium to realize the above method.

[0006] Based on the above aspects, the application can obtain a historical dialogue information set of a target user in an e-commerce platform, which contains multiple rounds of user consultation and customer service response, perform deep semantic mining processing on the historical dialogue information set, obtain dialogue intention features and dialogue theme association features of the target user, and in-depth analyze the potential demand direction of the user in each round of dialogue and the theme continuation and conversion relationship between different rounds of dialogue. The historical dialogue semantic understanding result generated according to the dialogue intention features and the dialogue theme association features contains user core demand induction information and dialogue theme development knowledge path. The adaptive dialogue interaction strategy generated according to the historical dialogue semantic understanding result clearly defines the response content focus, guidance direction and rhythm control mode, so that the customer service can provide more comprehensive, coherent and user potential demand conforming service. Finally, the dialogue interaction strategy is executed to realize the dialogue interaction operation with the target user, output a response sentence conforming to the strategy, and receive a new input sentence of the user, which effectively improves the customer service quality and user experience of the e-commerce platform. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 is an execution flow schematic diagram of the dialogue interaction method based on long historical dialogue semantic understanding provided by an embodiment of the application.

[0008] Figure 2 is a schematic diagram of exemplary hardware and software components of the dialogue interaction system based on long historical dialogue semantic understanding provided by an embodiment of the application. DETAILED DESCRIPTION

[0009] The application will be specifically described below in combination with the drawings of the specification, Figure 1 is a flow schematic diagram of the dialogue interaction method based on long historical dialogue semantic understanding provided by an embodiment of the application. The dialogue interaction method based on long historical dialogue semantic understanding will be described in detail below.

[0010] Step S110, obtain the historical dialogue information set of the target user in the e-commerce platform, the historical dialogue information set contains multiple rounds of user inquiry sentences and corresponding customer service response sentences, the user inquiry sentences involve product information inquiry, order processing request, after-sales problem feedback, and the customer service response sentences contain answer content and guidance information for the user inquiry sentences.

[0011] In this embodiment, the historical dialogue information set of a target user B on a certain comprehensive e-commerce platform is selected as the processing object. The historical dialogue information set covers 12 rounds of dialogue between user B and platform customer service within a month, each round of dialogue contains the inquiry sentence sent by user B and the response sentence made by the customer service for the sentence.

[0012] Among these user inquiry sentences, there are product information inquiries such as "As mentioned earlier, the product with multiple functions, what are the main differences in core performance between different versions? In addition, are there differences in the application scenarios corresponding to these versions?"; order processing requests such as "The order I submitted last week contains multiple categories, and it shows that some goods have been shipped, how is the restocking of the remaining goods? Can you speed up the processing progress to ensure that the whole is shipped as soon as possible?"; and after-sales problem feedback such as "One of the accessories in the goods I received is damaged, and the running state is different from the description when in use, how should this situation be handled for return or repair?"

[0013] The answer content of the customer service response sentence for the product information inquiry is "The main differences in core performance between different versions of the product are mainly reflected in running speed, endurance, and compatibility range. Among them, the basic version is suitable for daily simple use scenarios, the advanced version is more suitable for moderate use requirements, and the professional version can meet the high-intensity complex operation scenarios." and contains guidance information "If you need more detailed comparison parameters, please inform us of your main use requirements, and we will provide you with targeted recommendations." The response to the order processing request is "After inquiry, the remaining goods in your order are currently in a state of waiting for shipment, and we have fed back your needs to the warehouse, which will be processed in priority. It is expected to be completed within the next two days, and the logistics information will be updated after shipment." For the after-sales problem feedback, the customer service response is "Please first upload the photos or videos of the damaged parts of the goods and the running abnormal state to the platform after-sales channel, and our after-sales specialist will review it within 24 hours. After the review is passed, we will arrange return or repair service for you according to the specific circumstances, and the related expenses during the period will be borne by the platform." Step S120, deep semantic mining processing is performed on the set of historical dialogue information to obtain a dialogue intention feature and a dialogue theme association feature of the target user, the dialogue intention feature representing a potential demand direction of the target user in each round of dialogue, and the dialogue theme association feature reflecting a continuation and conversion relationship of themes between different rounds of dialogue.

[0014] In this embodiment, deep semantic mining processing is performed on the set of 12 rounds of historical dialogue information of user B. First, the user inquiry sentence and the customer service response sentence in each round of dialogue need to be processed in detail, key information is extracted therefrom, and the dialogue intention feature of the user is analyzed, which clearly presents the potential demand direction of the user in each round of dialogue, such as wanting to understand the characteristics of the product, being anxious to know the order progress, or seeking after-sales solutions, etc. At the same time, the theme association between different rounds of dialogue is analyzed to determine whether the theme is continued or converted, thereby obtaining the dialogue theme association feature.

[0015] Step S121, the user inquiry sentence and the customer service response sentence in the set of historical dialogue information are processed according to the natural pause and semantic integrity of the sentence, and a plurality of dialogue sentence units are divided, each dialogue sentence unit retaining the original context position information.

[0016] In this embodiment, taking a round of dialogue of user B as an example, the user inquiry sentence is "What are the main differences in core performance between different versions of the product with multiple functions mentioned earlier? In addition, are there differences in the corresponding application scenarios of these versions?". According to the natural pause and semantic integrity, it can be divided into two dialogue sentence units, the first unit is "What are the main differences in core performance between different versions of the product with multiple functions mentioned earlier?", and the second unit is "In addition, are there differences in the corresponding application scenarios of these versions?".

[0017] For the customer service response sentence "The main differences between different versions of the product in core performance are reflected in running speed, endurance, and compatibility range. Among them, the basic version is suitable for daily simple use scenarios, the advanced version is more suitable for moderate use needs, and the professional version can meet the needs of high-intensity complex operation scenarios. If you need more detailed comparison parameters, please tell us your main use needs, and we will provide you with targeted recommendations.", it is divided into three dialogue sentence units, which are "The main differences between different versions of the product in core performance are reflected in running speed, endurance, and compatibility range.", "Among them, the basic version is suitable for daily simple use scenarios, the advanced version is more suitable for moderate use needs, and the professional version can meet the needs of high-intensity complex operation scenarios.", and "If you need more detailed comparison parameters, please tell us your main use needs, and we will provide you with targeted recommendations."

[0018] Each divided dialogue sentence unit retains its position information in the original dialogue, such as belonging to the first dialogue, being the first unit of the user's inquiry sentence in the dialogue, or being the first unit of the customer service response sentence, etc., so as to analyze the context semantic association in the subsequent.

[0019] Step S122, extracting feature words related to e-commerce transactions from the plurality of dialogue sentence units, and determining a core feature word set through word frequency statistics and semantic importance evaluation, the feature words including product names, transaction process terms, and problem description words.

[0020] In this embodiment, feature words are extracted from all divided dialogue sentence units. For sentence units related to product information, product-related feature words such as "product", "different versions", "core performance", "running speed", "endurance", "compatibility range", "application scenario", "basic version", "advanced version", and "professional version" are extracted; from order processing-related sentence units, transaction process terms such as "order", "product", "inventory status", "processing progress", "shipping", and "logistics information" are extracted; from after-sales problem-related sentence units, problem description words such as "product", "accessory", "damage", "running status", "deviation", "return", and "repair" are extracted.

[0021] Next, word frequency statistics are performed to count the number of times each feature word appears in all dialogue sentence units. For example, the word "product" appears in multiple rounds of dialogue with a relatively high frequency; "core performance" mainly appears in product information inquiry dialogue with a medium frequency.

[0022] Then, semantic importance evaluation is performed to evaluate the importance of each feature word in expressing the semantics of the dialogue and the user's demand. Words such as "core performance", "return", and "logistics information" are directly related to the core needs of the user and are of high importance; and conjunctions such as "in addition" and "among" are of low importance.

[0023] Based on the results of the vocabulary frequency statistics and the semantic importance evaluation, feature words with high importance and a frequency that meets certain requirements are selected to form a core feature word set, such as "product", "different versions", "core performance", "running speed", "battery life", "application scenarios", "order", "logistics information", "damage", and "return".

[0024] In step S1221, part-of-speech tagging is performed on each dialogue sentence unit to identify nouns, verbs, and adjectives, and nouns and verbs that can be used as feature words are selected from among them.

[0025] In this embodiment, a natural language processing tool is used to perform part-of-speech tagging on each dialogue sentence unit. Taking the dialogue sentence unit "What are the main differences in core performance between different versions of the multi-functional product mentioned earlier?" as an example, after tagging, "product", "version", and "core performance" are nouns, "mention" and "have" are verbs, and "multiple", "different", and "main" are adjectives.

[0026] From the tagging results, nouns and verbs are selected as potential feature words. In the above sentence unit, the nouns "product", "version", and "core performance" and the verbs "mention" and "have" are selected. In combination with the e-commerce transaction scenario, "mention" and "have" have relatively weak roles in expressing core semantics, while "product", "version", and "core performance" are closely related to product information, and these nouns are ultimately determined as the feature words of the sentence unit.

[0027] In step S1222, the selected words are compared with a pre-set e-commerce domain vocabulary library to match candidate feature words that belong to product names, transaction process terms, and question description words.

[0028] In this embodiment, the pre-set e-commerce domain vocabulary library contains a large number of e-commerce transaction-related words, which are classified, such as product name-related words, transaction process terms related to orders, payments, and logistics, and question description words related to damage, failure, and return.

[0029] The screened words from the dialogue sentence units are matched with the word library, for example, the words such as “product”, “version”, “core performance” are matched with the product name type words in the word library to become product name type candidate feature words; the words such as “order”, “logistics information”, “delivery” are matched with the transaction process terminology type words to become transaction process terminology type candidate feature words; the words such as “damage”, “return”, “repair” are matched with the problem description words to become problem description word type candidate feature words.

[0030] In step S1223, the total number of occurrences of each candidate feature word in the historical dialogue information set is counted, and the distribution frequency of each candidate feature word in each round of dialogue is calculated to obtain the frequency distribution feature.

[0031] In this embodiment, the total number of occurrences of each candidate feature word is counted by traversing the entire historical dialogue information set. For example, “product” occurs 15 times, “order” occurs 10 times, and “return” occurs 8 times.

[0032] Then, the distribution frequency of each candidate feature word in each round of dialogue is calculated, that is, the ratio of the number of occurrences of the word in each round of dialogue to the total number of words in the round of dialogue. Taking “order” as an example, it occurs 2 times in the 3rd round of dialogue, and the total number of words in the round of dialogue is 50, so the distribution frequency in the 3rd round of dialogue is 2 / 50; it occurs 3 times in the 7th round of dialogue, and the total number of words in the round of dialogue is 60, so the distribution frequency is 3 / 60, and so on. The distribution frequency of “order” in each round of dialogue is obtained, and then the frequency distribution feature is formed. Other candidate feature words are calculated in the same way to obtain their respective frequency distribution features.

[0033] In step S1224, the semantic importance of the candidate feature words is evaluated according to the semantic weight of each candidate feature word in the sentence, the close degree of association with the dialogue theme, and the key degree in the e-commerce transaction scene, and each candidate feature word is given a corresponding importance score.

[0034] In this embodiment, the semantic importance of the candidate feature words is evaluated from three dimensions. In terms of semantic weight, the position of the feature word in the sentence (such as the beginning, middle, or end of the sentence) and whether it is the emphasized content of the sentence are determined. The feature word with high semantic weight is at the beginning of the sentence and is the emphasized content. In terms of close degree of association with the dialogue theme, if the feature word directly revolves around the current dialogue theme, the close degree of association is high, for example, in the dialogue discussing the order problem, “order” and “logistics” are closely associated with the theme. In terms of the key degree in the e-commerce transaction scene, the feature word directly affecting the transaction process or the core needs of the user has high key degree, such as “payment”, “return”, etc.

[0035] Based on the evaluation results of these three dimensions, each candidate feature word is assigned an importance score between 0 and 1. For example, in a product information inquiry dialogue, "core performance" has a high semantic weight, is closely related to the topic, and is crucial for users to understand the product, with a score of 0.85. "In addition," as a connective word, has a low semantic weight, is loosely related to the topic, and is less critical, with a score of 0.1.

[0036] Step S1225 , a screening threshold is set based on the frequency distribution characteristics and the importance score, and candidate feature words with importance scores higher than the screening threshold are selected to form a core feature word set, wherein each word in the core feature word set retains its original position mark in the dialogue sentence unit.

[0037] In this embodiment, the frequency distribution characteristics and importance scores of the candidate feature words are comprehensively considered to set a screening threshold. Assume that after analysis, the importance score threshold is set to 0.5. All candidate feature words are traversed and words with an importance score higher than 0.5 are selected, such as "product", "different versions", "core performance", "operating speed", "endurance", "applicable scenarios", "order", "logistics information", "damage", "return", etc., to form the core feature word set.

[0038] At the same time, in the core feature word set, each word retains its position mark in the original dialogue sentence unit. For example, "order" appears in the first unit of the user consultation sentence in the third round of dialogue, and its position mark is "3-U-1", where "3" represents the third round, "U" represents the user consultation sentence, and "1" represents the first unit.

[0039] Step S1226 , performing deduplication processing on the core feature word set, and performing relevance sorting on the deduplicated core feature word set according to the degree of semantic relevance between the core feature words, and determining a relevance sorting result.

[0040] In this embodiment, the core feature word set may contain repeated words. For example, if "commodity" appears in multiple sentence units, only one "commodity" word is retained after deduplication.

[0041] The semantic correlation between core feature words is then analyzed, measured by methods such as calculating word vector similarity. For example, "core performance" has a high degree of semantic correlation with "running speed" and "battery life," while "order" is closely associated with "logistics information" and "shipping." Based on these correlations, the deduplicated core feature word set is sorted, with highly correlated words grouped together to form a correlation ranking result, such as "product," "different versions," "core performance," "running speed," "battery life," "applicable scenarios," "order," "shipping," "logistics information," "damage," "return," and "repair."

[0042] Step S1227, optimizing the internal structure of the core feature word set through the relevance ranking result to form a final core feature word set.

[0043] In this embodiment, the internal structure of the core feature word set is optimized according to the relevance ranking result. Feature words with high semantic correlation are grouped into a group to form multiple feature word groups. For example, “product”, “different versions”, “core performance”, “running speed”, “endurance capability”, and “application scenarios” are grouped into a feature word group of commodity information; “order”, “delivery”, and “logistics information” are grouped into a feature word group of order processing; and “damage”, “return”, and “repair” are grouped into a feature word group of after-sales problems.

[0044] The above optimization makes the internal structure of the core feature word set more clear, facilitating subsequent semantic analysis and feature extraction, and the final core feature word set is presented in the form of these feature word groups.

[0045] Step S123, performing semantic analysis on the core feature word set in combination with the context to determine the basic intent category corresponding to each user inquiry sentence.

[0046] In this embodiment, for each user inquiry sentence, the semantic analysis is performed on the relevant feature words in the core feature word set in combination with the context, i.e., the customer service response sentences before and after the sentence and other related user inquiry sentences. For example, the user inquiry sentence “I submitted an order containing multiple categories last week, and it shows that some goods have been shipped. How is the restocking situation of the remaining goods? Can the processing progress be accelerated to ensure that the whole is shipped as soon as possible?” The core feature words include “order”, “goods”, “restocking situation”, “processing progress”, and “shipment”. In combination with the context, the user did not mention the order-related problem in the previous dialogue, and this time is the first inquiry, and the subsequent response of the customer service is also about order processing, so it can be analyzed that the intent of the user inquiry sentence is related to order processing.

[0047] Through the above analysis of each user inquiry sentence, the corresponding basic intent category of each is determined, such as commodity information inquiry, order processing request, and after-sales problem feedback.

[0048] Step S1231, constructing an association mapping table of basic intent categories and core feature words, and recording in the association mapping table the mapping relationship that the commodity name class feature word corresponds to inquiry commodity information, the transaction process terminology class feature word corresponds to inquiry transaction status, and the problem description vocabulary class feature word corresponds to feedback after-sales problems.

[0049] In this embodiment, the constructed association mapping table explicitly shows the correspondence between different types of core feature words and basic intent categories. Among them, the commodity name class feature words such as "product", "different version", "core performance", "application scenario" correspond to "consulting commodity information" in the basic intent category; the transaction process term class feature words such as "order", "delivery", "logistics information", "inventory situation" correspond to "inquiring transaction status"; the problem description vocabulary class feature words such as "damage", "operation state deviation", "return", "repair" correspond to "raising post-sale problems".

[0050] The association mapping table is stored in the form of a table, which facilitates subsequent quick query of the corresponding basic intent category of the feature word when analyzing the user consultation statement.

[0051] Step S1232, for each user consultation statement, extracting the core feature words contained in the user consultation statement from the core feature word set, and preliminarily determining the candidate basic intent category to which the user consultation statement may correspond according to the mapping relationship recorded in the association mapping table.

[0052] In this embodiment, taking the user consultation statement "Among the goods I received, one of the accessory components is damaged, and when using it, it is found that its operation state deviates from the description. How should this be handled for return or repair?" as an example, the core feature words contained in the statement are extracted from the core feature word set, including "commodity", "accessory component", "damage", "operation state", "deviation", "return", and "repair".

[0053] According to the association mapping table, these feature words belong to the problem description vocabulary class, so the candidate basic intent category to which the user consultation statement corresponds is preliminarily determined as "raising post-sale problems".

[0054] For user consultation statements containing multiple types of core feature words, such as containing both commodity name class and transaction process term class feature words, multiple candidate basic intent categories will be preliminarily determined.

[0055] Step S1233, based on the mapping logic of the association mapping table, analyzing the context of the user consultation statement, checking the core feature words and semantic tendencies involved in the adjacent customer service response statements and user consultation statements, and verifying whether the preliminarily determined candidate basic intent category conforms to the mapping relationship.

[0056] In this embodiment, taking a user consultation statement as an example, the user consultation statement contains core feature words such as "product", "core performance", and "order", and the preliminarily determined candidate basic intent categories are "consulting commodity information" and "inquiring transaction status".

[0057] The context of the user inquiry statement is analyzed. The user inquiry statement before the user inquiry statement mainly revolves around the function of the commodity, and the customer service response statement also answers the function of the commodity. The customer service response statement after that mentions the processing of the order. By viewing the core feature words in the previous and subsequent statements, such as "function" and "parameter" before and "order number" and "shipping time" after, and the semantic tendency, it is found that the statement is more focused on further inquiry of the commodity information, and therefore it is verified that the candidate basic intent category of "inquiry of commodity information" is more consistent with the mapping relationship.

[0058] In step S1234, if there are multiple candidate basic intent categories, the matching degree of each candidate and the context is calculated according to the association strength of the feature words and the intent categories in the association mapping table, and the candidate basic intent category with the highest matching degree is selected as the basic intent category corresponding to the user inquiry statement.

[0059] In this embodiment, the association mapping table also contains the association strength information of the feature words and the intent categories, for example, the association strength of "core performance" and "inquiry of commodity information" is 0.9, and the association strength of "order" and "inquiry of transaction status" is 0.85. For the user inquiry statement with multiple candidate basic intent categories, the matching degree of each candidate and the context is calculated.

[0060] Taking the user inquiry statement containing the core feature words "product", "core performance", and "order" as an example, "inquiry of commodity information" and "inquiry of transaction status" are candidate basic intent categories. When calculating the matching degree, each feature word in the statement is multiplied by the association strength of the corresponding intent category and the semantic weight of the feature word in the context (the semantic weight is determined according to the frequency and importance of the feature word in the context), and then the products are summed to obtain the matching degree of the candidate basic intent category.

[0061] Assuming that the semantic weights of "product" and "core performance" in the context are 0.7 and 0.8 respectively, and the semantic weight of "order" is 0.5. The matching degree of "inquiry of commodity information" is ("product" association strength x 0.7) + ("core performance" association strength x 0.8), and the matching degree of "inquiry of transaction status" is ("order" association strength x 0.5). After calculation, the matching degree of "inquiry of commodity information" is higher than that of "inquiry of transaction status", so "inquiry of commodity information" is selected as the basic intent category corresponding to the user inquiry statement.

[0062] In step S1235, the basic intent categories corresponding to each user inquiry statement are arranged in the order of the dialogue round, forming a sequence of basic intent categories.

[0063] In this embodiment, in the 12 rounds of conversation of user B, the user inquiry statement of each round determines the corresponding basic intent category, which is in turn: inquiry of commodity information, inquiry of commodity information, inquiry of transaction status, inquiry of commodity information, offer of post-sale question, inquiry of transaction status, inquiry of commodity information, offer of post-sale question, inquiry of transaction status, inquiry of commodity information, offer of post-sale question, inquiry of transaction status. The basic intent category sequence is formed by arranging these basic intent categories in order.

[0064] In step S1236, the proportion of each basic intent category in the basic intent category sequence is calculated, and the distribution characteristics of different intent categories are analyzed, which are positively correlated with the preset association strength of the association mapping table.

[0065] In this embodiment, the number of occurrences of each basic intent category in the basic intent category sequence is counted, and then divided by the total number of rounds 12 to obtain the proportion of each basic intent category. For example, "inquiry of commodity information" appears 5 times, accounting for 5 / 12; "inquiry of transaction status" appears 4 times, accounting for 4 / 12; "offer of post-sale question" appears 3 times, accounting for 3 / 12.

[0066] When analyzing the distribution characteristics, it is found that the basic intent category with a high proportion has a high preset association strength of the corresponding feature word in the association mapping table, such as the high association strength of the feature words "core performance" and "product" corresponding to "inquiry of commodity information", which meets the positive correlation between the distribution characteristics and the preset association strength.

[0067] In step S1237, the basic intent category sequence is adjusted based on the distribution characteristics to form the final basic intent category sequence.

[0068] In this embodiment, according to the distribution characteristics, if the proportion of a certain basic intent category is too low and the semantic weight of the corresponding feature word in the context is also low, it may be caused by analysis error, and the basic intent category sequence needs to be adjusted.

[0069] For example, the user inquiry statement of a certain round of conversation is initially determined as "offer of post-sale question", but the proportion of this category in the sequence is extremely low, and the semantic weight of the feature word "damage" in the context is low. Combined with the theme of the previous and subsequent conversations, it is found that the statement is more inclined to "inquiry of commodity information", so the basic intent category of this round is adjusted to "inquiry of commodity information" to form the final basic intent category sequence.

[0070] In step S124, the theme continuity between the statement units of different rounds of conversation is analyzed based on the basic intent category, and the theme similarity parameter is calculated to generate a theme association degree parameter, which is used to represent the association degree of adjacent rounds of conversation in theme.

[0071] In this embodiment, the continuity of the theme between the adjacent round dialogue sentence units is analyzed based on the basic intent category. If the basic intent categories of the adjacent two rounds of dialogue are the same, it means that the theme may have continuity; if they are different, the theme may have changed. The above-mentioned correlation degree is quantified by calculating the theme similarity parameter, and then the theme correlation parameter is generated.

[0072] In step S1241, a theme feature vector is constructed for each basic intent category, which is composed of the core feature words corresponding to the basic intent category and the second weight thereof, and the second weight is determined according to the importance score of the feature word.

[0073] In this embodiment, for the basic intent category of “consulting commodity information”, the corresponding core feature words are “product”, “different versions”, “core performance”, “running speed”, “endurance ability”, “application scenario”, etc., and the importance scores of these feature words are 0.8, 0.75, 0.85, 0.7, 0.72, and 0.68, respectively.

[0074] These core feature words are arranged in a certain order, and the importance score corresponding to each feature word is taken as its weight to construct a theme feature vector. For example, the theme feature vector of “consulting commodity information” is [ (“product”, 0.8), (“different versions”, 0.75), (“core performance”, 0.85), (“running speed”, 0.7), (“endurance ability”, 0.72), (“application scenario”, 0.68) ].

[0075] Similarly, the corresponding theme feature vectors are also constructed for “inquiring transaction status” and “asking post-sale questions”, wherein the theme feature vector of “inquiring transaction status” contains the core feature words “order”, “delivery”, “logistics information” and their weights, and the theme feature vector of “asking post-sale questions” contains the core feature words “damage”, “return”, “repair” and their weights.

[0076] In step S1242, the dialogue sentence units of adjacent rounds are selected, and the theme feature vectors of the basic intent categories corresponding to the dialogue sentence units of adjacent rounds are extracted.

[0077] In this embodiment, the sentence units of the second and third rounds of dialogue are selected, the theme feature vector V2 of the second round of user consultation sentence corresponding to the basic intent category of “consulting commodity information” is extracted, and the theme feature vector V3 of the third round of user consultation sentence corresponding to the basic intent category of “inquiring transaction status” is extracted.

[0078] For example, the sentence units of the 5th round and the 6th round of dialogue are selected, the basic intent category corresponding to the 5th round is "sell after question", and the theme feature vector is V5; the basic intent category corresponding to the 6th round is "ask transaction status", and the theme feature vector is V6.

[0079] In step S1243, the cosine similarity between the two theme feature vectors is calculated to obtain a theme similarity parameter, and the larger the theme similarity parameter is, the closer the theme relevance of the two round dialogues is.

[0080] In this embodiment, when calculating the cosine similarity between the theme feature vectors, first, the core feature words in each theme feature vector are converted into word vectors, and the dimension of the word vector is determined according to a preset word embedding model, such as 100 dimensions. Then, the word vectors of the core feature words in each theme feature vector are weighted and averaged according to the weight to obtain a comprehensive vector of the theme feature vector.

[0081] For example, the cosine similarity of V2 and V3 is calculated, first, the comprehensive vector A of V2 and the comprehensive vector B of V3 are obtained, the calculation method of the cosine similarity is the dot product of A and B divided by the product of the module length of A and the module length of B, and the result obtained is the theme similarity parameter. If the parameter is 0.3, it indicates that the theme relevance of the 2nd round and the 3rd round of dialogue is general; if the cosine similarity of the theme feature vectors of the 1st round and the 2nd round of dialogue is calculated to be 0.8, it indicates that the theme relevance of the two rounds of dialogue is relatively close.

[0082] In step S1244, the association level is divided according to the size of the theme similarity parameter, and different association levels correspond to different theme relevance parameter values.

[0083] In this embodiment, the range of the theme similarity parameter is divided into multiple intervals, and each interval corresponds to an association level and a theme relevance parameter value. For example, when the theme similarity parameter is between 0.7-1.0, the association level is "close association", and the theme relevance parameter value is 0.9; when it is between 0.4-0.7, the association level is "general association", and the theme relevance parameter value is 0.5; when it is between 0-0.4, the association level is "weak association", and the theme relevance parameter value is 0.2.

[0084] For example, the theme similarity parameter of the 1st round and the 2nd round of dialogue is 0.8, which corresponds to "close association", and the theme relevance parameter value is 0.9; the theme similarity parameter of the 2nd round and the 3rd round of dialogue is 0.3, which corresponds to "weak association", and the theme relevance parameter value is 0.2.

[0085] In step S1245, the theme relevance parameter of the adjacent round of dialogue is recorded in the order of the round, to form a theme relevance sequence, and the total number of theme transitions in the entire historical dialogue information set is calculated to obtain the theme transition frequency in combination with the total number of rounds of dialogue.

[0086] In this embodiment, the theme correlation degree parameters of adjacent turns in 12 turns of dialogue are recorded in order, such as 0.9 for the first and second turns, 0.2 for the second and third turns, 0.3 for the third and fourth turns, 0.4 for the fourth and fifth turns, 0.2 for the fifth and sixth turns, 0.5 for the sixth and seventh turns, 0.3 for the seventh and eighth turns, 0.4 for the eighth and ninth turns, 0.2 for the ninth and tenth turns, 0.3 for the tenth and eleventh turns, and 0.5 for the eleventh and twelfth turns, forming a theme correlation degree sequence [0.9, 0.2, 0.3, 0.4, 0.2, 0.5, 0.3, 0.4, 0.2, 0.3, 0.5].

[0087] The total number of theme transitions is calculated. When the basic intent categories of adjacent turns are different, it is considered as a theme transition. The total transition number is 7. The total number of turns is 12, and the theme transition frequency is 7 / 11 (because there are 11 adjacent turn pairs in 12 turns of dialogue).

[0088] In step S1246, the theme correlation degree sequence is analyzed by setting a fixed length window. The average correlation degree parameter in the window is calculated by traversing the theme correlation degree sequence.

[0089] In this embodiment, the window length is set to 3, that is, each window contains 3 adjacent theme correlation degree parameters. The window is slid from the first element of the theme correlation degree sequence. The first window contains [0.9, 0.2, 0.3], and the average correlation degree parameter is (0.9+0.2+0.3) / 3. The second window contains [0.2, 0.3, 0.4], and the average correlation degree parameter is (0.2+0.3+0.4) / 3. Similarly, the average correlation degree parameter of each window is obtained until the window slides through the entire sequence.

[0090] In step S1247, the theme correlation degree sequence is adjusted according to the average correlation degree parameter, and the final theme correlation degree parameter is generated by combining the adjusted theme correlation degree sequence and the theme transition frequency.

[0091] In this embodiment, if a theme correlation degree parameter is significantly different from the average correlation degree parameter in the window, such as exceeding the range of ±0.2 of the average correlation degree parameter, the parameter is adjusted to be close to the average correlation degree parameter. For example, the average correlation degree parameter of a window is 0.4, and one of the theme correlation degree parameters in the window is 0.1, which is significantly different. It is adjusted to 0.3.

[0092] The adjusted theme relevance sequence is smoother and can more accurately reflect the overall trend of theme relevance. In combination with the adjusted theme relevance sequence and the theme transition frequency, the final theme relevance parameter of each adjacent round of dialogue is comprehensively determined. For example, for the adjusted parameter, the parameter is fine-tuned according to the high or low of the theme transition frequency. If the theme transition frequency is high, the part of the relevance parameter is appropriately reduced to reflect the frequency of theme transition.

[0093] In step S125, the dialogue intention feature and the dialogue theme relevance feature of the target user are constructed according to the basic intention category and the theme relevance parameter. The dialogue intention feature includes an intention category sequence and an intention intensity descriptor, and the dialogue theme relevance feature includes a theme transition frequency and a theme duration parameter.

[0094] In this embodiment, the final basic intention category sequence is taken as the intention category sequence in the dialogue intention feature. At the same time, an intention intensity descriptor is added for each basic intention category. The intention intensity descriptor is determined according to the semantic weight and the appearance frequency of the intention category in the dialogue. For example, “consulting commodity information” occupies an important position in multiple dialogues, and its intention intensity descriptor is “high intensity”. “Asking about the sale problem” appears less frequently, and its intention intensity descriptor is “medium intensity”.

[0095] The theme transition frequency in the dialogue theme relevance feature is 7 / 11, which is calculated previously. The theme duration parameter is calculated by counting the number of rounds in which each theme (i.e. basic intention category) appears continuously. For example, “consulting commodity information” appears continuously for 2 rounds, and its duration parameter is 2. “Inquiring about the transaction status” appears continuously for 1 round, and its duration parameter is 1.

[0096] In step S126, the dialogue intention feature is weighted and integrated according to the importance of each round of dialogue in the overall historical dialogue. Different rounds of dialogue are assigned a first weight. The higher the first weight of a round of dialogue, the greater the impact of the round of dialogue on the final dialogue intention feature.

[0097] In this embodiment, when evaluating the importance of each round of dialogue, factors such as the position of the round in the dialogue (e.g. newer rounds may be more important), the detail of the dialogue content, and whether it involves the user's core needs are considered. Higher first weights are assigned to rounds with higher importance, such as the 10th round of dialogue, which involves the user's final confirmation of the product, with a first weight of 0.9. Lower first weights are assigned to earlier rounds, such as the 1st round, which has a first weight of 0.5.

[0098] The intention category sequence elements of each round of dialogue are multiplied by their corresponding first weights to obtain a weighted intention category sequence, thereby achieving weighted integration of the dialogue intention feature and making the final dialogue intention feature more accurately reflect the intentions of important rounds.

[0099] Step S127, the dialogue topic association feature is time calibrated, the weighted integrated dialogue intention feature and the time calibrated dialogue topic association feature are cross-verified, if the target user dialogue tendencies reflected by the two are consistent, the dialogue intention feature and the dialogue topic association feature are confirmed, if the target user dialogue tendencies reflected by the two exist deviation, the feature extraction parameter is readjusted and the deep semantic mining processing is performed again until the dialogue intention feature and the dialogue topic association feature are obtained.

[0100] In the embodiment, the dialogue topic association feature is time calibrated, that is, according to the time sequence of the dialogue round, the time mark of the theme duration time parameter and the theme association degree parameter is adjusted to ensure that it is consistent with the time line of the actual dialogue.

[0101] In the cross-verification, whether the high-frequency intention category in the weighted integrated dialogue intention feature matches the theme with longer duration time and higher association degree in the dialogue topic association feature is viewed. For example, the "consulting commodity information" is a high-frequency and high-intensity intention in the dialogue intention feature, the theme corresponding to the "consulting commodity information" has longer duration time and higher association degree in the dialogue topic association feature, which indicates that the dialogue tendencies reflected by the two are consistent, and the final dialogue intention feature and dialogue topic association feature are confirmed.

[0102] If there is deviation, for example, the dialogue intention feature shows that "inquiring transaction status" is the main intention, but the duration time of the theme is short and the association degree is low in the dialogue topic association feature, the importance score parameter in the feature extraction, the weight of the theme feature vector and the like need to be readjusted, and the deep semantic mining processing is performed again until the tendencies of the two are consistent.

[0103] Step S130, the historical dialogue semantic understanding result is generated according to the dialogue intention feature and the dialogue topic association feature, the historical dialogue semantic understanding result includes user core demand induction information and dialogue theme development knowledge path.

[0104] In the embodiment, the intention category sequence and the intention intensity descriptor in the dialogue intention feature are used to induce the core demand of the user in the whole historical dialogue, the theme conversion frequency, the theme duration time parameter and the theme association degree parameter in the dialogue topic association feature are used to construct the knowledge path of the dialogue theme development, and the two parts of the content are integrated to form the historical dialogue semantic understanding result.

[0105] Step S131, the intention category sequence in the dialogue intention feature is statistically analyzed, the intention category with the highest appearance frequency is identified, and the main intention direction of the target user is determined in combination with the intention intensity descriptor.

[0106] In this embodiment, the sequence of intent categories is counted, "consulting commodity information" appears 5 times, "inquiring about transaction status" appears 4 times, and "raising post-sale questions" appears 3 times, so "consulting commodity information" is the intent category with the highest frequency. The intent intensity descriptor is "high intensity", and the main intent direction of the target user is determined to be in-depth understanding of commodity-related information for making purchase decisions.

[0107] In step S132, the relevant user consultation sentences and customer service response sentences are integrated based on the main intent direction, the key information in the sentences is extracted, and the core demand of the target user in the entire historical dialogue is summarized.

[0108] In this embodiment, all relevant user consultation sentences and customer service response sentences are integrated around the main intent direction of "consulting commodity information". The user consultation sentences involve core performance, application scenarios, and other information of different versions of the product, and the customer service response sentences contain the differences in running speed, endurance, and other aspects of each version.

[0109] The key information in these sentences, such as the differences in core performance of different versions and the division of application scenarios, is extracted, and the core demand of the target user is summarized as: clearly understanding the core performance and application scenarios of different versions of the product to choose the version that best meets their needs.

[0110] In step S133, the theme correlation degree sequence and the theme transition frequency in the dialogue theme correlation features are analyzed to analyze the appearance order, duration, and transition nodes of different themes in the dialogue.

[0111] In this embodiment, the appearance order of different themes is analyzed based on the theme correlation degree sequence, theme transition frequency, and theme duration parameters: consulting commodity information - consulting commodity information - inquiring about transaction status - consulting commodity information - raising post-sale questions - inquiring about transaction status - consulting commodity information - raising post-sale questions - inquiring about transaction status - consulting commodity information - raising post-sale questions - inquiring about transaction status.

[0112] The duration of each theme is determined according to the number of consecutive rounds, "consulting commodity information" appears continuously for 2 rounds, the duration is 2; "inquiring about transaction status" appears continuously for 1 round, the duration is 1; "raising post-sale questions" appears continuously for 1 round, the duration is 1.

[0113] The theme transition nodes are the positions where the adjacent rounds of themes are different, such as the 2nd-3rd round, the 3rd-4th round, the 4th-5th round, and other rounds.

[0114] In step S1331, the theme correlation degree sequence is traversed, and the consecutive significantly correlated dialogue segments are determined according to the theme correlation degree parameter value, and each significantly correlated dialogue segment corresponds to a continuous theme.

[0115] In this embodiment, when the theme correlation degree parameter value is greater than 0.6, it is determined as significant correlation. The theme correlation degree sequence is traversed to find a segment with continuous parameter values greater than 0.6. For example, the theme correlation degree parameter values of the first and second rounds of dialogue are 0.7 and 0.8, the third round is 0.5, and the fourth to sixth rounds are 0.7, 0.65, and 0.72. Then, the first and second rounds and the fourth to sixth rounds can be determined as continuous significant correlation dialogue segments. The first and second rounds of dialogue are focused on the basic functions of the product, and the corresponding continuous theme is "consultation on the basic functions of the product". The fourth to sixth rounds of dialogue focus on the use method of the product, and the corresponding continuous theme is "discussion on the use method of the product".

[0116] In step S1332, the starting round and the ending round of each significant correlation dialogue segment in the historical dialogue are recorded, the difference between the ending round and the starting round is calculated, and the duration of the continuous theme is obtained.

[0117] In this embodiment, for the significant correlation dialogue segment of the first and second rounds, the starting round is 1 and the ending round is 2. The difference between the ending round and the starting round is 2-1=1, that is, the duration of the theme "consultation on the basic functions of the product" is one round interval. For the significant correlation dialogue segment of the fourth to sixth rounds, the starting round is 4 and the ending round is 6. The difference is 6-4=2, and the duration of the theme "discussion on the use method of the product" is two round intervals. In the same way, all significant correlation dialogue segments are processed, their starting rounds and ending rounds are recorded respectively, and the durations of their respective themes are calculated.

[0118] In step S1333, the order of the appearance of different themes in the historical dialogue is determined according to the order of the starting rounds of the significant correlation dialogue segments, and the theme names are arranged in order.

[0119] In this embodiment, the starting rounds of the significant correlation dialogue segments are 1, 4, 8, and 11 in turn. According to the order of the starting rounds from early to late, the corresponding themes are "consultation on the basic functions of the product", "discussion on the use method of the product", "order logistics inquiry", and "after-sales problem solving". Therefore, the order of the appearance of these themes in the historical dialogue is determined as "consultation on the basic functions of the product"— "discussion on the use method of the product"— "order logistics inquiry"— "after-sales problem solving", and the theme names are arranged in this order.

[0120] In step S1334, the position where the correlation degree parameter value suddenly decreases in the theme correlation degree sequence is found as a potential conversion node of theme conversion.

[0121] In this embodiment, the subject relevance sequence is [0.7, 0.8, 0.5, 0.7, 0.65, 0.72, 0.4, 0.68, 0.7, 0.55, 0.69, 0.71]. By observation, it is found that at round 3, the relevance parameter value suddenly decreases from 0.8 in the second round to 0.5; at round 7, it suddenly decreases from 0.72 in the sixth round to 0.4; and at round 10, it suddenly decreases from 0.7 in the ninth round to 0.55. The positions where the parameter values suddenly decrease, i.e., rounds 3, 7, and 10, are marked as potential conversion nodes of subject conversion.

[0122] At step S1335, the rationality of the potential conversion nodes is verified in combination with the subject conversion frequency, and after confirmation of rationality, the potential conversion nodes are marked as formal conversion nodes, and the subjects corresponding to the nodes before and after are determined.

[0123] In this embodiment, it is statistically found that the subject conversion frequency of the entire historical dialogue is 3 times. The potential conversion nodes are rounds 3, 7, and 10, a total of 3, which is consistent with the subject conversion frequency, indicating that these potential conversion nodes are reasonable. After marking them as formal conversion nodes, it is determined that the subject before round 3 is “consultation on basic functions of goods”, the subject after round 3 is “discussion on use methods of goods”, the subject before round 7 is “discussion on use methods of goods”, the subject after round 7 is “query on order logistics”, the subject before round 10 is “query on order logistics”, and the subject after round 10 is “solution to after-sales problems”.

[0124] At step S1336, the appearance order, duration, and conversion nodes of different subjects are arranged into structured data.

[0125] In this embodiment, the appearance order 1, duration 1 round interval, and corresponding conversion node front position (none, because it is the first subject) of the subject “consultation on basic functions of goods” are arranged in; the appearance order 2, duration 2 round interval, and conversion node round 3 of the subject “discussion on use methods of goods” are arranged in; the appearance order 3, duration 2 round interval (rounds 8-9), and conversion node round 7 of the subject “query on order logistics” are arranged in; the appearance order 4, duration 2 round interval (rounds 11-12), and conversion node round 10 of the subject “solution to after-sales problems” are arranged in, forming structured data. The structured data is presented in the form of a list, and each element contains subject name, appearance order, duration, and conversion node information.

[0126] At step S1337, the duration of each subject is normalized, and after converting the duration into the proportion occupied in the entire dialogue duration, the duration parameter in the structured data is adjusted according to the normalization result.

[0127] In this embodiment, the total number of turns of the entire conversation is 12, and the total duration is 11 turn intervals. The duration of the topic "consultation of commodity basic function" is 1 turn interval, and after normalization, it is 1 / 11; the duration of the topic "discussion of commodity use method" is 2 turn intervals, and after normalization, it is 2 / 11; the duration of the topic "order logistics inquiry" is 2 turn intervals, and after normalization, it is 2 / 11; the duration of the topic "solution of after-sales problem" is 2 turn intervals, and after normalization, it is 2 / 11. According to these normalized results, the duration parameters of each topic in the structured data are adjusted to the corresponding proportion values.

[0128] In step S134, the topic development knowledge path of the entire historical conversation is constructed according to the appearance order, duration and conversion node of the topic.

[0129] In this embodiment, the duration and conversion node of each topic are marked in the path with the topic appearance order as the time axis. The starting topic is "consultation of commodity information", which is switched to "inquiry of transaction status" at the 2-3 turn conversion node after 2 turns; this topic is switched back to "consultation of commodity information" at the 3-4 turn conversion node after 1 turn; and so on, the switching node and duration of each topic are recorded in turn to form a complete topic development knowledge path: consultation of commodity information (turns 1-2) - inquiry of transaction status (turn 3) - consultation of commodity information (turn 4) - raising of after-sales problem (turn 5) - inquiry of transaction status (turn 6) - consultation of commodity information (turn 7) - raising of after-sales problem (turn 8) - inquiry of transaction status (turn 9) - consultation of commodity information (turn 10) - raising of after-sales problem (turn 11) - inquiry of transaction status (turn 12).

[0130] In step S135, the user core demand summary information and the conversation topic development knowledge path are integrated to generate the historical conversation semantic understanding result.

[0131] In this embodiment, the user core demand summary information "clear the core performance and application scenarios of different versions of the product, and select the version that meets the needs" is integrated with the topic development knowledge path. In the knowledge path, the topic stages related to the core demand are marked, such as the "consultation of commodity information" stage corresponding to the main exploration process of the core demand, and the "raising of after-sales problem" stage corresponding to the use guarantee link of the demand extension, to preliminarily form the historical conversation semantic understanding result.

[0132] In step S136, the user core demand summary information is hierarchically divided, the core demand is decomposed into multiple levels of demand, and the hierarchical relationship between the levels of demand is determined to obtain the hierarchical division result.

[0133] In this embodiment, the core requirement "clearly understand the core performance and applicable scenarios of different versions of the product, and select the version suitable for the user's needs" is divided into three levels of requirements: the first level of requirement is "product selection decision support"; the second level of requirement is "core performance comparison" and "applicable scenario matching"; and the third level of requirement is "running speed difference", "endurance capability comparison", "daily scenario adaptability", "high-intensity scenario adaptability", etc. The dependency relationship of each level of requirement is: the third level of requirement supports the second level of requirement, and the second level of requirement serves the first level of requirement, forming a hierarchical division result.

[0134] In step S137, the topic development knowledge path is divided into stages according to the topic transition nodes, the entire dialogue process is divided into different stages, each stage corresponds to one or more closely related topics, and a stage division result is obtained.

[0135] In this embodiment, the entire dialogue process is divided into four stages with the topic transition nodes as the dividing points: stage 1 (rounds 1-2) corresponds to the "consulting product information" topic, which is the demand preliminary exploration stage; stage 2 (rounds 3-4) includes the "inquiring about the transaction status" and "consulting product information" topics, which is the demand and process parallel stage; stage 3 (rounds 5-8) includes the "raising post-sale questions", "inquiring about the transaction status" and "consulting product information" topics, which is the demand deepening and problem feedback stage; and stage 4 (rounds 9-12) includes the "inquiring about the transaction status", "consulting product information" and "raising post-sale questions" topics, which is the decision confirmation and guarantee stage, and a stage division result is obtained.

[0136] In step S138, the historical dialogue semantic understanding result is optimized based on the hierarchical division result and the stage division result, and a final historical dialogue semantic understanding result is formed.

[0137] In this embodiment, according to the hierarchical division result, the specific content and the association relationship of each level of requirement are supplemented in the historical dialogue semantic understanding result; and according to the stage division result, the core requirement level corresponding to each stage is marked in the topic development knowledge path, such as the preliminary exploration of the second level of requirement "core performance comparison" in stage 1 and the in-depth confirmation of the third level of requirement "running speed difference" in stage 3. Through the above optimization, the historical dialogue semantic understanding result clearly presents the hierarchical structure of the user's requirements, and clearly presents the corresponding relationship between the topic development and the requirement evolution, and a final historical dialogue semantic understanding result is formed.

[0138] In step S140, an adapted dialogue interaction strategy is generated according to the historical dialogue semantic understanding result, and the dialogue interaction strategy includes response content focus information, guidance direction and rhythm control mode information.

[0139] In this embodiment, the user core demand summary information in the historical dialogue semantic understanding result and the dialogue topic development knowledge path are used to generate the dialogue interaction strategy. The user core demand summary information shows that the user mainly focuses on product functions, order logistics and after-sales problems, and the dialogue topic development knowledge path presents the sequence and conversion of the topics. In view of this, it is determined that the response content should focus on detailed description of product functions, real-time status of order logistics and solutions to after-sales problems; the guidance direction is to gradually transition from product function introduction to order tracking and then to after-sales guarantee; and the rhythm control mode needs to be maintained at a moderate response speed and sentence length according to the user's previous dialogue habits.

[0140] In step S141, the user core demand summary information in the historical dialogue semantic understanding result is analyzed to determine the information field covered by the response content and the focus information of the response content.

[0141] In this embodiment, the user core demand summary information is analyzed to find that the user core demand includes understanding the functional characteristics of the product, mastering the logistics progress of the order and solving problems in the use of the product. It is thus determined that the response content needs to cover the product function field, the order logistics field and the after-sales problem field. In these fields, the user's inquiry about product functions is the most frequent and detailed, and the demand for solving after-sales problems is the most urgent, so the product function introduction and the after-sales problem solution are taken as the focus information of the response content, and the order logistics information is taken as the secondary focus.

[0142] In step S142, the dialogue topic development knowledge path is analyzed to determine the development stage and reference extension direction of the current dialogue topic, and the guidance direction of the dialogue is determined in combination with the user core demand.

[0143] In this embodiment, the dialogue topic development knowledge path shows that the current dialogue topic is in the "product use method discussion" stage, has experienced the "product basic function consultation" stage and may extend to the "order logistics inquiry" stage. In combination with the user core demand, after understanding the product use method, the user is likely to be concerned about when the order will be delivered, so the guidance direction of the dialogue is determined to gradually guide from the detailed explanation of the product use method to the notification of the order logistics information, while taking into account the after-sales problems that the user may raise.

[0144] In step S143, the dialogue rhythm habits of the target user are evaluated by referring to the consultation frequency of the target user in the historical dialogue, the length of the user consultation sentence and the response speed of the target user, and the rhythm control mode information to be used in the new dialogue interaction is determined, the rhythm control mode information including control information of the response speed and the response sentence length.

[0145] In this embodiment, statistical history dialogue finds that the target user sends an inquiry sentence every 20 minutes on average, the inquiry frequency is moderate; the user inquiry sentence contains an average of 25 Chinese characters, the length is moderate; the average time from receiving the customer service response to sending the next inquiry is 15 minutes, the response speed is moderate. Based on these data, the target user's dialogue rhythm habit is evaluated as moderate rhythm. Therefore, in the new dialogue interaction, the response speed is controlled within 5-10 minutes, the response sentence length is controlled between 30-50 Chinese characters, and the corresponding rhythm control mode information is formed.

[0146] For example, in step S1431, the number of user inquiry sentences sent by the target user in a unit of time in the historical dialogue is counted to obtain the inquiry frequency of the target user.

[0147] In this embodiment, the historical dialogue in the past 2 hours is selected as the statistical time period, and in this time period, the target user has sent 6 inquiry sentences. The inquiry frequency of the target user is calculated as 6 / 2 hours=3 / hour, that is, the target user sends 3 inquiry sentences on average per hour, which is taken as the inquiry frequency data of the target user.

[0148] In step S1432, the number of characters in each user inquiry sentence of the target user in the historical dialogue is measured, and the average sentence length is calculated.

[0149] In this embodiment, the number of characters in each inquiry sentence of the target user is measured, which is 22, 28, 24, 30, 20, and 26 respectively. The sum of these numbers of characters is 22+28+24+30+20+26=150, and divided by the number of inquiry sentences 6, the average sentence length is 150 / 6=25 characters.

[0150] In step S1433, the time interval from receiving the customer service response sentence to sending the next user inquiry sentence of the target user in the historical dialogue is calculated to obtain the response speed of the target user.

[0151] In this embodiment, the time interval from receiving the customer service response sentence to sending the next inquiry sentence of the target user is recorded, which is 12 minutes, 18 minutes, 14 minutes, 16 minutes, 13 minutes, and 17 minutes respectively. The sum of these time intervals is 12+18+14+16+13+17=90 minutes, and divided by the number of intervals 6, the average time interval is 90 / 6=15 minutes, that is, the response speed of the target user is 15 minutes on average. reply once.

[0152] In step S1434, the inquiry frequency, the average sentence length, and the response speed are integrated to construct a user dialogue rhythm evaluation model, and the dialogue rhythm habit of the target user is divided into different types.

[0153] In this embodiment, the constructed user conversation rhythm evaluation model contains three input dimensions, namely consultation frequency, average sentence length and response speed. Set the consultation frequency high (> 5 per hour), medium (3-5 per hour), low (< 3 per hour); average sentence length long (> 40 words), medium (20-40 words), short (< 20 words); response speed fast (< 10 minutes), medium (10-20 minutes), slow (> 20 minutes). The consultation frequency of the target user is 3 per hour (medium), the average sentence length is 25 words (medium), and the response speed is 15 minutes (medium). The user conversation rhythm evaluation model outputs the target user's conversation rhythm habit type as "medium rhythm type".

[0154] Step S1435, preset corresponding rhythm control parameters for different conversation rhythm types, the rhythm control parameters including response speed interval and response sentence length suitable range.

[0155] In this embodiment, for "fast rhythm type" users, the preset response speed interval is 1-5 minutes, and the response sentence length suitable range is 10-20 words; for "medium rhythm type" users, the response speed interval is 5-10 minutes, and the response sentence length suitable range is 30-50 words; for "slow rhythm type" users, the response speed interval is 10-15 minutes, and the response sentence length suitable range is 50-80 words. These preset rhythm control parameters are stored in the parameter library of the system, which is convenient for calling according to the rhythm type of the user.

[0156] Step S1436, select the corresponding rhythm control parameters according to the conversation rhythm type to which the target user belongs, and determine the rhythm control mode information to be adopted in the new conversation interaction.

[0157] In this embodiment, the target user belongs to "medium rhythm type", and the corresponding rhythm control parameters are selected from the parameter library, that is, the response speed interval is 5-10 minutes, and the response sentence length suitable range is 30-50 words. These parameters are determined as the rhythm control mode information to be adopted in the new conversation interaction, so as to ensure that the response matches the rhythm habit of the user.

[0158] Step S1437, compare and analyze the conversation rhythm habits of the target user in different theme stages, and identify the stage difference information of the rhythm habit.

[0159] In this embodiment, the consultation frequency, average sentence length and response speed of the target user in the four theme stages of "commodity basic function consultation", "commodity use method discussion", "order logistics query" and "after-sales problem solving" are counted respectively. It is found that in the "order logistics query" stage, the user consultation frequency is 4 per hour, the average sentence length is 20 words, and the response speed is 10 minutes, and the rhythm is relatively fast; in the "commodity use method discussion" stage, the consultation frequency is 2 per hour, the average sentence length is 30 words, and the response speed is 20 minutes, and the rhythm is relatively slow. Through comparative analysis, the stage difference information of these rhythm habits is identified.

[0160] In step S1438, the rhythm control mode information is adjusted according to the stage difference information to form the final rhythm control mode information.

[0161] In this embodiment, based on the stage difference information, in the "order logistics query" stage, the rhythm control mode information is adjusted to a response speed interval of 3-7 minutes and a response sentence length suitable range of 20-40 words; in the "commodity use method discussion" stage, it is adjusted to a response speed interval of 7-12 minutes and a response sentence length suitable range of 40-60 words; the other stages keep the original "medium rhythm type" parameters. After the above adjustment, the final rhythm control mode information is formed.

[0162] In step S144, the corresponding dialogue interaction strategy is constructed according to the response content focus information, the guidance direction and the rhythm control mode information, and the priority of the response content focus information is set for different user core demand levels, and the strength of the guidance direction is adjusted according to the difference of the theme development stage, and the dialogue interaction strategy is optimized based on the set priority and strength adjustment result to form the final dialogue interaction strategy.

[0163] In this embodiment, first, the preliminary dialogue interaction strategy is constructed according to the response content focus information (commodity function introduction and after-sales problem solving scheme are the focus, and order logistics information is secondary), the guidance direction (from commodity use method to order logistics, and considering after-sales) and the rhythm control mode information. Then, the priority of the user core demand level is set, and the core demand "commodity function understanding" and "after-sales problem solving" have high priority, and the "order logistics query" has medium priority. According to the theme development stage, in the "commodity use method discussion" stage, the strength of guiding to order logistics is set to medium, and after entering the "order logistics query" stage, the strength is improved to high. Based on these priority and strength adjustment results, the preliminary dialogue interaction strategy is optimized to clearly define the response focus, guidance strength and rhythm control details in different stages to form the final dialogue interaction strategy.

[0164] Step S150, the dialogue interaction strategy is executed to realize the dialogue interaction operation with the target user, and a response sentence conforming to the dialogue interaction strategy is output and a new input sentence of the target user is received.

[0165] In this embodiment, the operation is performed according to the final dialogue interaction strategy. When the user asks about a certain detail of the use of the commodity, the use detail is explained in detail according to the response content focus, and the reply is given within 7 minutes according to the rhythm control mode, the sentence length is controlled to be about 40 words, and the user is appropriately guided to pay attention to the logistics information to be updated. After outputting the response sentence, the new input sentence of the user is waited in real time, and once it is received, it is included in the new dialogue process and the interaction is continued according to the strategy.

[0166] Step S151, according to the response content focus information in the dialogue interaction strategy, a matching response template is selected from a preset response template library, and an initial response sentence is generated by filling in information related to the core demand of the user and a guiding expression corresponding to the guiding direction.

[0167] In this embodiment, the response content focus information is the introduction of the function of the commodity and the solution to the after-sales problem. From the preset response template library, the template related to the introduction of the function of the commodity is selected, that is, "Regarding [the function of the commodity], its main features include [feature 1], [feature 2], and it can play a good role in [use scenario]. If you want to know more about [related function], please let us know at any time." and the template related to the solution to the after-sales problem is selected, that is, "Regarding the [after-sales problem] you mentioned, our solution is [solution step], which can solve [problem effect] after execution. We will notify you in time if there is any progress in the future."

[0168] When the user asks about a certain function of the commodity, the information related to the function of the commodity is filled in, such as "Regarding the energy-saving mode of the commodity, its main features include low-power consumption operation, automatic performance adjustment, and it can play a good role in long-time use scenario. If you want to know the switching method of the mode, please let us know at any time.", and the expression corresponding to the guiding direction is added, that is, "In addition, the logistics information of the commodity you purchased is expected to be updated today, and you can pay attention to it at that time.", to generate an initial response sentence.

[0169] Step S152, the initial response sentence is adjusted according to the rhythm control mode information, and an adjusted response sentence is output.

[0170] In this embodiment, the rhythm control mode information requires a response sentence length of 40-60 words in the "product use method discussion" stage, and the response speed is within 7-12 minutes. The initial response sentence generated has a length of 70 words, which exceeds the range and needs to be adjusted. Some descriptions need to be simplified to make the length about 50 words, such as "Regarding the energy-saving mode of the product, its characteristics are low power consumption, automatic adjustment, and suitable for long-term use. If you want to know the switching method, please inform. The logistics information is expected to be updated today, please pay attention." After confirming that the adjusted sentence meets the rhythm control requirements, output the response sentence within 10 minutes.

[0171] Step S153, after outputting the adjusted response sentence, real-time monitoring the input state of the target user, waiting to receive the new input sentence of the target user, recording the content and receiving time of the new input sentence of the target user.

[0172] In this embodiment, after outputting the adjusted response sentence, the user's input interface can be monitored in real time, and the user's input interface is in standby state. When the user inputs a new sentence "How to switch to energy-saving mode? Can I check the logistics information now?" after 30 minutes, the sentence is immediately received, and the content is recorded as "How to switch to energy-saving mode? Can I check the logistics information now?" and the receiving time is 2:15 pm on the same day.

[0173] Step S154, adding the new input sentence of the target user to the historical dialogue information set.

[0174] In this embodiment, the new input sentence "How to start energy-saving mode?" of the user is added to the historical dialogue information set. During the adding process, the receiving time of the sentence and its position in the dialogue round can be automatically recorded. Since the sentence is issued by the user in the 13th round of dialogue, it is marked as the 13th round of user consultation sentence and associated with the corresponding customer service response sentence.

[0175] At the same time, in order to ensure the integrity and consistency of the historical dialogue information set, the format of the added set can be checked to ensure that the storage format of the sentence is consistent with the previous dialogue sentence, including the format of the sentence content, position mark, timestamp and other information. In addition, considering that the content involved in the sentence may be related to the previous product function consultation, the association information between the sentence and the related theme can be updated in the internal index, so that subsequent deep semantic mining can quickly locate and analyze.

[0176] Step S155, the effect of the adjusted response sentence is evaluated by analyzing the response speed of the target user after receiving the adjusted response sentence and the content of the new input sentence of the target user. If the evaluation result shows that the response sentence does not meet the expectation of the target user, the response content focus information or the rhythm control mode information in the dialogue interaction strategy is adjusted, and the response sentence is regenerated.

[0177] In this embodiment, after receiving the adjusted response sentence, the target user sends a new input sentence "How to turn on the energy saving mode?" within 1 minute. First, the response speed is analyzed, and the response time this time is compared with the average response speed of the user in the historical dialogue. If the historical average response speed is 2 minutes, the response speed of 1 minute this time is faster than the average level, indicating that the user has a fast feedback to the response sentence.

[0178] Then, the content of the new input sentence is analyzed. The sentence is a query about the start mode of the energy saving mode, and the previous adjusted response sentence mainly focuses on the basic functions of the product and does not involve the energy saving mode related content. Combining the response speed and the content of the sentence, it is judged that the response sentence has triggered the user's fast feedback, but has not completely covered the user's potential needs. The effect evaluation result is not in line with the expectation of the target user.

[0179] Based on this, the response content focus information in the dialogue interaction strategy can be adjusted, and the energy saving mode related content can be added to the focus. At the same time, considering that the user responds faster this time, the current rhythm control mode information is appropriately maintained. Then, the response sentence is regenerated according to the adjusted strategy, such as "The start mode of the energy saving mode of the product is: long press the function key for 3 seconds, and the screen displays the word 'energy' to indicate that the start is successful. If you have any other operation questions, you can inform at any time." Step S156, the regenerated response sentence is output to the target user, and the new input sentence of the target user is received again until the dialogue interaction operation is completed.

[0180] In this embodiment, the regenerated response sentence "The start mode of the energy saving mode of the product is: long press the function key for 3 seconds, and the screen displays the word 'energy' to indicate that the start is successful. If you have any other operation questions, you can inform at any time." is output to the target user.

[0181] The system continues to monitor the input state of the user and waits to receive the new input sentence of the user. If the user subsequently inputs "I know, thank you", it indicates that the user's problem is solved, and the dialogue interaction operation is completed. If the user continues to ask other questions, such as "How much longer will the endurance time be extended in the energy saving mode?", the above steps can be repeated to generate, output and receive the response sentence again until the user has no new input or explicitly indicates the end of the dialogue.

[0182] During the whole process, the historical dialogue information set is constantly updated to ensure that the content of each interaction is accurately recorded. At the same time, all data related to user dialogue are stored in an encrypted manner, and data desensitization technology is used to process user identification and other private information to prevent privacy leakage and protect user data security.

[0183] Figure 2 A schematic diagram of exemplary hardware and software components of the dialogue interaction system 100 based on long historical dialogue semantic understanding that can implement the idea of the present application provided by some embodiments of the present application is shown. For example, the processor 120 can be used in the dialogue interaction system 100 based on long historical dialogue semantic understanding and used to perform the functions in the present application.

[0184] For example, the dialogue interaction system 100 based on long historical dialogue semantic understanding can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. Exemplarily, the dialogue interaction system 100 based on long historical dialogue semantic understanding can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The methods of the present application can be implemented according to these program instructions. The dialogue interaction system 100 based on long historical dialogue semantic understanding also includes an I / O interface 150 between the computer and other input / output devices.

[0185] In addition, the present application also provides a readable storage medium, in which computer executable instructions are preset, and when a processor executes the computer executable instructions, the dialogue interaction method based on long historical dialogue semantic understanding is implemented.

[0186] It should be noted that, in order to simplify the description of the present application and to help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, drawing, or description thereof.

Claims

1. A conversational interaction method based on semantic understanding of long-history conversations, characterized in that: The method comprises: Obtain a collection of historical conversation information of a target user on an e-commerce platform. The collection of historical conversation information includes multiple rounds of user inquiry statements and corresponding customer service response statements. The user inquiry statements include inquiries about product information, order processing requests, and after-sales feedback. The customer service response statements include answers to the user inquiry statements and guidance information. Performing deep semantic mining on the historical conversation information set to obtain the target user's conversation intention features and conversation topic association features. The conversation intention features reflect the target user's potential needs in each round of conversation, and the conversation topic association features reflect the continuation and conversion relationship between topics in different rounds of conversation. Generate historical conversation semantic understanding results based on the conversation intention features and the conversation topic association features, wherein the historical conversation semantic understanding results include user core demand summary information and conversation topic development knowledge path; Generate an adaptive dialogue interaction strategy based on the semantic understanding results of the historical dialogue, wherein the dialogue interaction strategy includes information on the focus of the response content, the guidance direction, and the rhythm control method; The dialog interaction strategy is executed to implement a dialog interaction operation with the target user, a response statement that complies with the dialog interaction strategy is output, and a new input statement from the target user is received.

2. The method for dialogue interaction based on semantic understanding of long-history dialogues according to claim 1, characterized in that: The deep semantic mining process of the historical conversation information set to obtain the conversation intention features and conversation topic association features of the target user includes: Sentence processing is performed on the user consultation sentences and customer service response sentences in the historical dialogue information set according to the natural pauses and semantic integrity of the sentences, and multiple dialogue sentence units are divided, and each dialogue sentence unit retains the original context position information; Extracting feature words related to e-commerce transactions from the multiple dialogue sentence units, and determining a core feature word set through word frequency statistics and semantic importance evaluation, wherein the feature words include product names, transaction process terms, and problem description words; Perform semantic analysis on the core feature word set in combination with the context to determine the basic intent category corresponding to each user consultation statement; Analyzing the topic continuity between the sentence units of different rounds of dialogue based on the basic intention category, and generating a topic relevance parameter by calculating a topic similarity parameter, wherein the topic relevance parameter is used to indicate the degree of relevance between adjacent rounds of dialogue in terms of topic; Constructing the target user's conversation intention feature and conversation topic relevance feature based on the basic intention category and the topic relevance parameter, wherein the conversation intention feature includes an intention category sequence and an intention strength descriptor, and the conversation topic relevance feature includes topic switching frequency and topic duration parameters; The conversation intention features are weighted and integrated according to the importance of each conversation round in the overall historical conversation, and a first weight corresponding to each conversation round is assigned. The higher the first weight, the greater the impact of the conversation round on the final conversation intention feature; The conversation topic association features are time-calibrated, and the weighted integrated conversation intention features and the time-calibrated conversation topic association features are cross-validated. If the target user conversation tendencies reflected by the two are consistent, the conversation intention features and the conversation topic association features are confirmed to be obtained; if there is a deviation in the target user conversation tendencies reflected by the two, the feature extraction parameters are readjusted and deep semantic mining processing is performed again until the conversation intention features and the conversation topic association features are obtained.

3. The method for dialogue interaction based on semantic understanding of long-history dialogues according to claim 2, characterized in that: The step of extracting feature words related to e-commerce transactions from the multiple dialogue sentence units and determining a core feature word set through vocabulary frequency statistics and semantic importance evaluation includes: Perform part-of-speech tagging on each dialogue sentence unit to identify nouns, verbs, and adjectives, and select nouns and verbs that can be used as feature words; Compare the selected words with the preset e-commerce domain vocabulary to match candidate feature words belonging to product names, transaction process terms, and problem description words; Counting the total number of times each candidate feature word appears in the historical conversation information set, calculating the distribution frequency of each candidate feature word in each round of conversation, and obtaining a frequency distribution feature; Based on the semantic weight of each candidate feature word in the sentence, its closeness to the conversation topic, and its criticality in the e-commerce transaction scenario, the candidate feature word is evaluated for semantic importance and assigned a corresponding importance score. A screening threshold is set based on the frequency distribution characteristics and importance score, and candidate feature words with importance scores higher than the screening threshold are selected to form a core feature word set. Each word in the core feature word set retains its original position mark in the dialogue sentence unit; Deduplication is performed on the core feature word set, and relevance ranking is performed on the core feature word set after deduplication according to the degree of semantic relevance between the core feature words to determine the relevance ranking result; The internal structure of the core feature word set is optimized according to the relevance ranking result to form a final core feature word set.

4. The method for dialogue interaction based on semantic understanding of long-history dialogues according to claim 2, characterized in that: The semantic analysis of the core feature word set is performed in combination with the context to determine the basic intent category corresponding to each user consultation statement, including: Construct an association mapping table between basic intent categories and core feature words, which records the mapping relationship between product name feature words corresponding to product information inquiries, transaction process terminology feature words corresponding to transaction status inquiries, and question description vocabulary feature words corresponding to post-sales questions. For each user inquiry statement, extract the core feature words contained in the user inquiry statement from the core feature word set, and preliminarily determine the candidate basic intent category that the user inquiry statement may correspond to based on the mapping relationship recorded in the association mapping table; Analyze the context of the user inquiry statement based on the mapping logic of the association mapping table, check the core feature words and semantic tendencies involved in the adjacent customer service response statements and user inquiry statements, and verify whether the initially determined candidate basic intent category meets the mapping relationship; If there are multiple candidate basic intent categories, calculate the matching degree between each candidate and the context according to the strength of association between the feature words and the intent categories in the association mapping table, and select the candidate basic intent category with the highest matching degree as the basic intent category corresponding to the user's consultation statement; Arrange the basic intent categories corresponding to each user's consultation statement in the order of conversation turns to form a basic intent category sequence; Calculating the proportion of each basic intent category in the basic intent category sequence, and analyzing the distribution characteristics of different intent categories, wherein the distribution characteristics are positively correlated with the preset association strength of the association mapping table; The basic intent category sequence is adjusted based on the distribution characteristics to form a final basic intent category sequence.

5. The method for dialogue interaction based on semantic understanding of long-history dialogues according to claim 2, characterized in that: The analyzing the topic continuity between the conversation sentence units of different rounds based on the basic intention category and generating the topic relevance parameter by calculating the topic similarity parameter includes: Constructing a topic feature vector for each basic intent category, wherein the topic feature vector is composed of the core feature words corresponding to the basic intent category and their second weights, where the second weights are determined according to the importance scores of the feature words; Selecting adjacent rounds of dialogue sentence units, and extracting the topic feature vectors of the basic intent categories corresponding to the adjacent rounds of dialogue sentence units; Calculate the cosine similarity between the two topic feature vectors to obtain the topic similarity parameter. The larger the topic similarity parameter, the closer the topic connection between the two rounds of dialogue. The relevance level is divided according to the size of the topic similarity parameter, and different relevance levels correspond to different topic relevance parameter values; The topic relevance parameters of adjacent conversation rounds are recorded in turn order to form a topic relevance sequence. At the same time, the total number of topic transitions in the entire historical conversation information set is calculated, and the topic transition frequency is obtained by combining the total number of conversation rounds. Perform sliding window analysis on the topic relevance sequence by setting a fixed-length window, traverse the topic relevance sequence, and calculate the average relevance parameter within the window; The topic relevance sequence is adjusted according to the average relevance parameter, and the final topic relevance parameter is generated by combining the adjusted topic relevance sequence and the topic conversion frequency.

6. The method for dialogue interaction based on semantic understanding of long-history dialogues according to claim 1, characterized in that: Generating the historical conversation semantic understanding result based on the conversation intention feature and the conversation topic association feature includes: Performing statistical analysis on the intent category sequence in the conversation intention feature to identify the most frequently occurring intent category, and combining it with the intent strength descriptor to determine the main intent direction of the target user; Based on the main intention direction, relevant user inquiry statements and customer service response statements are integrated to extract key information and summarize the core needs of the target user in the entire historical conversation; Analyze the topic relevance sequence and topic conversion frequency in the conversation topic relevance features to analyze the order of appearance, duration, and conversion nodes of different topics in the conversation; Construct the knowledge path of the theme development of the entire historical dialogue based on the order of appearance, duration and transition nodes of the theme; Integrate the user's core demand summary information and the conversation topic development knowledge path to generate historical conversation semantic understanding results; Performing hierarchical division on the summarized information of the user's core needs, decomposing the core needs into multiple levels of needs, determining the subordinate relationship between the needs of each level, and obtaining a hierarchical division result; The knowledge path of topic development is divided into stages according to the conversion nodes of the topic, and the entire dialogue process is divided into different stages. Each stage corresponds to one or more closely related topics, and the stage division results are obtained; The historical conversation semantic understanding result is optimized based on the hierarchical division result and the stage division result to form a final historical conversation semantic understanding result.

7. The method for dialog interaction based on semantic understanding of long-history dialogs according to claim 6, characterized in that: The analyzing of the topic relevance sequence and topic conversion frequency in the conversation topic relevance feature to analyze the order of appearance, duration, and conversion nodes of different topics in the conversation includes: Traversing the topic relevance sequence, determining continuous significantly related conversation segments according to topic relevance parameter values, each significantly related conversation segment corresponding to a continuous topic; Record the start and end rounds of each significantly related conversation segment in the historical conversation, calculate the difference between the end round and the start round, and obtain the duration of the continuous topic; Determine the order in which different topics appear in historical conversations based on the order of the start turns of each significantly related conversation segment, and arrange the topic names in that order; Find the position where the relevance parameter value suddenly decreases in the topic relevance sequence as the potential transition node of topic transition; Verify the rationality of the potential transition node based on the frequency of topic transitions. Once confirmed to be rational, mark it as a formal transition node and determine the topics before and after the node. Organize the order of occurrence, duration, and transition nodes of different topics into structured data; The duration of each topic is normalized, and after converting the duration into a proportion of the entire conversation duration, the duration parameter in the structured data is adjusted according to the normalization result.

8. The method for dialog interaction based on semantic understanding of long-history dialogs according to claim 1, characterized in that: Generating an adapted dialogue interaction strategy based on the historical dialogue semantic understanding results includes: Analyze the user's core needs summary information in the historical conversation semantic understanding results, determine the information field covered by the response content, and determine the key information of the response content; Analyze the knowledge path of conversation topic development, determine the current conversation topic's development stage and reference extension direction, and determine the conversation's direction based on the user's core needs; By referring to the target user's inquiry frequency, inquiry sentence length, and response speed in historical conversations, the target user's conversation rhythm habits are evaluated, and the rhythm control method information to be adopted in the new conversation interaction is determined. The rhythm control method information includes control information of response speed and response sentence length; A corresponding dialogue interaction strategy is constructed based on the response content focus information, the guidance direction and the rhythm control method information, and priorities are set for the response content focus information according to different user core demand levels. The intensity of the guidance direction is adjusted according to different stages of topic development, and the dialogue interaction strategy is optimized based on the set priority and intensity adjustment results to form a final dialogue interaction strategy.

9. The method for dialog interaction based on semantic understanding of long-history dialogs according to claim 1, characterized in that: The executing the dialogue interaction strategy to implement a dialogue interaction operation with a target user, outputting a response statement that complies with the dialogue interaction strategy, and receiving a new input statement from the target user includes: Selecting a matching response template from a preset response template library based on the response content focus information in the dialogue interaction strategy, and filling in information related to the user's core needs and a guiding statement corresponding to the guidance direction to generate an initial response statement; adjusting the initial response statement according to the rhythm control mode information, and outputting the adjusted response statement; After outputting the adjusted response statement, monitoring the input status of the target user in real time, waiting to receive a new input statement from the target user, and recording the content and reception time of the new input statement from the target user; Adding the target user's new input sentence to the historical conversation information set; By analyzing the response speed of the target user after receiving the adjusted response statement and the content of the target user's new input statement, the effect of the adjusted response statement is evaluated. If the evaluation result shows that the response statement does not meet the target user's expectations, the response content emphasis information or rhythm control method information in the dialogue interaction strategy is adjusted, and a new response statement is generated; The regenerated response statement is output to the target user, and a new input statement from the target user is received again until the dialogue interaction operation is completed.

10. A conversational interaction system based on semantic understanding of long-history conversations, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the dialogue interaction method based on long history dialogue semantic understanding as described in any one of claims 1 to 9.

Images